Innovate UK picks firms that were already pulling ahead. Their head start explains most of the growth that follows, and their extra survival doesn't track the grant money.

1. The programme and what is already known

Three serious evaluations have looked at Innovate UK's support for business, and all three found that it worked. Why run a fourth? Because only one of them compared winners with firms that had also applied, and the international evidence says that choice decides the answer. This chapter sets out what's funded, what's been found, and where we think the UK evidence falls short.

What Innovate UK funds

Innovate UK, part of UK Research and Innovation (UKRI), gives grants to businesses for research, development and early commercialisation, mostly through competitions, each carrying its own product type. Among the firms in this study, the largest by number of first-time winners is Collaborative R&D, where firms work with partners on a shared project (6,153 of the 11,150 started there). Feasibility Studies follow (1,766), then Grant for R&D (1,056) and Studies (417). Smart grants, the open competition for any idea, sit across these and have their own 2026 evaluation. In most cases the grant covers only part of a project's cost; the firm finds the rest.

These are small firms. Very small. UKRI records 9,726 of them as micro or small (its labels vary: 8,648 "micro/small", 761 "micro", 294 "small" and a few others) and just 686 as large. Half were offered less than £92,300 across all their projects since 2016. The mean is far higher, at £371,000, pulled up by a handful of big awards (the largest was £39.2 million). One year stands out: 2,120 firms had their first project in 2020, during the Covid response, more than in any other year.

Who wins

Innovate UK's winners aren't a cross-section of British companies, and the difference shows before any money arrives. Here they are against the 1-in-100 sample of filing companies, each seen in the same calendar years and only where accounts were filed for the year before (descriptive, added after the plan; Appendix N):

That last row is the selection problem in a single line. Before any grant was paid, about half the winners were already adding staff, against one firm in five in the companies around them. Every comparison in this report is built to account for that, and the placebo test measures how far it succeeds.

What the UK evaluations found

We read each of these in full before writing a word of this report, and each figure below comes from the document itself.

There's a fourth, UKRI's evaluation of ICURe (2020), but it studies university teams rather than firms, so it sits outside our population; we read it for its method.

Look at the comparison column. Frontier and the Enterprise Research Centre set winners against firms that never applied; only the Smart evaluation used firms that applied and lost. And the growth it found over the evaluation period (4.3% more employment) is below the 5.8–6.0% the 2018 study reported over three years. Different programmes and years, admittedly. Still, we don't think the gap is a coincidence, and Chapter 7 shows why.

What the strongest designs abroad found

The most convincing evidence on R&D grants comes from funders that rank applicants and fund down to a line. Firms just above the line and just below it were judged almost equally promising, so the difference in what happens to them next is close to the grant's own effect. That's called a regression discontinuity design. Why does it matter here? Because it's the one design in which a funder's skill at picking winners can't pass itself off as the effect of its money. Four such studies matter here, and we've checked each against its own text.

Read together, they tell a consistent story. R&D support can work. But for whom? When it does, it works mostly by giving money to firms that can't easily raise it, and it does little for firms that could have funded the work themselves. That gives us two tests to apply to Innovate UK, and both can be run on public accounts: if its grants work the same way, effects should rise with the money, and they should be largest for the firms that were shortest of cash. Chapters 4 and 7 run both.

What this study adds

A fair reviewer might ask what a desk study of public data can add to work done inside the ONS secure lab. Four things, we think:

• Every Innovate UK product type, from 2016 to 2026, rather than one competition.
• Public data and published analysis. Studies run in the secure lab can't be rerun outside it. This one can: we publish the company-level data and the analysis code, so anyone can rerun the analysis behind our findings, and every value names the filing it came from. Only the first profit split, built straight from the filings, can't be rerun from what's published (Appendix B).
• The balance sheet and survival. Filed accounts show net assets, fixed assets, cash and whether debts exceed assets. In other words, they tell us whether firms got stronger as well as bigger.
• A formal placebo test of the pre-grant trend for every result, reported whichever way it points.

What it can't add is a better comparison than UKRI's own Smart evaluation. We'd rather say so at the start than have a reviewer say it for us.

2. How the grants were meant to work, link by link

A grant is meant to set off a chain: a project gets done, the firm builds capacity, it comes out stronger, and then it survives and grows. Which links hold, on the evidence of the firms' own accounts? Our answer, link by link, is below. The short version: the first link holds, the middle ones break, and the last one shows up without the links that should have led to it.

The first link holds

The money goes where it's meant to. Almost nine in ten project places are closed, very few were withdrawn, and closed projects spent nearly all of what they were offered. Whatever else this report says, it isn't a story of projects abandoned or money left unspent.

The middle links break

If projects built capacity, funded firms should end up with more staff and more equipment than firms like them, beyond anything they were heading for already. They do have more of both. Yet the gap in each had opened before the money arrived, at about the size of the later effect, which means the accounts can't credit any of it to the grant however the comparison is cut. Then comes the link that matters most to a lender or an investor. Did firms come out stronger? Not that the accounts can show. Here the chain doesn't just fail to show. It points the other way among survivors, with more funded firms owing more than they own. In our view that's the clearest sign the programme doesn't work the way its logic says.

The last link appears from nowhere

Funded firms do survive more, by a wide margin. But survival is meant to come from a stronger firm, and the firms aren't stronger. Nor does survival rise with the size of the grant, or with how short of money the firm was before it (Chapter 4). Whatever keeps funded firms open, in other words, it isn't the chain in the table. The likeliest explanation is the oldest one: the firms that apply for and win R&D competitions are going concerns with plans, which firms that merely resemble them in their accounts often aren't.

What can break the chain

Four things can. Selection is the big one, and it's what this study finds: stronger or faster-growing firms apply, and win, and the placebo test measures how much. Substitution, where a firm would have run the project anyway, can't be seen in accounts, though selection on this scale makes it likely. Time lags beyond five years are out of reach for most firms so far. And Covid hit the 2020 and 2021 cohorts hard, which is why they're tested separately in Appendix E.

In short, a theory of change that reads well on paper doesn't survive contact with the accounts. That's the puzzle the rest of this report works through.

3. How the study was done

Here's the method in brief. The full specification, the matching diagnostics and every result table sit in the appendices.

Data

• Who was funded: UKRI, "Innovate UK funded projects, 2016/17 to present" (September 2026). 39,175 participant rows carry a company number, covering 15,782 companies. We used the award offered, project dates, product type, competition, region, size and the withdrawn flag.
• Outcomes: Companies House bulk accounts data (inline XBRL), every monthly archive from 2014 to September 2026, read by company number. Each value is kept with the filing it came from. A year's value comes from that year's own accounts; the next year's comparative figure only fills a gap. Where both exist, they agree for 92% of net-assets and 94% of employee firm-years.
• Firms that closed: status, closure date, sector and incorporation date come from the Companies House register, or from its company profiles for closed firms, read on 11–12 October 2026. A closed firm with no closure date is dated to the year after its last accounts.
• Data protection: we use company-level figures only. Directors' and other people's personal details in the filings aren't read or kept.

Who is compared with whom

Funded firms are businesses of every size (universities, public bodies, Catapults and charities left out) whose first Innovate UK project started in April 2016 or later, with no project in the 2004–16 records, an award above zero and no withdrawal. A firm's grant year is the start year of its first project.

Comparison firms come from a 1-in-100 sample of every company that filed accounts in any year from 2013 to 2026, drawn by a fixed hash of the company number (71,515 firms). None appears in either grant file, and 45% have since closed. Why does that matter? Had we drawn them from today's register, we'd have sampled only survivors, and every firm that later failed would have vanished from the comparison.

Three populations are reported separately and never pooled:

• Established firms with six years of accounts before the grant (median age 10): the main net-assets design, which also removes each firm's own earlier trend (2,300 funded firms matched).
• Firms with three years of accounts before the grant: the plain matched design (4,818 for net assets; 3,614 for employees, which small companies only began to disclose for periods from 2016).
• Young firms without that history (median age 3 at the grant), matched on incorporation year, sector, region and size in their first accounts. It's a weaker design, reported in Chapter 7.

Matching and the comparison

Within each grant year and two-digit sector (or the wider sector section, where fewer than 25 comparison firms were available), each funded firm is matched to up to five comparison firms with the closest propensity score. That score comes from the three years of accounts before the grant, their level and their slope, together with the firm's age. Balance is checked for every cohort, and where any standardised difference exceeds 0.1 the matching is tightened, with both results reported (Appendix C). The effect is the difference in change from the year before the grant (t−1) between funded firms and their matches, traced from six years before the grant to five after, with year three (t+3) as the headline. Net assets are in pounds and employees in heads, each trimmed at the 1st and 99th percentiles every year. Intervals come from 500 bootstrap draws of funded firms with their matches. Crucially, we chose this estimator on simulated firms with a known answer, before any real outcome was seen. Two designs failed that test.

The placebo test, and how results are combined

Matching makes funded and comparison firms alike in the years it uses. It can't promise they'd have stayed alike. So for each result we move the grant date two years earlier, match on the years before that, and measure the gap at the real t−1, a year before any money changes hands. A method that works gives a gap near zero. A gap means the funded firms were already pulling away, and part of any later effect belongs to them rather than to the grant.

The plan also fixed, before any result was seen, one rule for combining what we found. Each result is shown for every comparison we can build, with its Maryland level and its own placebo. The more an effect is matched by a pre-grant gap of the same size, the less of it we credit to the grant; where comparisons disagree, the claim shrinks to what they share. Findings are then worded shown (a level 3 design or better, with a clean placebo), suggested (consistent, but not testable this way) or not established.

Power

For each headline we report the minimum detectable effect: the smallest true effect the design would catch four times in five at the 5% level (2.8 times the bootstrap standard error). It matters most when nothing shows up. A null result then reads "no effect larger than this", which isn't the same as "no effect".

How the plan was fixed

Version 1.0 of the plan was fixed and fingerprinted at 11:52 UTC on 12 October 2026, before any funded-versus-comparison outcome had been computed. It wasn't published first, so a reader has only our record that it came first, and we'd rather say so than leave it implied. Its first results, failed placebo included, were then seen. The plan was reviewed against our reporting standard and against UKRI's own evidence, and it changed eight times that day (versions 1.1 to 2.4); the last two, 2.3 and 2.4, only record analyses added after the results, marked as such. Every change made after results is marked, with its reason, in Appendix M. Nothing is hidden. Every version's exact text is archived beside its fingerprint, and is published with the code. Nothing in version 1.0 was removed without a line saying so. Our later studies register the plan independently, under embargo, before any outcome is seen.

4. Survival

Question 2, first part. Were funded firms more likely to still be open three and five years after their grant year? They were, by a wide margin. The harder question is why, and our answer is less flattering to the programme than the headline. We start here because survival changes how every growth result must be read: the growth comparison can only see firms that are still filing accounts.

The result

Of all the results in this report, this is the one that most resembles what came before. Frontier Economics found 11 and 16 points for Innovate UK support in 2008–12, and its comparison firms survived at 84% and 75%, almost exactly the rates of ours. Different data, a different decade, a different team, and still almost the same answer.

Both groups count only firms still open at the end of the grant year. That rule matters more than it looks. Funded firms have to exist to receive their grant, whereas comparison firms don't, so counting from the start of the year would hand the funded side a head start that has nothing to do with the grant and everything to do with the calendar. Our first run made exactly that mistake and overstated the three-year gap (+14.9 and +11.5 points). The definition was corrected before anything further was read, and both figures are shown in Appendix G.

Is it the grant, or who wins?

Here the placebo test can't help. Every firm in the comparison was open before the grant, by construction, so there's no earlier gap to measure, and the result rests on matching alone. A sceptical reviewer would ask straight away what else could produce a gap this size. It was our first question too. If it's the grant money keeping firms going, as the strongest studies abroad suggest grants can (Chapter 1), three things should be true. Each was tested after the plan's results were in, and all three are reported as exploration in Appendix N.

The employee sample tells the same story, with one hint the other way: there, firms in the lowest cash third gained 3.9 points more than those in the highest, an interval that just reaches zero. It's not enough to rest a mechanism on.

Put simply, the survival gap doesn't behave like an effect of the money. A gap that ignores the size of the grant, ignores the firm's finances and keeps widening years later looks like a difference in the firms themselves. Firms that bid for R&D competitions, and win them, are trading, planning and backed by people who expect them to last; some of their matches, with similar accounts, were already winding down. That's our reading, and we hold it firmly. None of this says the grants have no effect on survival. The EU evidence finds a real one at the funding line, and some of our gap is probably that. Our point is that the 12 points mostly measure Innovate UK's choice of firms, and that the earlier 11-point finding for 2008–12 likely does too.

One suspect can be ruled out. 4.4% of matched comparison firms in the net-assets design now file dormant accounts (2.3% in the employee design), against 0.7% and 0.4% of funded firms. You might expect these dormant shells to inflate the gap. They do the opposite. Dormant companies are still registered, so they count as open, which makes the comparison look more likely to survive and shrinks the gap (Appendix K).

Survival also has a knock-on effect for everything that follows. When funded firms survive more, the growth comparisons in Chapter 5 set a larger share of funded firms against a smaller, surviving share of their matches. Every growth headline therefore comes with two bounds: one counting closed firms at zero from their closure year, and one trimming the extra funded survivors (Lee bounds). Is that enough to rescue the growth results? Chapter 5 answers that.

5. Growth: net assets and employees

Question 1. Did funded firms' net assets and employees change more than those of comparable firms that weren't funded, in the five years after the grant? On the face of it, yes. Our finding is that the face of it is misleading: the winners were already on that path before the grant, and the accounts can't credit the grant with any of the gain.

Net assets

Net assets, total assets less all liabilities, are filed by almost every company, small ones included, for every year since 2014. That makes them the most complete measure we've got, and the one an investor or a lender would look at first when asking whether a firm had come out of a project stronger or weaker than it went in. The main design (B) uses established firms with six years of accounts before the grant: it matches on the three years before the grant, then removes each firm's own trend, measured over the three years before that. Why go to that trouble? In the plain matched design (A), funded firms' net assets were already climbing far faster than their matches' right through the pre-grant years, from a gap of £168,400 six years out to zero at t−1. Design B exists to strip that out.

Of the 2,300 funded firms matched, 1,038 are observed three years after the grant. Firms first funded from 2023 onwards haven't filed that year's accounts yet, and a few others have stopped filing. Results at four and five years rest on fewer, older cohorts still, most of them funded before 2020, so they describe the programme as it was rather than as it is now.

Three years after the grant, funded firms' net assets had risen £35,600 more than their matches' (95% −£30,000 to +£100,600). It's a modest figure, and the typical firm tells a quieter story still: the median funded firm sat just £5,500 above its matches, and 52% of funded firms were above theirs, barely more than a coin toss. On the inverse hyperbolic sine scale, which weighs every firm more equally, the estimate even turns negative (−1.31, interval −2.83 to +0.23). By four and five years the levels estimate grows (£79,500 and £111,200), though on fewer, older cohorts.

Then the placebo. With the grant date moved two years earlier, funded firms had gained £57,200 more than their matches by the real t−1 (95% £21,600 to £89,100), before any grant was paid. That's larger than the effect at three years. Under the plan's combining rule, none of the net-assets gain can be credited to the grant.

A reviewer would rightly push on this: isn't the placebo built on different firms? It is, partly. It needs a longer history, so it covers 1,380 of the 2,300. On exactly those firms, though, the effect at three years is smaller still, £16,500 (−£71,700 to +£100,600), against their own pre-grant gap of £57,200. Like for like, the conclusion only hardens.

The later-winners comparison (C) points the same way as B for the cohorts both cover, with a larger estimate. Unfortunately its own placebo could only be built for 36 firms, and their matches were badly balanced, so it adds no weight either way. All six robustness checks leave the estimate inside an interval that includes zero (Appendix E). Among firms with a single grant, it's essentially nil (−£1,900).

Employees

Average employees in the year come from the notes to the accounts. Small companies only had to disclose them for periods from 2016, so the employee design needs three years of disclosed figures before the grant, which in practice means grant years from about 2019. It includes directors, too. Here the plain matched design (A) has to be the main one, because only 112 firms had the six years of disclosures design B needs; its estimate for them is in Appendix D.

It's a striking pattern. Funded firms had 0.77 more employees than their matches in the grant year itself, 1.66 a year later, 2.06 at two years and 1.98 at three (95% +1.35 to +2.62). At the median, a funded firm was 0.75 of an employee above its matches, and 58% of funded firms were above theirs. On the log scale the gap is 0.26, about 29% more staff, a large share for firms this small. That jump in the grant year is exactly what hiring for a funded project would look like.

Still, we'd been warned. The plain design is the one our simulation flagged before any real data were seen: when winners are picked partly on growth the accounts don't show, it reported 0.48 employees at three years where the true effect was zero (Appendix L). Our trend-adjusted design corrects for that, but it needs six years of disclosed employees, and few firms have them.

And the placebo fails here too. With the grant moved two years earlier, funded firms had added 1.58 more employees than their matches by the real t−1 (95% +1.21 to +2.07). In other words, they were pulling ahead before the grant by an amount close to the effect itself. On exactly the 1,767 firms the placebo could match, the effect at three years is 1.39, smaller than their pre-grant gap of 1.58. Could some of that run-up be firms hiring in expectation of a grant they'd applied for? Some of it, perhaps, but not most. Splitting the placebo by year, an exploratory check we added after the results (Appendix N), about half the gap (0.84 of 1.58 employees) had already opened two years before the grant, too early for hiring after an application, unless firms commonly applied more than a year before their projects began. Net assets point the same way: 71% of their pre-grant gap was there by then, and hiring in hope costs money, so it should pull net assets down, not up. Even that first year's pace, if it simply carried on, would cover the whole effect at three years. What the public file still can't do, without application or decision dates, is pin down how long firms waited (Recommendation 3).

One technical point a reviewer will look for. Comparison firms can be matched to more than one funded firm, and three in four employee matches use a firm that serves elsewhere too. Allowing for that widens the intervals a little, to 1.17 to 2.80 (net assets: −£32,300 to +£107,600), and changes no conclusion. It also holds across the robustness checks, from 1.25 without the Covid-era cohorts to 1.98 in the main design. That steadiness is less reassuring than it sounds, mind you: every check shares the same comparison, so every check inherits the same pre-grant gap. Later winners give a similar figure with a wide interval, and their placebo could be built for only 4 firms.

Fixed assets and capital employed

If projects build capacity, it should show up in what firms own. Fixed assets (design B) rose by £88,200 more in funded firms at three years (95% £38,100 to £137,500), against a pre-grant gap of £48,900. Total assets less current liabilities rose by £125,400, against a pre-grant gap of £69,100. Both effects are bigger than the gaps that came before them, though their intervals overlap, and neither passes its own test: each placebo is well clear of zero, so the firms were already building assets before the money came (Appendix I).

What the growth results say

Put together, the picture is consistent, and it isn't the one a quick look would give. Funded firms grew more than firms that resembled them in their accounts: in staff, in fixed assets and, by four and five years, in net assets. Yet every one of those gaps had begun to open before the grant, at roughly the size of the later effect. Innovate UK picks firms that are already on a faster path. A competition designed to back promising firms ought to do that, of course. But it means a comparison with never-funded firms can't show how much of their growth the grant caused, and any evaluation that presents such a comparison as impact, without a pre-grant test, is very likely presenting selection. One more number makes the point: an unmeasured difference that made a firm 1.5 times as likely to win, and to grow, would be enough to explain the employee result (E-value 1.54; for net assets just 1.20).

We don't conclude that the grants cause no growth. We conclude that, on the public record, there's no growth left over once the winners' head start is counted, and that the programme's growth case therefore rests on evidence nobody outside UKRI can test.

6. Financial resilience: cash and negative net assets

Question 2, second part. Did funded firms end up with more cash, and were fewer or more of them left owing more than they own? Here, for once, a result passes its own test. It still doesn't mean what it first seems to, and we'll say plainly what we think it does mean.

Negative net assets

A company has negative net assets when what it owes exceeds everything it owns. For an established trading company, that's a warning sign. For a young, equity-backed technology firm it can be perfectly normal for a while, especially where founders or investors lend money that will later convert into shares, and an accountant reading such a balance sheet would look first at who the creditors are. Either way, it's the opposite of the strength the theory of change expects.

This is the only balance-sheet or growth result in the study that shows a difference and passes its own placebo. Before the grant, funded firms were no more likely than their matches to have negative net assets. Three years after it, among firms still filing, 2.8 points more of them did. Taken at face value, grants would seem to leave firms weaker. A careful reviewer would object straight away that this compares only the firms still standing. They'd be right to.

Is it a survival effect?

Mostly, yes. The comparison only sees firms still filing accounts, and funded firms survive more (Chapter 4), so a weak comparison firm is likely to close and drop out of the sample while a weak funded firm is more likely to stay, and to stay with negative net assets on its books. To test this, we counted closed firms back in, among firms open at the end of their grant year, in two ways:

Most firms that close have run out of money, so the first of these is nearer the truth than the second. On that reading, funded firms are less often in distress, and the higher share among survivors comes from more weak funded firms staying open. We should be candid about one thing. The placebo for that row isn't clean, and it can't be: funded firms had to be open to receive their grant, so in the shifted placebo years only comparison firms are able to close, which tilts the placebo before anything is measured. We describe the result as it stands and don't claim it as a grant effect.

What keeps those weak funded firms open, then? Not the grant money itself, on Chapter 4's tests, because survival doesn't track it. Our reading is that many funded firms are R&D-heavy companies that make losses by design, charge their research costs as they go, and are carried by founders and investors who expect a payoff later. Such firms show negative net assets and survive anyway; the firms they're matched with, on similar accounts, have nobody to carry them. Splitting net assets into profit kept and capital put in tests this directly (Chapter 8), and the split fits it. After the grant, funded firms kept less profit than their matches, about £222,000 less by three years among firms whose every balance sheet carries the figure, while the extra capital investors put in was mostly there before the grant. Among young firms the pattern is starker still (Chapter 7). That's a test added after the results, so it supports our reading; it doesn't prove it.

Cash

Funded firms also held more cash at the year end than their matches: £113,500 more at three years in design A, and £93,300 in design B. Both exceed their pre-grant gaps (£34,900 and £54,400), so some of the cash gain does arrive after the grant. What is that cash, though? It can include grant payments and partners' contributions not yet spent, or equity raised on the back of the grant, and a year-end balance can't tell us which. With net assets flat and more survivors in negative net assets, more cash doesn't add up to stronger firms. We report it as not established as a gain in strength.

7. For whom: grant size, programme type, Smart grants, young firms

Question 3. If grants cause growth, bigger grants should cause more of it, and the effect should vary in sensible ways across programmes. Does it? Our answer is no: the variation follows the firms Innovate UK chooses, not the money it gives them. This chapter also asks whether public data can reproduce UKRI's own Smart grants estimate, and what happens to young firms, which the main designs can't cover.

Grant size

A dose-response test is one of the evaluator's best defences against selection: if larger grants, relative to the firm, bring larger effects, selection alone becomes a harder story to tell. Ipsos leaned on exactly this test in its evaluation of the Employment Allowance. There's a catch, however, and the plan anticipated it. A rising pattern only argues against selection if the pre-grant gap doesn't rise with it, so we ran the placebo within each third of the grant range too.

The thirds are cut at a first-year grant of £48,500 and £99,400. Our first run split firms by every grant they were offered across the whole file, including later grants, which fast-growing firms are more likely to win; the table uses the first year's grant instead (Appendix I). Both versions show the same thing (0.33, 1.64, 3.28 in the first run, with a placebo of up to 2.97). In the largest third, most of the effect had opened before the grant. Net assets, meanwhile, show no rising pattern at all (Appendix I), and survival doesn't either (Chapter 4).

Only the middle third offers a reader anywhere to look for a grant effect, with 1.93 after against 1.01 before. One band out of three, though, is a thin thread, and we wouldn't hang a programme's case on it. What the table shows clearly is how assessors allocate money: the largest grants go to the firms with the strongest run-up.

Programme type

Grant for R&D is the most telling row in the report. It's the one product type whose winners weren't already pulling ahead, and it's the one that shows no employee effect at all. It also proves the placebo isn't rigged to fail: where funded firms weren't ahead before their grant, the test says so plainly. Put the two halves together and the lesson is uncomfortable. Where Innovate UK didn't pick growers, we see no growth in jobs. For Collaborative R&D and the other types, the employee effect does exceed its placebo, by about one employee in Collaborative R&D and 1.6 in the rest, yet the intervals overlap. No product type shows a net-assets effect.

Smart grants: can public data reproduce UKRI's estimate?

This is the test we most wanted to run. UKRI's 2026 evaluation, by Ipsos with Technopolis and George Barrett, found Smart grants raised employment by 4.3%, comparing winners with high-scoring applicants who were turned down, on ONS data. We applied our design to the 398 firms whose first grant was a Smart grant between 2019 and 2024.

Against never-funded firms, Smart winners appear to grow far faster than UKRI's estimate allows, but their gap before the grant is larger still. The public data, in short, can't reproduce the 4.3%. Why not? Winners of an open competition judged on promise differ sharply from firms that merely look like them in their accounts, whereas high-scoring declined applicants share that promise. Ipsos compared 1,096 winners with 5,849 such applicants. The remaining differences are smaller: Ipsos measured employment from the Business Structure Database (a yearly snapshot of the ONS business register) where we use average employees in filed accounts, directors included, and its estimator lets effects differ by cohort. Smart firms' net assets, for their part, show nothing either way (−£50,800 at three years, interval −£242,500 to +£138,600).

Is a failed replication bad news for UKRI? We'd argue the reverse. It shows that UKRI's choice of comparison was the right one, and by how much a cheaper, never-funded comparison would have overstated the effect. It also sets a benchmark for everything else: a 4.3% employment effect, measured against firms that applied and lost, is the most credible growth figure Innovate UK has, and it's a long way below what a never-funded comparison suggests.

Young firms

Most funded firms are young; the typical firm outside the main designs was three years old at the grant. Young firms lack the accounts history the main designs need, so they're matched on incorporation year, sector, region and size in their first accounts, then compared from their first filing. With no pre-grant years to test, this is the weakest design in the report (Maryland level 2).

• Net assets rose £357,900 more from first filing to three years after the grant (95% £259,900 to £489,600; 2,916 firms); £330,000 with closed firms counted at zero.
• Employees rose 3.27 more (2.93 to 3.62; 2,286 firms).
• Survival was 20.4 points higher at three years and 24.0 points at five.

These are the largest differences in the study, and the least protected against selection. A young firm that wins a competitive grant has been judged promising by expert assessors, while its matches have only been judged similar in size and age. Then there's the money. A rise in a young firm's net assets can come from new shares sold to investors rather than from profit, and UKRI's Smart evaluation records £3.6bn of equity raised by funded firms. We'd treat these figures as a description of the firms Innovate UK backs, and nothing more. Splitting it shows where the gain comes from. It's money put in by investors. Funded young firms' capital put in rose £1.44m more than their matches' from first filing to three years after the grant, while the profit they kept fell £1.09m further. On the firms that carry retained earnings from their first filing, net assets rose £345,000 more. Not every filing carries every line, so the three figures cover slightly different firms and don't add up exactly. They're untrimmed averages, as in the rest of this section, so a few large fundraises weigh heavily in them. The young firms Innovate UK backs, in other words, spend heavily and are funded by investors to do it. That's a description of who they are; this design has no pre-grant test, so it can't say what the grant added (exploratory, Appendix N).

Sector and region

Results by sector and by region are in Appendix J, shown as description rather than tested one by one. In three regions the employee effect is clear of zero while the gap before the grant isn't: the North East (3.41, on 32 firms at three years; gap 0.29), the East of England (2.55, on 109; gap 1.02) and the East Midlands (2.25, on 65; gap 0.81). Tempting as it is to read something into that, these are small groups among 21 splits, and with that many some will look clean by chance. We draw no conclusion from them. Most net-assets estimates have intervals that include zero.

The question this chapter opened with, then, gets an awkward answer. Effects do vary, by grant size above all, but the variation follows the firms Innovate UK chooses as closely as it follows the money.

8. Turnover and profit: what the public record can show

Question 4. Sales and profit are what most people would ask about first. We read every set of accounts filed electronically at Companies House since 2013, so why can't this report say whether funded firms sold more or earned more? Because, for most small companies, the law lets them leave those figures out.

Why the figures are missing

Small companies may choose not to file their profit and loss account at all (so-called filleted accounts), and most do. It isn't a gap in our collection. Only 708 of the 11,150 funded firms (6.3%) filed a turnover figure in the electronic accounts data in any year from the year before the grant to three years after. You might expect large firms to fill the gap. They don't: among firms UKRI records as large, the share is only 8%, because many larger companies' full accounts reach Companies House as documents rather than tagged data. In the comparison sample, 5.3% ever filed turnover.

Paper filings are a smaller problem than they sound. A year missing between two filed years, the trace a paper or image filing leaves, accounts for 4.7% of funded firm-years and 3.1% of comparison firm-years. Both groups lose about the same, so it doesn't tilt the comparison.

Profit, read from the balance sheet

There's a way round the missing profit and loss account, and we think it's the most useful addition a public-data study can make. Almost every balance sheet, filleted or not, shows the firm's retained earnings: the running total of profits kept in the business. Its change from one year to the next is the year's profit after tax, less any dividends paid. That separates two things net assets mix together: money the firm earned, and money investors put in. It's a cleaner test. Our first reading of the filings skipped the figure, because most filers tag it as one part of an equity breakdown, and we've corrected that for every filing. So what does it show?

After the grant, funded firms kept less profit than their matches. On the trend-adjusted design, their retained earnings were £107,000 lower three years on, though that interval still allows no difference at all. Could uneven coverage be doing this? Retained earnings sit on fewer comparison filings than funded ones, 42% against 64%, so we reran both designs on only the firms whose every balance sheet carries the figure. The gap held, and firmed: £222,000 lower. Those are firms filing fuller accounts, 42% of funded firms and 28% of comparison firms, so the figure describes them rather than every firm. Before the grant, the placebo points the other way (£73,000 higher, an interval that includes no difference), so this isn't a slide that was already under way.

Money put in by investors tells the opposite story, and a weaker one. Funded firms raised more, but most of that gap was there before the grant: on the same firms, the placebo already shows £66,000 of the £78,000 effect. We read it this way: after the grant, funded firms spend more than they earn, which is what research looks like on a balance sheet, and their investors were backing them before Innovate UK did. A lower retained-earnings figure can also mean bigger dividends, which the balance sheet can't tell apart from losses, and it can't tell us either whether the extra spending is the grant's project or just the kind of firm Innovate UK picks. Both checks were added after the plan's results had been seen, and they're reported in full in Appendix N.

Turnover is another matter. A turnover estimate built on so few, and such unusual, firms wouldn't describe the programme, so we built none. UKRI's own evaluations take turnover from the ONS business register, which covers every VAT- or PAYE-registered firm. For this question that's the right source, and we'd point any reader there.

9. Additionality and value for money

How much of what happened to funded firms would have happened anyway? Most of it, on our evidence. And what does the balance sheet show for each pound of grant? That the money was spent, which is all filed accounts can say.

Deadweight, displacement and substitution

Deadweight is growth that would have happened without the grant, and the placebo tests are this study's measure of it. For employees, the gap before the grant (1.58) is about four-fifths of the effect at three years (1.98). For net assets, it's larger than the effect. Can we turn those into a deadweight percentage? Not precisely, because it depends on what the run-up would have done next. If it stopped dead the moment the grant arrived, all of the 1.98 would be the grant's; if it carried on at its pre-grant pace for the four years to t+3, it would more than account for the whole gap. Which is likelier? Firms with two years of faster growth behind them don't usually stop the day a grant arrives, so our judgement is that deadweight is high: most, and possibly all, of the measured gain.

Displacement, where funded firms take staff, customers or market share from other firms, can't be measured from firm accounts. Neither can substitution, where a firm uses the grant for a project it would have funded itself. Both would pull the programme's net effect below anything estimated here. With winners this far ahead before the grant, we'd expect substitution to be common: firms already growing fast are exactly the ones that could have found the money elsewhere, which is what the evidence abroad finds for firms that aren't short of finance.

Value for money on the balance sheet

Start with how a grant reaches the accounts. Grant income is recorded as income, so the grant's own money passes into retained earnings, and from there into net assets, unless it's spent. A rise of £1 in net assets for each £1 of grant would therefore mean the firm was no stronger than before; it would simply be holding the grant. Below £1, the money was spent, and for an R&D grant that's the whole purpose, so the ratio measures what's left on the balance sheet. It doesn't measure benefit. A Treasury reader would object that this isn't value for money in the Green Book sense at all, and we agree; we report it because it answers a narrower question that the accounts can answer, namely whether the money stayed in the firm or was spent.

At the top of its interval, funded firms keep about half of each pound of grant as extra net assets by year three; at the central estimate, they keep 17p. And the net-assets effect itself isn't established (Chapter 5). Retained earnings tell the same story from the other side (Chapter 8). Grant income passes through them, yet funded firms ended up keeping less profit than their matches, so the money was spent rather than kept. The grant's real value has to be judged on what the money bought, which filed accounts simply don't record.

What does the ratio leave out? R&D carried out, new products and processes, skills, knowledge passed to other firms and universities, benefits to customers, and wider effects on the economy. It's not comparable with the £1.27 of productivity gains per £1 of social cost in UKRI's Smart evaluation, which measures productivity from ONS data for one programme. Nor do we give a cost per job. Dividing grant money by an employee effect that isn't established would produce a figure with no meaning, however precise it looked.

Where, then, does that leave value for money? In our view, the honest answer for Innovate UK as a whole is that it hasn't been shown either way, and can't be from public data. The £1.27 for Smart is the best figure there is, and it covers one competition.

10. What this study cannot tell us

Every study has edges. Which of ours matter most, and which would a reviewer push hardest on? We've listed them all here, each also placed in the chapter where it bites, and said which of them could change our conclusions.

• Differences the accounts can't see. Ambition, management, the quality of an idea: the placebo tests show these matter, and the E-values (1.20 for net assets, 1.54 for employees) show that quite modest unmeasured differences could explain the results. This cuts one way only. It makes the growth effects smaller, so it strengthens our conclusion rather than threatening it.
• Turnover and profit. Small companies may leave the profit and loss account out of their filed accounts, and most do (Chapter 8), so turnover is assessed only from coverage, and profit only through retained earnings on the balance sheet. Years filed on paper or as image documents are missing from the electronic data, for funded and comparison firms alike.
• Firms without a company number. 19% of participant rows in the UKRI file (universities, public bodies, sole traders, partnerships) are out of scope.
• Employees. Small companies weren't required to disclose average employees before 2016 periods, so the employee design covers later grant years only, and the figure they report includes directors.
• Time. Accounts are complete to 2025, so effects five years on rest on cohorts up to 2020, and nothing here reaches further than five years, which is short for research that may take a decade to reach the market. This is the limit that could most plausibly change our view: a grant that pays off after year five would be invisible here.
• Covid. 2020's cohort is the biggest by far. Leaving out 2020–21 lowers the employee estimate to 1.25 and raises the net-assets estimate (Appendix E).
• Firm-level only. No spillovers to other firms, no effects on the wider economy, and nothing about the people employed. The UK tax-relief evidence finds spillovers to nearby firms (Chapter 1), so a firm-level study can understate the total.
• No interviews or case studies. This is a desk study of public data, so it can't say how firms used their grants, why some grew and others didn't, or what the projects produced. The evaluations we set ourselves against, by SQW and Technopolis among others, interview firms and write case studies; ours doesn't, and our own reporting standard asks for mixed methods where the questions need them. A reader should weigh our findings on mechanism, in Chapters 4 and 6, with that in mind.
• Reused comparison firms. Matching lets one comparison firm serve several funded firms (up to 44 times). Our main intervals resample funded firms only; intervals that also resample comparison firms are slightly wider, and they're reported beside them (Appendix E).
• Many comparisons, and some added late. We report dozens of splits and checks. Each carries its own interval and placebo, and none is singled out after the fact. The survival tests in Chapter 4 were added after the plan's results were seen, and they're kept apart in Appendix N for that reason.
• No declined applicants. Until UKRI releases them, our comparison is weaker than the one its own Smart evaluation could use.

That last limit matters most. It's also the only one on this list that a single decision could remove, and the decision belongs to UKRI, which already holds the declined applicants, already links them to ONS data for its own evaluators, and would need to change nothing about how it runs its competitions to let others do the same.

11. Recommendations

What should change? We've six recommendations, each addressed to the body that can act on it, each saying what's wrong now, and each with its cost stated plainly. The first two matter most. Together they'd let anyone answer the question this report can't.

1. UKRI: open the declined applicants to independent researchers. A comparison that separates a grant's effect from Innovate UK's choice of firms already exists, but only UKRI's commissioned evaluators can use it: its own Smart evaluation linked 5,849 high-scoring declined applicants to ONS data in the Secure Research Service. Make that list (company numbers, competition, score) a standing dataset there, open to any accredited researcher, for every competition since 2016. The first step costs nothing: UKRI's privacy notice for applicants (updated 11 August 2026) names, for evaluation, only third parties UKRI contracts itself, so adding accredited researchers working in a secure setting, on declined applications as well as funded ones, would tell applicants plainly from the start. What it costs: won't applicants object that their bids were meant to stay confidential? A secure setting protects that, though each study then needs accreditation and secure-lab time. That's a small price for evidence on £4 billion.
2. Innovate UK: measure at the funding line, and break ties there by lot. Innovate UK ranks applicants and funds down to a budget, which is exactly the setting the strongest evidence abroad comes from (Chapter 1), yet none of the three UK evaluations in Chapter 1 uses it that way. Keep each application's final score and the funding line for every competition, in a form an evaluator can use in the secure lab. Where several applications sit tied at the line, judged equally good, decide between them by lot, as the Swiss National Science Foundation has done in all its schemes since 2021. Doing both turns the closest calls into a fair test. Innovate UK has run this setting before: in the Agri-Tech Catalyst only a third of applications scoring 70 or more, the level it deemed fundable, were funded, highest scores first, and its evaluator surveyed the applicants who fell short (SQW, 2020). With 103 projects, though, the evaluator judged a statistical comparison not viable. That's why the record matters for every competition: pooled across years, the numbers are there, and the Smart evaluation alone linked 5,849 high-scoring declined applicants. Won't other considerations, such as balance across a portfolio, muddy the line? They will, and the design allows for it. Scores needn't decide funding on their own; it's enough that the chance of funding jumps at the line, which is what economists call a fuzzy regression discontinuity (Imbens and Lemieux, 2008). What it can't do without is the score and the decision for every application. From the next round, as a first step, store each application's final score and the funding line, before anything else changes. One catch: a few applicants will lose by lot rather than by judgement, and some will resent it. Yet they'd be applications assessors couldn't separate on merit, and a published rule is easier to defend than an invisible tie-break.
3. Innovate UK: date the applications and decisions in the public file. Right now the funded-projects file gives only project dates, so no evaluator can tell hiring in expectation of a grant from a firm's own momentum. Publishing the month of application, decision and first payment would fix that. One more column would do as a first step: the month of the funding decision, in a file Innovate UK already publishes. Against it: would dates give away when a firm was seeking money? Giving the month rather than the day limits that, and there's a small cost to extend the file. It's a modest ask.
4. Innovate UK and DSIT: stop calling growth against never-funded firms impact. A comparison with firms that never applied mostly measures who wins, as Chapters 5 and 7 show, and a survival figure read alone counts a firm carrying losses for longer as a success. Every published growth claim of that kind should be labelled an upper bound, and monitoring should track survival, net assets and the share with negative net assets side by side, for funded firms and a comparison, from filed accounts. Labelling the next published growth figure against never-funded firms as an upper bound would be a first step, and it costs nothing. Its price: won't the programme look worse? Its published numbers will look smaller, yes. Smaller numbers that stand up are worth more to it than large ones that don't.
5. DSIT and the Evaluation Task Force: make the placebo test a condition of sign-off. Evaluations can match on the past and never show whether the groups were already diverging; in this study, that one test changed every growth conclusion. Every impact evaluation of business support should publish a pre-grant placebo beside its headline, whichever way it points, with its plan registered before outcomes are seen. A single line in the Evaluation Task Force's guidance, asking for the placebo beside the headline, is the first step. Its cost: isn't this one more burden on evaluators? It's a rerun of a model they've already built, but some evaluations will report weaker findings than they otherwise would, for the same money. We think that's worth paying.
6. Innovate UK: test whether its largest awards buy more than its smaller ones. First-year awards over £99,410 take 84% of first-year money, £1.68 billion since 2016, or about £153 million a year. Yet the employee gap those firms show three years on (3.39) is close to the head start they already had (2.64), so on the public record the extra money buys little the firms weren't already heading for. Put it to a test: in chosen competitions, offer one in ten of the largest awards, picked by lot, at a lower ceiling, and fund the next applications below the line with what's saved. A single competition, announced in advance, would be a first step. Its cost: won't smaller awards sink projects that need the full amount? Some might, which is why it's a trial on a tenth of them, each firm tracked against the full awards. At one in ten it moves about £15 million a year, enough to fund more applicants at the line and to show whether size matters.

12. Future evaluation and monitoring

What would settle the questions this study leaves open? Mostly, data that already exists. Here's what we'll do next, and what others could.

• Declined applicants. We asked UKRI on 20 October 2026 for unsuccessful applicants' company numbers by competition, product type and score band. If UKRI releases them, comparison D gets added, every result is rerun against it, and the update is published beside this version, with its own plan registered first; if it refuses, the refusal and its grounds go on the record too.
• A test at the funding line. With scores and the funding line (Recommendation 2), the design behind the strongest evidence abroad could be run on Innovate UK for the first time. It would settle the selection question this report keeps meeting.
• What to record. Appendix O sets out, field by field, what UKRI would need to keep for every application, funded or not, so that the next evaluation can make the comparison this one couldn't.
• Linking to ONS data. Business register employment and turnover cover every firm, which would close the gaps in Chapters 5 and 8. An accredited researcher could rerun this design inside the Secure Research Service.
• Where net assets come from. Retained earnings separate profit kept in the firm from capital raised. That split is now run (Chapter 8 and Appendix N). Rerun each year, it'll show whether the profit gap closes as funded firms' research reaches the market.
• A study that updates itself. Because the data are public, we can rerun every result as new accounts are filed, and the five-year results will cover more cohorts with every year that passes, which is exactly where this study is thinnest today. Our plan is to rerun it annually and publish what changes.
• Mixed methods. Interviews with funded and declined applicants, through associate evaluators, would show how grants were used and why the effects differ. That's beyond a desk study.

This report opened with a simple observation: firms that win Innovate UK grants go on to employ more people and close less often. That's still true. What's changed, by the end, is what it tells us. The grant picks out firms on the way up, and on the public record there's nothing left over to show it lifts them. Whether it does is a question UKRI's own data could answer, and we hope it'll let someone ask.