Using Data and AI to Improve Government Programme Assurance

Most programme trouble shows in the data before it reaches a board. How data and AI give government programme assurance earlier sight of delivery risk.

What you need to know in 60 seconds

Programme assurance in government still leans on point-in-time reviews, and those reviews often confirm what the delivery team already suspected. At 31 March 2026 the Government Major Projects Portfolio held 189 projects with a whole-life cost of £924.2bn, and 34 of them were rated red[1].

Data and AI can move assurance earlier. Programmes already produce the signals: milestones that keep moving, risks nobody has touched in months, spend that tracks to plan while delivery does not. Read together, they flag trouble before the next gate. NISTA now runs an AI early warning tool on portfolio data for exactly this reason[1].

What AI should not do is decide the rating. Judgement, evidence and accountability stay with people. This article covers where data helps, where it doesn't, and how Mayfair's Programme Healthcheck gives a senior leader an indicative view in about fifteen minutes.

What is programme assurance, and why does it so often arrive late?

Programme assurance, like project assurance on a smaller scale, is the independent check that a programme can deliver what it promised, for the money and time it promised, and that the people running it see the same picture as the people paying for it. In central government it follows the project delivery functional standard, GovS 002[2], and the review products the Infrastructure and Projects Authority built and NISTA now runs, from Gate 0 to Gate 5[3], which many people still call Gateway reviews.

The model is sound. The timing is the weak point. A gate review is a snapshot, booked weeks ahead, and the evidence it reads is mostly what the programme chose to write down. By the time reviewers arrive, the slipped dependency or the supplier that has gone quiet is usually known to someone on the programme. It just hasn't reached a board paper yet.

The National Audit Office traced the recurring causes of trouble on major programmes to four areas: scope, planning, managing interdependencies and oversight[4]. Three of those four leave a trail in programme data long before they show up as a red rating.

What do the 2026 portfolio figures say about delivery risk?

NISTA's annual report for 2025-26 gives the clearest current picture. Of the 189 projects on the portfolio at 31 March 2026, 29 were rated green, 109 amber and 34 red, with 17 exempt from publication[1].

Two points matter for anyone who owns assurance. First, NISTA is explicit that a red rating does not mean failure: it signals significant issues that need urgent attention[1]. Second, the weight of the portfolio sits in amber, where delivery appears feasible but significant issues need management attention[1]. Amber is where early sight pays off most, because there is still time to act on what you find.

The same report describes NISTA's Early Warning System, a digital AI tool that uses existing portfolio data to flag projects at risk of moving to red[1]. The centre of government has reached the conclusion many delivery directors reached years ago: the information needed to spot trouble usually exists. It is just scattered.

Where can data and AI make programme assurance earlier and sharper?

Most programmes already hold the raw material. The schedule, the risk and issue log, the change log, finance actuals, the resourcing plan and board minutes all carry signals. Each one alone says little. Read together, month on month, they say more than any single status report.

SignalWhat it can tell youWhere it usually sits
Milestones re-baselined more than onceThe plan is being managed, not deliveredSchedule and board papers
Risks whose score hasn't moved in monthsNobody is actively working themRisk register
Dependencies opening faster than they closeInterdependency risk is building upRAID log
Spend on plan while milestones slipMoney is buying activity, not progressFinance actuals against the plan
Key roles vacant or on short contractsCapacity is weaker than the plan assumesResourcing tracker
Green status for three periods while open issues riseReporting is more optimistic than the factsStatus reports and issue log

None of these needs advanced AI. A tidy data model and a handful of agreed rules will surface most of them. AI earns its place in two jobs: finding patterns across many programmes that no single reviewer would see, which is what NISTA's tool does with portfolio data, and reading unstructured material such as board minutes and risk descriptions at a volume no assurance team has time for.

One caution belongs alongside that. The State of Digital Government Review reported that only 8% of AI projects showed measurable benefits and only 16% showed forecast costs[5]. Assurance tools are not exempt. If an AI-assisted review can't show what it found that a person would have missed, it is a cost, not a control.

What should AI never decide in a programme review?

The rating. A delivery confidence assessment is a judgement about people, plans and evidence, and someone has to own it. The government's AI Playbook asks for meaningful human control at the right stages[6], and assurance is a clear case for it: AI can gather, sort and flag, and a named reviewer decides what it means.

Three rules keep data-led assurance honest:

How does Mayfair's AI programme healthcheck work?

We built the Programme Healthcheck, which runs on our Assure AI service, for the moment before formal assurance: when a senior responsible owner or programme director wants an honest outside view and can't wait weeks for one.

It works as a structured conversation. You talk or type for about fifteen minutes, and it takes you through eight areas of your programme, from scope and users to governance, technology and data, risks and benefits. Behind the conversation sit seven scored dimensions mapped to GovS 002, from strategy clarity and delivery realism to commercial and supplier risk. Each is scored from 0 to 5 against the evidence you gave, and the red, amber or green rating is worked out from those scores by a fixed rule. The model never chooses the overall rating.

It is anonymous by design. It never asks who you are, where you work or what the programme is called, and the only thing it collects is an email address to send the report. You see a headline view on screen when you finish, and the written report follows by email.

It is indicative, not a gate review. What it gives you is a clear view of where to look first, in words you can take into your next board. Try the Programme Healthcheck.

Common questions about data-led programme assurance

What is a Delivery Confidence Assessment?

It is NISTA's assessment of how likely a project is to deliver its objectives to time and cost, given on a red, amber or green scale. Amber means delivery looks feasible but significant issues need management attention. Red means significant issues need urgent attention, which is not the same as saying the project will fail.

Can AI replace a NISTA gate review?

No. A gate review is an independent, accountable judgement by experienced reviewers, and nothing in current government guidance hands that to a model. AI is useful before and between reviews: reading programme data, spotting patterns and pointing reviewers at the areas most likely to matter.

What data does a programme need to start data-led assurance?

Usually what it already has: the schedule, the risk and issue log, the change log, finance actuals and a resourcing view. The first job is rarely new data. It is getting those sources to agree on dates, owners and definitions, so that trends can be read across them.

Is an anonymous healthcheck useful if nobody knows which programme it is?

Yes, for the person who takes it. Anonymity is what lets a programme director describe problems frankly, and the report goes only to them. They decide what to share, and with whom.

Where should a programme team start?

Pick three signals from the table above that your programme can measure today, and track them monthly for a quarter before your next formal review. You will find out quickly whether your reporting tells the same story as your data.

If you want an outside view first, the Programme Healthcheck takes about fifteen minutes and costs nothing. For programmes that need the data work behind continuous assurance, joining up schedule, risk and finance data and building the reporting on top, Mayfair IT Consultancy's data transformation and digital transformation teams do that work across central government with SC and DV cleared people.

Related reading