Future Public Administration

Public administration is the part of government that has to actually do the thing: run the major-project portfolio, keep the computer systems underneath it alive, buy what it cannot build, deliver services through contracts and offices and application forms, and operate the performance regimes that are supposed to tell everyone how it is going. It is the least glamorous subject in this category and by a wide margin the best evidenced, because supreme audit institutions on three continents have been taking it apart for decades and publishing numerators and denominators rather than opinions.

What that evidence says is not what the commissioning intuition expects, and it separates into four findings that have to be kept apart. Performance regimes work, in the narrow and important sense that a high-powered target reliably moves the quantity it names — and it reliably reorganises the timing and composition of everything around that quantity, which almost nobody evaluates. The delivery machinery cannot learn, because evaluation is discretionary and stopping is nearly impossible: two thirds of a £456 billion slice of portfolio has no adequate plan to find out whether it worked, and fewer than one project in forty is stopped in a year. Procurement has become load-bearing and is the part the state cannot measure: four official and commercial sources give four irreconcilable figures for one year of United Kingdom consultancy spend, and a Canadian auditor could not establish what a single mobile application cost. And the whole accountability apparatus is bounded by bookkeeping — where the invoices do not say which project they belong to, no amount of oversight machinery downstream can recover the answer.

The scope boundary with the sibling slot is the unit of analysis, and it is worth stating precisely because the two briefs share evidence deliberately. Future Civil Services owns the workforce and its rules: headcount, pay structure, age and skill composition, appointment and tenure protections, politicisation, and the evidence on whether machines substitute for administrative staff. This brief owns the apparatus and its practices: how public organisations are structured and where a delivery unit sits in the machinery, how programmes are procured and delivered, how performance is measured and gamed, how accountability operates, and how administration fails. Where the two meet — the Canadian pay system is simultaneously a labour-relations catastrophe and a programme-delivery failure, and a border application is simultaneously a contracting-out decision and a project — the sibling takes the pay backlog and the contractor day rates, and this brief takes the programme decisions, the procurement records and the project management.

One deliberate omission follows from that cut. The published evaluations of general-purpose artificial-intelligence assistants inside civil services belong to the sibling, which reads them as evidence about administrative labour. They are not restated here, and where this brief needs them it points across rather than requoting numbers.

A selection bias should be named at the outset, because it shapes every section below. The countries whose administrative failures are documented to audit standard — the United Kingdom, Canada, Australia, the United States — are the countries with well-functioning supreme audit institutions, statutory inquiries and royal commissions. The absence of a comparable record elsewhere is not evidence of a better record. It is evidence of a thinner one, and this brief has no verified figures at all from the two jurisdictions most often held up as models of modern public administration.

Established Start with the portfolio, and get the names and the scale right, because most commentary does not. The projects authority merged with the national infrastructure commission on 1 April 2025 to form a single successor body; and the familiar five-point traffic-light scale was abolished in June 2021 in favour of a three-tier delivery-confidence assessment plus an exempt category, so any question phrased around “amber/red” has been unanswerable for five years.

Established The time series is the load-bearing table. At 31 March 2023: 244 projects, £805 billion whole-life cost. At March 2024: 227 projects, £834 billion, 27 red (12%). At March 2025: 213 projects, £996 billion, 31 red (15%) — and those 31 carried £198 billion, very nearly a fifth of the portfolio's whole-life cost. At March 2026: 189 projects, £924.2 billion, 34 red (18%). Over two years the portfolio shrank by 17% while the red count rose from 27 to 34 and the red share went from 12% to 18%. Whether that is deterioration or composition is genuinely unresolved on public data; the responsible minister asserts composition, and notes correctly that a red rating signals issues needing urgent attention rather than failure.

Established Then the finding that sits upstream of everything else in this brief. The national audit office, February 2024: “government does not routinely look at what happens after major projects are completed”, with professionals reporting that the emphasis falls on identifying potential benefits and delivering to budget and schedule rather than on ensuring projects achieve their intended purpose.

Established And that finding has a number attached. Assessed against the 227-project, £834 billion portfolio: 144 projects (63%) had some evaluation plan; 77 (34%) had a robust or adequate one; 150 (66%), carrying £456 billion of whole-life cost, did not. Of that 66%, 55% provided no evidence of any evaluation plan at all. The source is government reviewing its own evaluation practice and publishing the answer, which is what makes the figure credible.

Established The complementary number explains why nobody minds. Of 184 projects tracked over twelve months, 4 had been stopped or closed early — 2.2% — and 7 were replaced. A portfolio in which fewer than one project in forty is stopped in a year has no use for evidence about whether projects work. The evidence gap is downstream of the incentive structure, not upstream of it.

Established The second half of this brief's argument is about records, and the cleanest case in the OECD is Canadian. The Auditor General of Canada's February 2024 report on the ArriveCAN border application put its cost at $59.5 million — $53.3 million for the health component and $6.2 million for customs and immigration digitisation — and was explicit that this is the auditor's own reconstruction rather than the government's figure: “We were unable to determine a precise cost for the ArriveCAN application because of poor documentation and weak controls at the Canada Border Services Agency.” Eighteen per cent of tested contractor invoices lacked documentation sufficient to establish whether the expense belonged to the application at all. A two-person intermediary firm that performed no development work itself received $19.1 million, roughly a third of the total. There were 177 application releases with little to no documentation of testing, and no governance agreement between the two responsible agencies from April 2020 to July 2021. The audit's conclusion was that three federal organisations repeatedly failed to follow good management practices in the contracting, development and implementation of the application.

Established Read that as an administration finding rather than a scandal and it says something more useful. Every downstream accountability mechanism — parliamentary questions, value-for-money review, contract dispute, prosecution — depends on a bookkeeping layer that in this case did not exist. The auditor could reconstruct an estimate; the government could not produce a figure. What was missing was not oversight but the record that oversight consumes.

Established The same defect appears at national scale in the United Kingdom's consultancy spending, and the auditor says so in terms. Its November 2025 report gives four figures for central government consultancy spend in 2022-23 — £1.34 billion from departmental annual reports, £1.36 billion from the Treasury's own consolidated data, £1.68 billion and £2.23 billion from two commercial spend-analytics providers — with an average variance between sources of £270 million a year across 2017-18 to 2022-23. The causes are administrative: inconsistent definitions, contracts that bundle consultancy with professional services and contingent labour, departments not following the central definition, and arm's-length bodies reporting to different standards. The conclusion is unhedged: “The government does not have a clear picture of how much is being spent on consultants or how this spending has changed over time.”

Established The finding inside that report which matters most for this brief is about capability rather than cost. Knowledge transfer from consultants to departments was found to be systematically neglected rather than merely imperfect: officials typically wait until a project ends before discussing what was learned, consultants' intellectual property protections limit what can be shared, departmental turnover breaks continuity, and contracts specify learning objectives ambiguously. The auditor's formulation is the diagnosis — “Knowledge transfer should not just be a transfer of documents at the end of a project but should be an active process throughout the engagement” — and its recommendation is a default rather than a control: government “should always start with the assumption that using their own staff will be the best use of resources.”

Established And the same report contains the reason the default will not hold. Eighty-six per cent of surveyed government officials rated consultants valuable and 40% rated them extremely valuable. That is not a contradiction of the capability finding; it is the mechanism by which the capability finding persists. The people making each individual buying decision are satisfied with what they bought, and the aggregate consequence — a capability that migrates outward and does not come back — is nobody's decision.

Frontier The delivery record's most instructive current case is a government replacing a system it already knows how it broke. Canada's Phoenix pay system had 233,653 pay transactions in backlog at 30 September 2025, 155,217 of them more than a year old, affecting 133,619 employees, on a system paying $38 billion a year to more than 430,000 current and former public servants. Its replacement, Dayforce, carries a preliminary estimate of over $4.2 billion excluding transition costs. In January 2026 the schedule was shortened by three years, from a 2034 to a 2031 horizon — and schedule compression on a government-wide pay replacement is the specific decision that produced the failure being replaced. The auditor's other two flagged risks are that an uncleared backlog will import existing errors into the new system, and that slow progress on simplifying the underlying pay rules forces expensive customisation of commercial software.

Established Finally, the organisational question, which is this brief's and not the sibling's. The units built to modernise delivery from inside were reorganised out of existence without any of the decisions citing an evaluation. 18F, founded in March 2014, peaked above 250 staff in 2018 and ran on cost recovery — client agencies had to choose to pay for it, which is about as close to a market test as a public body gets — and its remaining roughly 90 staff were terminated as “non-critical” in a reduction in force on 1 March 2025. Canada's digital service was not abolished but relocated in July 2023 from the Treasury Board Secretariat, a central agency, into a service-delivery department of 32,000 staff, several layers down; a co-founder's assessment of the move is that what is lost is the whole-of-government perspective and a voice at the cabinet table, and a former staff member described the unit's structural problem as “all carrots, no sticks” — it could advise departments and could require nothing of them. Where a delivery unit sits in the machinery determines what it is able to do, and placement is decided by machinery-of-government reshuffle rather than by performance.

The genuinely open questions in public administration are narrower than the discourse, and separating them from the questions that merely sound open is most of the work. Three things usually presented as contested are settled; four things usually presented as settled are open.

Established The strongest causal evidence in this brief is on the most-mocked intervention, and it points in favour. A difference-in-differences comparing England against Scotland around the introduction of hospital waiting targets, over 1997/98 to 2003/04 with two policy-off and four policy-on years, found the numbers waiting over six months down about 20% and the numbers waiting twelve months or more down about 60%, from a baseline of roughly 23-week mean waits with maximum waits above eighteen months. The design has a real control group and the effect is large.

Established And the gaming evidence is equally real, and does not contradict it. A duration analysis of the same period across three specialties found that “hazard rates vary over time and peaks in them — high probabilities of admission — coincide with targets and change when targets change”. Separately, an audit office visiting 50 trusts found 6 (12%) with inappropriately adjusted waiting-list figures, later rising to nine identified in total; it could not assure itself of the complete accuracy of the lists; and validation typically removed 5–15% of patients. The aggregate measured outcome improved and the timing of admissions was reorganised around the threshold rather than around clinical need. Both are true, and picking one is not scepticism.

Established The statutory inquiry supplies the extreme case, and it is about displacement rather than falsification. A public inquiry into one hospital trust found management “dominated by financial pressures and achieving FT status, to the detriment of quality of care”, a focus on “reaching national access targets, achieving financial balance and seeking foundation trust status” that came “at the cost of delivering acceptable standards of care”, and standards measurement that failed because the metrics “did not focus on the effect of a service on patients”. The inquiry states no numerical excess-deaths figure, and the number widely attributed to it does not come from its findings.

Frontier The other natural experiment runs the other way and is the one nobody expects. Wales abolished school performance tables and England did not. The difference-in-differences finds “significant and robust evidence that this reform markedly reduced school effectiveness in Wales”, with top-quartile schools showing no measurable effect and the impact on segregation by ability or socioeconomic status negligible. Removing the accountability regime failed on its own terms and on its opponents' terms simultaneously. This brief carries the direction and not the magnitude: the working paper could not be retrieved and the journal version is paywalled, so the commonly quoted effect sizes are not printed here.

Established And the regime's own scorecard has a quarter of it missing. Of 366 targets in the first public service agreement round: 221 met (60.4%), 36 not met, 19 partially met — and 90 (24.6%) unassessable, 52 for want of data and 38 because no department ever published a final assessment. The widely quoted 79.1% success rate is computed after excluding those 90. The committee that produced it said the figures should be taken as indicative only, given the absence of any mechanism for independently verifying departmental reporting.

Frontier The first genuinely open question: can a state recover a capability it has contracted out? The auditor's knowledge-transfer finding says the mechanism that is supposed to prevent dependence compounding is not operating — not that it operates imperfectly, but that officials typically wait until a project ends, that intellectual property terms constrain sharing, and that contracts specify learning obligations ambiguously. If that is right, contracting out is not a reversible decision at the margin but a ratchet, and the empirical question of how far a department can climb back is open and, as far as this brief could establish, unstudied. The policy machinery is meanwhile being operated as though the answer were known: a halving target for consultancy spend, over £700 million a year of savings projected by 2028-29, and central controls reintroduced at £600,000 for ministerial approval and £100,000 for accounting-officer approval after having been withdrawn in 2023.

Frontier The second: whether performance monitoring inside a delivery organisation helps or hurts. In the only large-scale hand-coded study of completion across an entire federal bureaucracy — 4,700 public projects across 63 Nigerian federal organisations, with a companion covering 3,628 projects and tasks across 31 Ghanaian organisations in 2015 — an autonomy index is robustly positively correlated with project initiation, full completion and average completion rate, and an incentives-and-monitoring index is robustly negatively correlated with all three. The authors flag the contrast with private-sector management evidence themselves. This brief states the sign and declines to state a magnitude, because the source consulted reports directions and robustness and no coefficients. Set against the waiting-times result above, the honest position is that high-powered external targets on a named outcome and dense internal monitoring of process are different instruments with, on current evidence, different signs. Institutional Design draws the design lesson from the same study; the administrative reading is that four decades of public-management reform have been built on the opposite prior about the second instrument.

Frontier The third: whether a digital delivery unit is a durable organisational form or a temporary one. The record shows units founded, expanded, relocated and dissolved by executive instrument and machinery-of-government reshuffle, in every case without a published evaluation. The surviving United Kingdom unit is being expanded by prime-ministerial target rather than by measured performance, and its January 2026 roadmap reports a set of outputs unusually specific for this field: a single sign-on service at over 13 million users across more than 120 services, a government application at 316,000 downloads to end-December 2025, over 1,000 probation officers using bespoke transcription tools, over 70 published algorithmic transparency records, and a consultation-analysis tool with a median review time of 23 seconds per response. Those are self-reported by an interested party, and they are checkable in kind in a way the same document's £1.2 billion a year savings target is not.

Frontier The fourth: how much of the state's contracting is contracting at all. A two-person firm receiving $19.1 million to perform no development work is not a supplier in the sense procurement rules imagine; it is an intermediary that converts a staffing problem into a contract. Whether that is an isolated case or a structural feature of how governments buy technical capacity cannot be established from anything this brief consulted, because the spend data that would answer it are the same data four sources cannot reconcile.

Established What is not open, stated plainly because it is repeatedly presented as open. Whether a high-powered target moves the quantity it names: it does, on the strongest quasi-experiment in the subject. Whether it also distorts the timing and composition of what surrounds it: it does, in the same period and the same service. Whether governments know what they spend on consultants: they do not, and their own auditor says so. Whether major projects are evaluated after completion: two thirds of a £456 billion portfolio slice has no adequate plan to be. And whether the digital delivery units were closed because they failed: no evaluation is cited in any closure examined here.

Established The first bottleneck is the legacy estate and it is quantified on both sides of the Atlantic. 28% of central government systems were classified as legacy in 2024, up from 26%; 22% of legacy systems are red-rated for both likelihood and impact; and 28% of those red-rated systems have no remediation funding. Two large departments spend 70–85% of their technology budgets on upkeep rather than modernisation, against 67–70% in the regulated private sector and about 60% among digital leaders. An audit office on the other side reports that agencies typically spend about 80% of a >$100 billion annual IT budget on operations and maintenance. The two figures agree.

Established The systems are older than the debate assumes. A criminal-records database introduced in 1974 still carries around 150 million searches and updates a year; a pensions system from 1988 pays around £104 billion of state pension annually. An audit review of 69 federal legacy systems across 24 agencies selected 11 as most critical: they ranged from 23 to 60 years old, the oldest commissioned in 1964, 8 of 11 used legacy programming languages, and only 3 of 11 had complete modernisation plans. The practitioner testimony in government's own review names the reason nothing gets fixed: “Every year at least one … reprioritisation exercises, they always end up deprioritising the ‘fix the stack’ work.”

Established The second is that legacy destroys the market that would have disciplined its cost. When the 1974 system's support contract was renewed, “no other suppliers came forward to challenge the incumbent, who was awarded a new £48 million contract in 2022”. That is the mechanism in one sentence: the estate does not merely cost money to keep, it removes the competitive pressure on what keeping it costs. Alongside it sits a capacity arithmetic that is stark on its own terms — fifteen people in the government commercial function focused on relationships with nineteen strategic digital suppliers, in a market third parties put at at least £14 billion a year, a figure the audit office flags as incomplete because “government has not provided a more reliable figure”, and in which three cloud providers hold over 60% share.

Established The third binds harder than any of the above and is almost never listed as a constraint: the records. Where 18% of tested invoices do not identify their project, where four sources give four figures for one year's consultancy spend, and where 177 releases carry little or no documentation of testing, the binding constraint on accountability is not the will to hold anyone to account. It is that the question cannot be answered from the state's own books. Every remedy proposed for the other bottlenecks in this list — better appraisal, tougher approvals, post-completion review — consumes records of exactly the kind that were found missing.

Frontier The fourth is the capability ratchet in contracting. Knowledge transfer from consultants was found to be systematically neglected, and the sequence is self-reinforcing: work goes outside because the capability is thin, the transfer that would rebuild it does not happen, and the capability is thinner at the next decision. The individual buyers are content — 86% of officials rated consultants valuable and 40% extremely valuable — which is why the ratchet has no natural opponent inside the organisation.

Established The fifth is that the portfolio cannot stop things. Four of 184 tracked projects were stopped or closed early in twelve months. At a 2.2% stop rate, the portfolio changes by completion and reclassification rather than by portfolio management, and evaluation evidence has no decision to feed.

Frontier The sixth is schedule compression, which recurs because it is the only lever that appears to cost nothing. A government-wide pay replacement estimated at over $4.2 billion had its schedule shortened by three years in January 2026, with a backlog of 233,653 transactions still uncleared and the auditor flagging in advance that an uncleared backlog will import existing errors into the new system. Compression is chosen at the moment when a programme is already behind, which is the moment its estimates are least reliable.

Frontier The seventh is that some of the standard remedy may be part of the problem. Across 4,700 Nigerian and 3,628 Ghanaian public projects and tasks, autonomy predicts completion positively and incentives-and-monitoring predicts it negatively, robustly and in the same direction across both instruments. The sign is what is reported and the sign is all this brief states. If it generalises to rich-country bureaucracies — which is untested — then adding approval gates to a delivery organisation that is already missing deadlines is a treatment with the wrong sign.

Established What is not a bottleneck, stated plainly. Evidence that performance regimes move the measured quantity is not lacking; it is among the better-identified findings in social science. Technology is not the constraint on the legacy estate — migration routes are well understood and are deprioritised annually by choice. And political appetite is not the constraint on evaluation: the evaluation review, the audit reports and the portfolio data are all published, funded and read. What is missing is a rule that makes any of them consequential.

Established Nothing in this brief waits on a research result. There is no discovery pending whose arrival would change what a public administration can deliver. What it waits on is rules, records and money, and the absence of the second is the most consequential because it disables everything downstream of it.

Established It depends first on evaluation being required rather than discretionary. Two thirds of a £456 billion portfolio slice has no adequate evaluation plan and 55% of that group produced no evidence of any plan, in a system where nobody routinely looks at what happens after completion. This is a funding rule, not a capability.

Established It depends on a portfolio in which projects can actually be stopped. At a 2.2% stop rate, evaluation evidence has no decision to inform, which is why the two dependencies are one dependency stated twice.

Established It depends on contracting records good enough to audit. One auditor could not determine what a single application cost because 18% of tested invoices did not say which project they belonged to; another found four irreconcilable figures for one year's consultancy spend, with an average variance of £270 million a year. Bookkeeping is the binding input to accountability and is treated everywhere as a clerical matter.

Frontier It depends on an active knowledge-transfer obligation rather than a documentary one. The auditor's finding is that transfer is neglected as a matter of practice, and its recommended default — start from the assumption that own staff are the best use of resources — is a presumption rather than a control. Whether a presumption can hold against 86% of buyers who are satisfied with what they bought is the open half of this dependency.

Established It depends on funded remediation for the legacy estate and on more than one bidder for a legacy support contract. Both are absent: 28% of the highest-risk systems have no remediation funding, and a £48 million renewal was awarded uncontested because nobody challenged the incumbent.

Established What depends on it is most of the rest of this map. Civic Technology inherits the legacy estate as a constraint. AI-Assisted Governance inherits the procurement structure and the quality-assurance capacity documented in the sibling brief, since what a government can deploy is set by what it has already bought. Megaproject Governance shares the delivery engineering. And every brief on this map whose constraint is delivery rather than discovery depends on there being an administration capable of delivering — a dependency this brief declines to assert as an enabling edge, for the reason given in the tech tree.

Established The cheapest high-value intervention is to make evaluation a condition of funding rather than a nice-to-have. The measurement already exists: 34% of the portfolio has an adequate plan, 66% does not, and 55% of the shortfall group produced no evidence of a plan at all. That figure is published annually by government about itself. Making it a gate rather than a report would generate, within one project cycle, the outcome evidence this entire subject lacks.

Established Second, and specific: re-base the optimism-bias uplifts. The empirical basis for a national capital-appraisal regime is a single consultancy study from 2002, and so far as this brief could establish it has never been re-based in over two decades. The outturn data to do it — two decades of completed major projects with decision-to-build estimates and final costs — is held by the body that publishes the portfolio.

Frontier Third: publish the post-completion review. An audit office that has to write that government does not routinely look at what happens after projects finish is describing a policy choice, not a technical obstacle. A requirement to publish a short outcome assessment three years after completion would cost a rounding error against £924 billion and would make the next twenty years of this brief writable.

Frontier Fourth, and the one experiment in this brief that has never been run anywhere in a rich-country administration: measure management practice against completion. The Nigerian and Ghanaian studies hand-coded engineering assessments of project completion across 63 and 31 public organisations respectively and matched them to a survey-based index of autonomy and of monitoring. Nothing equivalent exists for a European or North American bureaucracy, which is remarkable given that the same countries publish project-level delivery-confidence ratings and reset records annually. The instrument exists, the outcome data exist, and nobody has joined them.

Frontier Fifth: the reimbursement-adequacy readout that is already complete. A spectrum repacking programme moved nearly a thousand broadcast stations over 39 months in ten phases against a fixed $1.75 billion reimbursement fund that stakeholders doubted in advance, with an identified physical constraint — the supply of experienced tower crews — and an identified excluded class, low-power stations and translators, that received no reimbursement at all. That is a rare object: a large implementation programme with a pre-registered cost envelope, a pre-registered doubt about it, and a completed outturn. Whether the fund sufficed, and what the excluded class actually bore, is answerable from public records and would calibrate every reimbursement-based transition programme that follows.

Established Sixth, a negative result already recorded and worth preserving as one. A 2012 shared-services strategy projected £128 million a year; over three years it delivered £90 million of savings against £94 million of investment, with four planned customers withdrawing and one contract terminated. Over its first three years the programme cost more than it saved, and the audit verdict was that it had not achieved value for money. That is a rare thing in this field: a quantified projection, an outturn, a set-up cost and an explicit verdict, all four in one document.

Established Seventh, the experiment that keeps being run by accident, and its result keeps being ignored. Simplifying the paperwork does not reduce administrative burden when the burden is set by the value of the prize. A national research funder cut its online application form from 180 to 68 fields and its applications from roughly 100 to 50 pages; measured preparation time per application rose from 34 working days to 38, and total researcher time from 547 to 614 working years, with the success rate unchanged at 21% in both periods. The authors attribute the result to effort tracking the expected value of the award and the level of competition rather than the length of the form. Scientific Governance Models reads this as a finding about research funding; the administrative reading is more general, and it applies to every burden-reduction programme that measures forms rather than incentives.

Established The cost-overrun literature has a real dataset and a real dispute, and both belong in any appraisal. The canonical study covers 258 transport projects worth $90 billion across 20 countries and 1927–1998: 86% overran, with mean cost escalation of 27.6% overall — rail 44.7%, fixed links 33.8%, roads 20.4%. Its author's conclusion is that underestimation “must be expected to be intentional”, best explained by strategic misrepresentation.

Frontier Two caveats travel with those numbers and are usually dropped. The author's own side of the published exchange concedes that the sample is probably not representative, naming four reasons. And the critics' substantive objection is about the baseline: measuring overrun from the decision-to-build estimate rather than from the contract price “may lead to inflated cost overruns being propagated”. Those are different questions — political accountability against contractor performance — and the literature routinely conflates them. How much of the 27.6% is an artefact of baseline choice is unresolved.

Established Note also that optimism-bias uplifts and reference-class forecasting are not the same instrument. The first is a fixed schedule of percentage adders by project type; the second is distributional, drawn from an empirical outturn distribution for the specific class. Different provenance, different mathematics, different failure modes. And the fixed schedule contains two numbers that deserve to be quoted more: for equipment and development projects the capital uplift at outline business case runs to 200% at the upper bound, so government's own guidance says the honest planning assumption is that cost may triple; and there is an explicit uplift for outsourcing, of up to 41% on operating expenditure. National guidance has encoded, since 2002, the expectation that outsourcing operating costs will be underestimated by up to two fifths, and it is almost never cited in the outsourcing debate.

Established The procurement layer beneath all of that has an anatomy, and one audit sets it out completely. The ArriveCAN record is worth reading as a specification of the parts rather than as a scandal, because each part is a normal administrative practice operating without its complement.

Established The consultancy figures are tabular for the same reason and should always be printed as a set. These are four estimates of one quantity for one year, differing because the definitions and the collection routes differ, and quoting any one of them alone is quoting one of four incompatible numbers.

Established Average variance between sources across 2017-18 to 2022-23 is £270 million a year, and a reported year-on-year fall may therefore be a reclassification rather than a reduction. Any target expressed as a percentage cut against this baseline is a target against an unmeasured quantity.

Established For contrast, here is what a fully specified implementation programme looks like, and it is instructive that the example comes from a mechanism famous for its design rather than its delivery. The United States broadcast incentive auction closed on 30 March 2017, paying $10.05 billion to 142 winning broadcasters covering 175 stations, raising $19.8 billion gross from 50 forward-auction bidders winning 2,776 of 2,912 licences, and repurposing 84 MHz of low-band spectrum. The clearing ran in a single computational process. The administrative programme that followed did not. Nearly 1,000 stations had to be repacked over 39 months in ten phases from November 2018, against a Congressional reimbursement fund of $1.75 billion that the legislative research service records stakeholders as doubting would suffice, with additional risk from a shortage of experienced broadcast tower crews and no reimbursement at all for low-power stations and translators. The elegant part cleared in one run; the part that required physically visiting a thousand transmitter sites took more than three years and was budgeted by an authority nobody could verify. That ratio is the normal shape of public administration and it is the part the design literature under-describes; Institutional Design takes the mechanism, and the implementation programme belongs here.

The nearest neighbour is Future Civil Services, and the two briefs split by unit of analysis rather than by subject. That one owns the workforce and its rules; this one owns the apparatus and its practices. They share the Canadian pay system and the Canadian border application deliberately and take different halves: the pay backlog and the contractor day rates there, the programme decisions, the procurement records and the project management here. Anyone reading one without the other will mistake a bookkeeping failure for a staffing failure or the reverse.

Megaproject Governance is adjacent in the strong sense that it owns the delivery engineering this brief borrows — the overrun distributions, the reference-class methods, the appraisal machinery. The division is that megaprojects are a class of project and this brief is about the portfolio, the procurement function and the performance regime that surround every project including the small ones.

Institutional Design is adjacent in a way worth naming precisely, because two of this brief's sources are shared with it and read differently. The Nigerian and Ghanaian management-practice study is a design finding there and a practice finding here. The spectrum incentive auction is a mechanism there and an implementation programme here. The general relationship is that design owns what was intended and this brief owns what the organisation then had to do about it.

Civic Technology inherits the legacy estate as its binding constraint and AI-Assisted Governance inherits the procurement structure, since the vendor relationships and enterprise agreements documented here are what determine which tools a government can deploy at all. Scientific Governance Models is adjacent through a single shared mechanism: the research-funder burden-reduction result in this brief's experiments section is the general form of a finding it treats as being about science.

Outside the map: project management and the economics of procurement; public financial management, which supplies the capability-trap literature; administrative law, which is where automated decision-making meets the question a royal commission answered the hard way; health services research, which supplies the targets evidence; and the education-economics literature behind the accountability natural experiment.

Established This is the best-evidenced subject in the category, and the reason is that audit institutions have been working it for decades. Almost every load-bearing figure here comes from a supreme audit institution, a parliamentary committee, a statutory inquiry, a royal commission or a peer-reviewed quasi-experiment. That is a materially stronger evidence base than any other governance brief on this map, and it is why this brief can state rates rather than anecdotes. The corollary is the selection bias named in the overview: the failures documented to audit standard are the failures in countries with functioning auditors.

Established Two government self-assessments are cited here and both are self-incriminating, which raises their weight. The evaluation review reporting that two thirds of the portfolio lacks an adequate evaluation plan is government reviewing its own evaluation practice. The value-for-money office reporting that its own control frameworks diluted accountability is the centre reviewing the centre. Self-assessment that publishes numbers against itself is worth more than self-assessment that does not, and this brief weights it accordingly.

Frontier One figure in wide circulation is a government self-review and should be labelled as such. The claim of over £45 billion a year in unrealised digital savings comes from government's own assessment of its own estate, published alongside the reform plan it justifies, with a method disclosed in a single sentence. A separate audit estimate of at least £14 billion of annual digital procurement is explicitly flagged by the auditor as coming from third parties and being incomplete. These are not alternative estimates of the same quantity and should never be presented as such. The same programme's January 2026 roadmap carries a £1.2 billion a year savings target for a new data-exchange platform on the same basis, alongside delivery figures — user counts, download counts, a 23-second median review time — that are at least the kind of number an outsider could in principle check.

Established And the reporting body assures the portfolio it reports on. Its checkable good-news number — 26 projects exiting the portfolio having delivered their objectives in 2025-26, up from 14 — is real and is a genuine improvement on its own terms. It is also self-reported by the body that assured those projects, and the same report notes that the average project rating has remained similar over four years.

Established A distinct institutional form appears in the procurement record and has no name in the rules that govern it. A two-person firm that performs no development work, receives $19.1 million and assembles subcontracted technical staff is not a supplier in the sense the competitive-tendering framework imagines, and it is not a staffing agency in the sense the employment framework imagines. It exists because a department needed capacity faster than its own hiring rules could supply it, and the contracting rules were the only instrument available. Intermediation of this kind is invisible to a procurement regime that classifies by contract type rather than by what the counterparty actually does, and nothing this brief consulted establishes how common it is.

Established Where a delivery unit sits determines what it can do, and nobody decides placement on that basis. One national digital service was moved out of a central agency into a service-delivery department of 32,000 staff in July 2023, several layers down; its structural problem was described from inside as all carrots and no sticks — power to advise, none to require. A provincial equivalent was led at deputy-minister rank. Another national unit ran on cost recovery, which gave client agencies a veto by wallet, and was dissolved by reduction in force. Mandate, rank and funding model are the three variables, and in every case examined here they were set by executive instrument or machinery-of-government reshuffle with no evaluation cited.

Frontier The institution that does not exist is a single authoritative definition of what the state buys. Four sources give four figures for one year's consultancy spend, and no body has the standing to declare which is right; the same absence is what allowed one application's cost to remain unknowable. A comparable gap sits under performance measurement across borders, where seven leading cross-national measures of state capacity correlate between 0.70 and 0.94 and load 86.91% onto a single component, and yet 45 countries diverge between two of them by more than 0.40 standardised units, with divergence largest in the middle of the distribution where most policy-relevant states sit. Convergent validity is high and interchangeability is low, which means an administration's measured capacity depends on which index its assessor happened to use.

Established What would change the institutional picture, stated as a testable condition. One jurisdiction attaching a mandatory, published post-completion outcome assessment to a defined class of projects — and a single enforced definition of consultancy spend across departments and arm's-length bodies — would within one cycle produce the two datasets this entire subject lacks. Both are administrative rules costing a rounding error against a £924 billion portfolio. That neither exists in a system that already publishes the numbers demonstrating their absence is the institutional finding this brief is actually about.

Established The sharpest ethical detail in this brief is a recommendation an audit office had to make. Following the waiting-list investigations, it recommended that the health department restrict “the use of confidentiality clauses in compromise agreements” in cases of alleged waiting-list irregularity, alongside ensuring independent external inquiries. An auditor had to recommend that a public service stop using non-disclosure agreements when settling with the people who reported that the figures were being manipulated.

Established The outsourcing record has a human ledger, and one collapse sets it out. A contractor holding about 420 public contracts failed with roughly £7 billion of liabilities and £29 million of cash, a £2.6 billion pension liability, £2 billion owed to suppliers, and a record £79 million dividend paid months before a £1,045 million writedown. Nineteen thousand United Kingdom employees; over two thousand job losses; £150 million committed by government to maintain essential services. A joint parliamentary report called it “a story of recklessness, hubris and greed”, and found that the government's own supplier-monitoring system “provided little warning of risks in a key strategic supplier”.

Frontier And the evidence for the alternative is worse, not better. The best available synthesis reports that early competitive tendering of technical services produced real savings of 20–30% that decayed over time, with weaker evidence in social services — and, on the return leg, that “the evidence is even weaker than the evidence on outsourcing”, with council-reported successes carrying “no studies to robustly assess whether these benefits were actually delivered”. Neither side of the outsourcing argument is carrying strong evidence, and the insourcing side is carrying less.

Established There is a public-money question in the procurement record that is not the usual one. The usual complaint is that too much was paid. The finding here is that the destination of the money cannot be traced: 18% of tested invoices on one application could not be assigned to a project, and a third of the reconstructed total went to an intermediary that performed no development work itself. Untraceable public spending is a distinct harm from excessive public spending, because it removes the possibility of the judgement rather than merely settling it the wrong way, and because it falls on the auditor and the legislature rather than on the department that kept the books.

Established The automated-decision case establishes who bears the error, and it should not be softened. A scheme that averaged annual income across fortnightly reporting periods produced results that a royal commission found inaccurate and non-compliant with the governing Act. The commissioner's assessment of the institution rather than the software is the part that belongs in an administration brief — how little interest there was in ensuring legality, and how rushed the implementation was. In a commercial setting an error rate of that kind is a business risk borne by the firm. In an administration it is borne by whoever the decision is about, and by construction those people are the ones least able to contest it.

Established Finally, obligations this brief takes on itself, stated as unresolved questions rather than as caveats. It cannot print an effect size for the school-accountability natural experiment, because the working paper returned an error and the journal version is paywalled, so it carries the direction only. It states the sign of the autonomy and monitoring result and no magnitude, because the source consulted reports no coefficients. It has not read the royal commission's final report, only a government agency's published overview, and it therefore carries none of the widely quoted figures on debts raised or settlement value. It has verified the cost-recovery model of a dissolved digital service unit and not that model's financial performance, so it makes no claim about whether the unit paid its way. It carries no figure at all from the two jurisdictions most often cited as models of digital public administration, because no verified source for either was obtained. And it does not carry the figure often attached to the Canadian pay rules — the number of distinct rules arising from collective agreements — because it could not be verified, which matters because that number is the usual explanation for why commercial software must be customised. Each of those is a claim this brief could have made and has not.

Established The full-cycle case is the one to hold on to, because it has every link. Probation services were contracted out in 2014; the audit office found in 2019 that the programme “has achieved poor value for money for the taxpayer”, with at least £467 million in additional payments connected with terminating contracts fourteen months early on a £2.3 billion programme; the contracts ended in December 2020 and the service was unified in June 2021. And the outcome detail is the sharpest in the brief: the number of reoffenders decreased while “the average number of reoffences they commit has increased significantly” — a reform improving the headline count while worsening the thing underneath it.

Established The recurring general principle is that the measured quantity moves and the unmeasured one absorbs the difference. Waiting times fell hugely; admission timing bunched at the thresholds. Reoffenders fell; reoffences per reoffender rose. A hospital trust hit its access targets and a public inquiry found the pursuit of them displaced care. A quarter of a target regime's own targets were never assessed at all. These are four instances of one mechanism, and the corpus's recurring question — what stands between a stated intention and a built world — is answered here by measurement design rather than by capability.

Established The second general principle is that accountability is bounded by bookkeeping, and this is the finding most likely to generalise beyond government. An auditor could not determine what an application cost because the invoices did not say what they were for. A national government cannot state what it spends on consultants because four collection routes use four definitions. Neither is a failure of oversight institutions, which functioned exactly as designed and produced the finding. They are failures of the record layer that oversight consumes, and no quantity of downstream scrutiny substitutes for it. A civilisation's capacity to hold its administration to account has a floor set by its filing.

Established And oversight machinery does have an audited track record, which cuts against the pessimism. About 53% of the areas placed on a national high-risk list since 1990 have been removed or narrowed in scope, with roughly $759 billion in attributed financial benefits over 2006–2024. That is a sustained, externally audited record of an oversight mechanism moving things — published alongside 463 of 1,881 IT recommendations still unimplemented and 32 of 69 priority ones. Both halves are the same institution's honest scorecard.

Established The terminal finding is not that the evidence is thin. It is that evaluation is discretionary and stopping is nearly impossible, and those two properties together guarantee that the evidence stays thin. A brief that says we do not know much about whether delivery reform works is understating it. The machinery is configured so that it will continue not to know, and the configuration is a set of funding rules that could be changed in a single spending round by people who already publish the numbers proving the point.

Established What already happened, because the chronology in this subject usually starts too late. A criminal-records system entered service in 1974 and a pensions system in 1988; both are still running and one of them attracted no competing bidder when its support contract was renewed in 2022. The optimism-bias schedule still in force rests on a consultancy study from 2002. Probation services were contracted out in 2014 and unified again in June 2021. Phoenix went live in 2016. A strategic supplier holding about 420 public contracts collapsed in January 2018. The five-point delivery-confidence scale was abolished in June 2021. The Auditor General of Canada reported on ArriveCAN in February 2024; an American digital service unit was dissolved by reduction in force in March 2025, three years after Canada's had been relocated out of its central agency; the projects authority merged into a successor body on 1 April 2025; the consultancy report landed in November 2025 and the pay-modernisation audit in March 2026. None of these is a pending milestone and most writing about the future of public administration references none of them.

Frontier 2026 to 2028: the marking window. The evaluation-plan share — 34% adequate against a 66% shortfall carrying £456 billion — is published annually and is the single most informative number in this subject; whether it moves is a cleaner test of reform than any savings claim. Consultancy spend is to fall by over £700 million a year by 2028-29 against a baseline the government's own auditor says cannot be measured, which is a prediction about what the marking will find. A national digital exchange is to deliver £1.2 billion a year in savings on the same self-assessed basis as the £45 billion opportunity figure that justifies it.

Frontier 2027 to 2031: the pay system, which is the scheduled test of whether an administration can learn from a documented failure of the same shape. The Dayforce planning phase is to complete by June 2027; the replacement horizon was shortened in January 2026 from 2034 to 2031; Phoenix vendor support runs to as early as 2036, with cloud extension costing at least $4 million a year in the meantime. The auditor's three flagged risks are schedule compression, an uncleared backlog importing errors, and unsimplified pay rules forcing customisation — which is to say, the three causes of the failure being replaced, stated in advance and in public.

Frontier Late 2020s: the legacy estate grows before it shrinks. Systems classified as legacy rose by 26% year on year and red-rated ones by 16%, with 28% of the worst carrying no remediation funding, in departments spending 70–85% of their technology budgets on upkeep. Nothing in the current funding structure forces a migration, and the annual reprioritisation exercise is documented as the mechanism that prevents one.

Speculative 2030s: the plausible split. Delivery evidence improves where audit institutions already work — projects, contracts, information technology — and stays absent where they do not, because that is exactly the pattern in this brief. A 1974 system that no supplier will bid against is a 2049 system on the same reasoning.

Handwave Any date beyond that. The units in this subject are contract terms, spending reviews and audit cycles; the longest reliable series here is seventy years of cost overruns, and the oldest system found in an audit review was commissioned in 1964 and is still in service. Nothing licenses a fifty-year forecast about software that has already outlived one.

Speculative The fork: either post-completion evaluation becomes a funding condition, or the audit offices keep writing the same finding. In the first case this subject acquires the outcome literature it has never had. In the second, the base rate holds — 8% of major-project spend robustly evaluated, and most top projects with no evaluation arrangements at all — and the reports of 2040 will be legible to anyone who has read the reports of 1990. The instrument is a condition of funding, which is cheap and available and has not been used.

Handwave Anything beyond that, and unusually so for this corpus. The units of this subject are contract terms and spending reviews, measured in single years, while the longest reliable series in the brief is seventy years of cost overruns that are high and constant. A field whose only durable finding is that its forecasts do not improve is a poor place to forecast from.

• Depends on Nothing on this map. Delivery capability waits on no result produced by another brief; the constraints are funding rules, contract structure, bookkeeping standards and an evaluation requirement nobody has imposed. That is an unusual sentence in a frontier-research corpus and it is the correct one here.
• Requires (not on this map) Evaluation required rather than discretionary: 150 of 227 projects carrying £456 billion had no adequate evaluation plan and 55% of those produced no evidence of any plan. A portfolio in which projects can be stopped: four of 184 tracked projects in a year, a 2.2% rate, which is continuation rather than portfolio management. Contracting records that identify their own project, since 18% of tested invoices on one application did not, and four sources give four figures for one year's consultancy spend. Knowledge-transfer obligations that operate during an engagement rather than as a handover of documents at the end, which is what the auditor found is not happening. Funded remediation for the legacy estate, absent for 28% of the highest-risk systems in departments spending 70–85% of technology budgets on upkeep. And more than one bidder for a legacy support contract — a £48 million renewal went uncontested. None is a research result: the first two and the fourth are rules, the third is a bookkeeping standard, and the last two are the consequences of not having them.
• Enables In principle, the ability to build what every other brief on this map assumes can be built. No typed enabling edge is claimed, and the reason is this brief's own standard of evidence: with two thirds of the portfolio carrying no adequate evaluation plan and no routine post-completion review, the enabling relationship is precisely the thing that has never been measured. Asserting it would assert the causal claim the brief says nobody has tested.
• Adjacent Future Civil Services, the workforce half of the same subject; Megaproject Governance, which owns the delivery engineering this brief borrows; Institutional Design, which takes the design lesson from two studies this brief reads administratively; Civic Technology, which inherits the legacy estate; AI-Assisted Governance, which inherits the procurement structure; and outside the map, project management, the economics of procurement, health services research and education economics, which supply the two natural experiments.

“The portfolio is rated red, amber/red, amber, amber/green or green.” Established That scale was replaced by a three-tier delivery-confidence assessment in June 2021, and the annual report states of the intermediate ratings that “this rating can no longer be given to projects”. Any figure quoted as “red or amber/red” after mid-2021 describes a scale that was abolished. The projects authority itself was merged into a successor body on 1 April 2025, so writing about it in the present tense is wrong as well.

“Targets are theatre.” Established Contradicted by the strongest quasi-experiment in the subject: waits over six months fell about 20% and waits of twelve months or more about 60% against a control country. “Targets worked.” Established Contradicted by the same period's hazard-rate spikes at the thresholds, six trusts in fifty adjusting figures, and a statutory inquiry finding target achievement displacing care. The sophisticated position holds all of them simultaneously, and the general form is that a high-powered measurement regime moves the measured quantity, reorganises what surrounds it, and is almost never evaluated on the second effect.

“The performance-target regime hit 79% of its targets.” Established Survivorship-filtered. The figure is computed after excluding the 90 of 366 targets — 24.6% — that could not be assessed at all, 52 for want of data and 38 because no department ever published a final assessment. The unfiltered figure is 60.4%, and the committee publishing both said the numbers were indicative only for want of independent verification. A quarter of the targets in a regime whose entire premise was measurement were never measured to a conclusion.

“England abolished school performance tables.” Established Wales did; England did not, which is why the natural experiment exists at all, and a sentence implying the reverse inverts the study design. The finding is that abolition markedly reduced school effectiveness in Wales with negligible effect on segregation — against both the reform's own objective and its opponents' expectation.

“Public projects overrun by 27.6% on average.” Frontier The figure is real and comes with two concessions its citers drop: the author's own side accepts that the sample is probably not representative, naming four reasons, and the critics' baseline objection — that measuring from the decision-to-build estimate rather than the contract price may propagate inflated overruns — is substantive and unresolved. Political accountability and contractor performance are different questions measured from different baselines.

“ArriveCAN cost $59.5 million.” Established That is the auditor's reconstruction, produced because the department's records could not answer the question: 18% of tested contractor invoices did not document which project they belonged to, and the audit's own words are that it was “unable to determine a precise cost”. Quoting the number as a known figure inverts the finding, which is about the absence of records rather than the size of a bill. The findings that survive precisely are the structural ones — $19.1 million to a two-person intermediary that did no development work, 177 releases with little or no testing documentation, and fifteen months with no governance agreement between the two responsible agencies.

“Government spends £X billion a year on consultants.” Established No such figure exists. Four sources give £1.34 billion, £1.36 billion, £1.68 billion and £2.23 billion for the same year, with an average variance of £270 million a year over six years, because definitions differ, contracts bundle consultancy with professional services and contingent labour, departments do not follow the central definition, and arm's-length bodies report to different standards. Any single figure quoted without a definition is one of four incompatible numbers, and a reported year-on-year fall may be a reclassification.

“Officials use consultants because they undervalue their own staff.” Established The opposite of what the survey in the same audit shows: 86% of surveyed officials rated consultants valuable and 40% rated them extremely valuable, which is the mechanism by which dependence persists rather than a defect in attitude. The auditor's recommendation is accordingly a default rather than a rule — that government should always start from the assumption that using its own staff will be the best use of resources — and it is a presumption placed against a population of buyers who are individually satisfied with their purchases.

“The digital service units were shut down because they failed.” Established No evaluation is cited in any closure examined here. One unit ran for eleven years on cost recovery, which required client agencies to choose to pay for it, and was terminated as “non-critical” in a reduction in force. Another was relocated from a central agency into a service-delivery department, where a co-founder's assessment is that the whole-of-government perspective and the voice at the cabinet table were what was lost, and where the unit's structural problem was described as “all carrots, no sticks”. A comparable provincial unit was led at deputy-minister rank; the federal one reported several levels below. Placement in the machinery is the variable, and it is set by reshuffle rather than by performance. Note the limit of the record: this brief has verified the cost-recovery model and has not verified the revenue that model produced, so it makes no claim about the unit's financial performance.

“Performance monitoring improves public sector delivery.” Frontier In the only large-scale hand-coded study of completion across an entire federal bureaucracy — 4,700 projects across 63 Nigerian organisations, with a companion of 3,628 projects and tasks across 31 Ghanaian ones — the incentives-and-monitoring index is robustly negatively correlated with project initiation, full completion and average completion rate, while autonomy is robustly positive. The authors flag the contrast with private-sector evidence themselves. The sign is stated here and no magnitude is, because the source consulted reports directions and robustness and no coefficients.

“Robodebt was a software failure.” Established It was an administrative decision rule that was unlawful. An Australian department raised social security debts by averaging annual tax data across fortnightly reporting periods; the royal commission found that the scheme “produced inaccurate results and did not comply with the income calculation provisions of the Social Security Act”. Debts raised solely on averaging ceased in November 2019 and a 2020 class action settlement reduced all of them to zero. The commissioner's assessment is of the administration and not the code: “It is remarkable how little interest there seems to have been in ensuring the Scheme's legality, how rushed its implementation was.” Its recommendation categories read as a specification for an administrative apparatus — mandatory legal risk assessment in policy proposals, oversight of automated decision-making, strengthened duties on government lawyers, better-resourced oversight agencies. This brief read a government agency's published overview of that report and not the report itself, and it therefore carries none of the widely quoted figures on the number of debts raised or the value of the settlement.

“Simplifying the paperwork reduces administrative burden.” Established Tested directly and refuted. A national research funder cut its application form from 180 to 68 fields and from about 100 pages to about 50; measured preparation time per application rose from 34 to 38 working days and total burden from 547 to 614 working years, with the success rate unchanged at 21%. Effort tracked the value of the prize and the level of competition, not the length of the form. Any burden-reduction programme that measures forms rather than incentives is measuring the wrong object.

“Best-practice administrative templates fail because of implementation capacity.” Established The better-supported diagnosis is that the failure is the equilibrium. Public financial management reform across Africa changed higher-level budgeting and accounting processes while, in the authors' words, “core processes determining how money was actually spent remained impervious to reform”; competitive-bidding procurement laws were among the first demands made of post-conflict Liberia, Afghanistan and Sudan and went unimplemented. The remedy those same authors propose comes, on their own statement, with no outcome evidence, and this brief does not assert that it works.