Farrans and Arup both blamed problem contracts. Can AI spot a losing job sooner?
A contract rarely becomes a loss on the day the provision is booked. The signs are written down months before: an instruction that was never priced, a notice that went out late or not at all, a forecast that has not moved while the cost has, a subcontractor's application that keeps running ahead of the valuation. By the time the number reaches the accounts, the story has been sitting in the correspondence for a year. The weekend's results are a reminder of what that costs, in a market with very little margin to absorb it.
Part 1 | The results
What did Farrans and Arup actually report?
Losses and provisions on contracts neither firm named. Farrans swung to a £36m pre-tax loss from a £3.2m profit the year before, on turnover down 22 per cent to £298m, after its directors booked further provisions on several older contracts, Construction Enquirer reported on 18 October. The Belfast contractor was bought by Sisk in October 2025, and its directors expect 2026 to finish positively on a strong framework pipeline.
The same day, Arup reported profit at a 13-year low for the year to March 2026. Pre-tax profit fell 41 per cent to £26m while revenue grew 5 per cent to £2.27bn. It booked nearly £26m of new provisions against loss-making contracts and £12m for legal costs, and its operating margin fell from 2.7 to 1.5 per cent.
None of this is a comment on either firm's projects, which were not identified. The point is how little room there is. The Office for National Statistics reported on 15 October that construction output fell 0.2 per cent in August, the fourth monthly fall in a row, with new work down 1.4 per cent over three months, as Construction Enquirer covered the same day. With less new work coming in, a bad job from two years ago has fewer good ones around it to hide behind, which is why construction insolvencies are a margin story long before they are a legal one.
Part 1 | The signs
Where does a losing contract show itself first?
In the commercial and site records, long before the cost report. Anyone who has run a package through a bad year will recognise the list below. The signs are rarely hidden. They are spread across inboxes, minutes, applications and programmes, and nobody has the hours to read them all together every week.
The early signs of a losing job, and where each one is written down
| The sign | Where it is written down | What an AI check looks for | Who acts on it |
|---|---|---|---|
| Change that was never priced | Instructions, RFI answers, minutes, emails | An instruction with no matching variation or early warning | Quantity surveyor |
| Notices late or missing | Correspondence, programme updates | An event that looks notifiable with nothing served in the period | Commercial manager |
| A forecast that does not move | CVR, cost reports | Cost to date rising faster than the forecast allows | Commercial director |
| Applications running ahead | Subcontract applications and valuations | The gap between applied and valued widening month on month | Quantity surveyor |
| Slip with no claim behind it | Programme revisions, site diaries | Activities sliding with no matching delay event | Project manager, planner |
| Rework creeping in | Non-conformance reports, snag lists, diaries | Rising rework with no cost code or recovery route | Package manager |
Each row is a pattern an experienced commercial manager already looks for. The value of software is reading all of it, on every job, every week, and putting the evidence in front of the right person with the source attached. The detail behind the first two rows is in tracking variations before they become disputes and JCT and NEC notices.
Part 2 | How it works
How would an AI early warning actually work?
As a weekly read of the records, producing a short list of evidenced flags rather than a dashboard of opinions. Each flag names the sign, links the email, minute or application behind it, and goes to a named owner. The owner accepts it, dismisses it with a reason, or acts, and that decision is logged. Over a few months, the log of dismissed flags tells you where the check is too sensitive, and the accepted ones tell you where the money was.
This is where this autumn's model releases matter. Reading every piece of correspondence on every job each week used to be too expensive to justify. At the prices now on offer it is a modest line in an overheads budget, and Google claims its new Gemini 4 Argon ranks first on the Vals Index, which covers finance, legal, tax and coding work. That is a vendor claim to test, not a reason to buy. The design principles are in AI and commercial management and discipline agents for construction oversight.
Would this have saved a legacy contract?
Sometimes, and it is worth being straight about which times. Earlier sight of unpriced change, a missed notice or a drifting forecast gives a commercial team months to act: price it, notify it, negotiate, re-plan. That is where most recoverable money is recovered. Software cannot fix a job that was underpriced at tender, a client that will not pay or a market that turns, and nobody should sell it as if it can.
Arup's results carry a second lesson. The firm has been spending on people, processes and systems, including AI and machine learning, and its chief executive acknowledged that the spending held back this year's numbers. Investment comes before return. That is an argument for starting small and measuring, not for waiting.
Part 2 | The limits
What should a commercial team not hand to software?
The provision, the forecast, any notice served and any negotiation. Software reads and flags; the commercial team decides and signs. For quantity surveyors in RICS-regulated firms this is a professional duty. The RICS standard on responsible AI use requires a written, named reliability decision on material AI output, a register of the AI systems in use and a risk register reviewed at least quarterly, and clients told in writing before AI is used on their work.
The CIOB's guidance on developing an AI strategy points the same way from the contractor's side: choose a specific use case, be honest about how good your data is, and run a small pilot with return measures defined before anything is scaled. Commercial early warning is a strong first use case for exactly that reason. The records exist, the outcome is measurable, and the person who acts on the flag is obvious.
Part 3 | Where to start
Where should a contractor start?
With a job that has already finished. Run the check over the records of a completed project whose problems you know, and see whether it would have flagged them, how many weeks earlier, and how many false alarms it raised along the way. If it would have caught the unpriced change in March rather than the final account in December, you have your business case. If it would not, you found out for the price of one back-test.
Then run it on one live job alongside the normal commercial meeting, as we describe in a safe 30-day pilot. Our first projects work the same way: one workflow, measured before and after. If a legacy job is keeping you up at night, book a call.
Sources
Which sources is this article built on?
Every figure above resolves to one of these, each read at source before publication.
- Construction Enquirer, Farrans plunges to £36m loss as legacy jobs bite, 18 October 2026
- Construction Enquirer, Arup profit slumps to 13-year low as project provisions mount, 18 October 2026
- Office for National Statistics, Construction output in Great Britain: August 2026, 15 October 2026
- Construction Enquirer, Construction output slides for fourth month running, 15 October 2026
- Google, Introducing Gemini 4 Argon, 13 October 2026
- RICS, first standard on responsible AI use now in effect, and the standard itself
- CIOB, Developing an artificial intelligence (AI) strategy
Photograph by Catia Climovich on Unsplash, used under the Unsplash Licence.