On 14 September Futurum, a US technology research firm, published the AI spending results from its 2H 2026 CIO & Technology Buyers survey. Futurum says 46.9% of enterprises are running over budget on AI. Only 5.6% came in under. A further 10% have no formal AI budget to measure against, and 5.6% do not know where their spending stands.
If you work in a consultancy, a contractor or a client-side PMO, this lands closer to home than an enterprise IT survey suggests. Over the past 18 months, most firms in our sector have invested in Copilot, ChatGPT, Claude and AEC-specific platforms. Many of these started as pilots or small-scale purchases. Those contracts are now coming up for renewal. As FY2027 budgets are built this autumn, finance directors will want to know what those investments have delivered.
The question at an individual level is simpler. When licences are cut, whose seat goes? The answer will depend on whether anyone can show that the tool has made a measurable difference to their work.
Most firms funded the overrun rather than pausing
Futurum's data covers 767 organisations that reported running over budget. Of those, 47.6% sought approval for supplemental funding and 43.3% absorbed the overrun into the next planning cycle. Around one in six reduced or paused AI scope. So the dominant 2026 response to blowing the AI budget was to find more money and carry on.
Mitch Ashley, VP and Practice Lead for the CIO & Technology Buyers practice at Futurum, and the report's author puts the cause down to a mismatch between how budgets are set and how AI costs behave.
AI budgets are set once a year, occasionally with a mid-year true-up, while the underlying costs remain highly variable and poorly understood.
For a QS that should sound familiar. It is the provisional sum problem: a figure entered once for something nobody has fully scoped, then a quantity that grows month by month.
Ashley's view is that the pattern will not hold. He argues that each quarterly review and annual planning cycle gives finance another chance to challenge the assumptions behind AI spend and that the FY2027 cycle is where enterprises with weak or no AI budget controls will feel pressure first. He also notes that a minority, 31.8%, kept AI spending approximately in line with budget and attributes that to having a real operating number and the management behaviour to run against it.
Futurum works with many of the vendors involved, so treat the interpretation as an analyst's view. The survey figures speak for themselves.
The bill is moving out of IT
The Futurum data contains one finding that is especially relevant to our sector. Among over-budget enterprises, 38.7% reallocated funds from elsewhere in the IT budget and 23.1% shifted spending to a non-IT business-unit budget. In a consultancy that means the cost management team or the project management service line ends up paying for its own tools out of its own P&L.
Who pays changes the question. If IT pays for the licence, the service line can treat it as a free resource. If the service line pays, its head will want to know what each team is actually using Copilot for and whether it is helping with fee-earning work. Futurum also expects vendors to move toward pricing tied to outcomes, and The New Stack reported in August, citing The Information, that OpenAI has begun testing billing for some enterprise customers only when an agent completes a task. OpenAI has not published details.
What the project delivery data says
The most useful AEC-adjacent evidence this month comes from Tempo Software, which released its 2026 State of AI in Portfolio Management report on 9 September. Tempo sells portfolio management and AI spend attribution software, so it has a product to sell on the back of the findings. The survey itself was run by Potloc, an independent research firm, among 300 senior project, portfolio and PMO leaders in June 2026, at organisations with 200 to 4,999 employees, 60% in North America and 40% in Western Europe.
Tempo found 91% of respondents piloting or using AI in project delivery, but only 33% delegating actual work such as documentation or quality assurance to AI agents. Organisations with agents in production reported that 71% of their projects finished within six months, against 55% at organisations without. The typical project at an AI-enabled organisation ran about three and a half months compared with nearly six across the full sample. Those are self-reported survey figures, not a controlled comparison, and firms that have got agents into production may simply be better run in the first place.
The finance-relevant numbers are that 42% of leaders said they cannot tie AI spend to return on investment, and 39% cannot distinguish AI-produced work from human work in their current tools. Vic Chynoweth, CEO of Tempo Software, summed it up in the release.
The results show that using AI is not, by itself, an advantage. What matters is whether organizations can put agents to work, and govern that work alongside their people, investments and strategic priorities.
For a reader in a PMO, the record of what the tool did is the argument for keeping it, and on Tempo's numbers most PMOs do not have one.
What a defensible licence looks like
Two further sources sharpen the point. Glean, which sells enterprise AI search, surveyed 6,000 full-time digital workers across the US, UK and Australia between December 2025 and January 2026 for its Work AI Index. Glean says workers spend an average of 6.4 hours a week on what it calls 'botsitting', meaning feeding AI context, checking outputs and cleaning up its mistakes, and that this accounts for 37% of all time spent with AI tools. If your firm has not measured that supervision time, the productivity case for the licence is incomplete in both directions.
At the other end, Bank of America's CEO Brian Moynihan told the bank's second-quarter earnings call in July that it has approved more than 300 AI use cases, "all of which have good economics", and that the 19,000 developers using AI coding assistance are seeing productivity gains above 20%. Whatever you make of the numbers, a bank that can say that has done the counting. Most of our sector cannot yet.
In practice, every licence in a construction consultancy or contractor should be able to show at least one of the following before FY2027 budgets close.
Hours saved on a recurring deliverable, such as the monthly cost report or the progress report, measured across at least three cycles by the person who produces it.
The number of RFIs or technical queries turned round with AI-drafted first responses, with the review time recorded alongside.
Time spent on programme quality checks, such as logic and float review, before and after the tool was introduced.
A record of the supervision time, in the Glean sense, so the net figure is honest rather than flattering.
A named person who owns the licence and will answer for it at the budget review.
Takeaway
Futurum's survey says nearly half of enterprises overspent on AI in 2026, and most of them funded the gap. Tempo's survey says 42% of project leaders cannot connect that spend to a result. Both firms have something to sell, but the direction of travel is the same, and it points at your finance director's desk this autumn. The licences that survive FY2027 will be the ones attached to a measured artefact. The ones that go will be the ones where the only evidence is that people seem to like them.
Pull the list of AI licences your team holds and find out who pays for each.
Attach each licence to one deliverable and one number, and start measuring this month.
Log supervision time as well as time saved.
Take that record into the FY2027 budget conversation before someone else arrives with a spreadsheet of seat counts.
Project Flux follows how AI spending in construction and the built environment turns into results or fails to. If you want the numbers before your finance team finds them, subscribe at projectflux.ai.
Links and Stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
