Construction is under pressure. Can AI help protect margins and productivity?
A contractor's margin is rarely lost in one big event. It goes in small places: the instruction nobody can find when the valuation is argued, the quotations compared by hand at nine at night, the revised drawing that reached the wrong person. AI can help protect margin, but in a narrow and specific way. It will not move material prices or fill a skills gap. What it can do is take the finding, copying and reconciling out of professional work, inside the systems a business already runs, with a named person still approving anything that matters.
Part 1 | The pressure
How much pressure is UK construction actually under?
Enough to make efficiency a margin question. The S&P Global UK Construction PMI, a monthly survey of construction purchasing managers, rose from 44.3 in August to 46.1 in September 2026. Below 50 means activity is shrinking, so September was a slower fall rather than a recovery: the least marked downturn for eight months. The detail matters more than the headline:
- Orders: new work fell at the fastest rate since June, with clients deferring decisions on major projects.
- People: employment has fallen in every month since January 2025, and use of subcontractors declined again.
- Costs: around a quarter of firms paid more for purchases, citing fuel, freight and raw materials. Inflation eased to a seven-month low, but S&P Global's economists said that seems unlikely to last.
- Supply: delivery times lengthened by the most since May, which firms put down to shipping delays and disruption linked to the Middle East conflict.
Currie & Brown's 2026 Construction Certainty Index, a survey of more than 1,300 construction and infrastructure decision-makers, found uncertainty had added an average of 12.4 per cent to pipeline costs over the past year. It is global rather than UK, and self-reported rather than measured. Respondents saw 30 per cent of projects delayed and 27 per cent descoped. Among the 32 per cent it classed as most digitally and data mature, 58 per cent were confident of meeting deadlines, against 41 per cent of the rest: confidence rather than delivery, and correlation rather than cause, but it points the same way as everything below.
Labour is a paradox. Firms are shedding staff now, yet CITB's Construction Workforce Outlook says the industry needs an extra 41,200 workers a year between 2026 and 2030. When workloads recover the people will not all be there, and output per person will matter more.
Part 1 | What AI can touch
Which of those pressures can AI actually do something about?
Fewer than the marketing suggests. AI does nothing about interest rates, fuel prices or the size of the order book. It works on the cost of handling information, which runs through all of them.
AI works on the information, not the market
| Pressure | The evidence | Where AI helps | Where it does not |
|---|---|---|---|
| Thin order books | New orders falling fastest since June | Better evidenced bids from your own library and rates | Winning work that is not there |
| Input costs | A quarter of firms paying more | Quotations compared against the specification, not just on price | The price of steel, fuel or freight |
| Supply delays | Delivery times lengthening most since May | Flagging orders and approvals that put an activity at risk | Shipping capacity |
| Labour | 41,200 extra workers a year to 2030 | Taking repetitive administration off experienced people | Skilled trades or professional judgement |
| Uncertainty | 12.4% added to pipeline costs, self-reported | Earlier sight of issues buried in correspondence | Political and economic risk |
Part 1 | Where the time goes
Where does the administrative time actually go?
Into finding, copying and reconciling information that already exists somewhere: emails, drawings, specifications, quotations, minutes, purchase orders, payment applications, photographs and reports, rarely in one place.
Project manager
Searches Outlook for an instruction given in passing three weeks ago.
Quantity surveyor
Reconciles three spreadsheets that should agree and do not.
Estimator
Reads a tender pack for days to find the handful of clauses that carry cost.
Buyer
Emails five suppliers and retypes their quotations into a comparison sheet.
All of it is necessary, and little of it uses the judgement the person is paid for. That is the part AI handles well, and construction has barely started: the Office for National Statistics found 13 per cent of construction businesses using AI in June 2026, against around 35 per cent of UK businesses with ten or more employees.
Part 2 | Your systems
Do you have to replace your existing systems to use AI?
No, and usually you should not. The useful applications connect what a business already runs: Outlook, SharePoint, Excel, Teams, Dropbox and the project platform. A new platform the whole team has to adopt is where most of these projects stall.
Take one routine event. A subcontractor emails revised drawings and an updated specification. Today someone downloads the attachments, works out what changed, updates the tracker, tells the people affected and replies. An AI-assisted workflow can do the preparation and stop exactly where judgement is needed.
The professional stays in control; the copying disappears. Where the job is answering questions across thousands of records, the same principle applies to secure AI agents over your project records: inside your own Microsoft 365 tenant, inheriting the permissions you already have, and citing every source.
Part 2 | Buying
Can AI help buyers make better procurement decisions?
Yes, by doing the comparison rather than the choosing. AI can read the specification, pull the relevant properties from each quotation and data sheet, and set the options side by side: performance, certification, lead time and price, with gaps and non-compliances flagged rather than smoothed over. That moves the question from "which is cheapest" to "which is the best value and still compliant". A cheaper product that fails a specification clause, or arrives three weeks late, was never cheaper.
The decision stays with a competent person, and on some projects the law says so in detail. On a higher-risk building in England, the Building Safety Regulator's guidance says a like-for-like or higher specification replacement is a recordable change for the change control log. One with a different reaction to fire specification, even a higher one, may be a notifiable change reported before work starts, and one with a lower reaction to fire specification is listed as a major change, which needs the regulator's approval first. A tool that suggests alternatives without knowing that is a liability, not a saving.
Part 2 | Risk
Can AI spot project risks before they become expensive?
It can surface them earlier, which is most of the value. A quotation that has not come back, a technical submission awaiting approval, an unanswered query and an order not yet placed are each small. Together, two weeks before an installation, they are a programme problem.
The warning signs are usually written down already, across email, minutes, trackers and reports. An AI-assisted check that reads them together finds the combination and sends it as an action to the right person, with the correspondence attached, instead of waiting for the next progress meeting. The measure that matters is the interval between a problem being created and being recognised, as set out in how to measure what AI actually changed.
Part 2 | Reporting
How much of the reporting and compliance burden can AI take on?
A large share of the assembly, and none of the sign-off. Diaries, inspection and condition reports, health and safety, quality and progress records all mean gathering information into a set template. AI can extract it from documents, photographs, voice notes and meeting notes, draft in your own template, check for missing sections and unsupported statements, organise the photographs, track open actions, and record each revision and who reviewed it.
A surveyor spends more of the day on findings, which is the idea behind turning inspection photographs into finished reports, and a site manager's voice notes become the daily record through our WhatsApp site assistant. The saving across a whole report is smaller than most suppliers quote, for reasons set out in what automated reporting actually saves.
Part 2 | Safeguards
What safeguards should construction AI have?
The ones you would expect of a capable new member of staff, written down. Construction documents carry commercially sensitive information, contractual obligations, personal data and safety-critical detail.
Five safeguards, written down before go-live
| Safeguard | What it means in practice |
|---|---|
| Controlled access | The system sees only what the person asking could already see |
| Human approval | Nothing consequential is sent, issued or ordered without a named person |
| A record of AI output | What was suggested, from which sources, and who accepted it |
| Data protection | Client and project data stays in your environment and is not used to train public models |
| Traceability | Every statement in a report can be followed back to where it came from |
An AI assistant can draft a contractual notice; it should not serve one. It can flag a safety concern; it cannot replace a competent person's assessment. The human in the loop matrix sets out where the line sits, a simple AI register shows it is kept, and our trust centre covers how we handle client data and Microsoft tenants.
Part 3 | Where to start
Where should a construction business start?
With one workflow, not a transformation programme. Pick something repetitive, owned by one team and frequent enough to measure: a weekly report, a drawing register, a quotation comparison, a site diary. Measure it now, build the automation into the systems you already run, run it on live work, and measure again.
That is how our first projects work: one workflow, measured before and after, live in about six weeks, from £1,500. If the saving is smaller than hoped, you found out for the price of one workflow rather than a platform.
Part 3 | How we help
How does AI Metric help construction businesses?
By starting from how the business actually runs, then building the smallest useful thing and proving it on live work.
Project management
Correspondence, documents and actions brought together and searchable, through secure agents.
Procurement
Specifications read and quotations compared side by side, as a workflow automation project.
Commercial management
Valuations, payment administration and deadlines, plus AI for estimating and take-off.
Site operations
Photographs, observations and the daily record, from the WhatsApp site assistant.
Technical reporting
Structured reports and evidence through inspection reporting and document automation.
Business administration
Less repetitive work across email and documents, starting from a clean SharePoint and CDE.
Part 3 | The limits
What will AI not fix?
AI will not bring material prices down, replace skilled tradespeople, or remove the need for experienced project managers, engineers and commercial professionals. It will not win work that is not there.
What it can do is remove hundreds of small frictions: fewer hours searching, quicker documents, better informed buying, earlier sight of risk and more consistent reporting. None is dramatic on its own. On margins this thin, across a year of projects, together they are the difference. The future of AI in construction is not replacing construction expertise. It is giving that expertise better tools.
Sources
Which sources is this article built on?
Every figure above resolves to one of these, each read at source before publication.
- S&P Global, UK Construction PMI, September 2026, published 6 October 2026
- Currie & Brown, Construction Certainty Index 2026, published 6 October 2026
- CITB Construction Workforce Outlook 2026 to 2030, as reported by Construction News, 18 June 2026
- Office for National Statistics, Artificial intelligence in UK businesses: 2023 to 2026, 20 July 2026
- Building Safety Regulator, Making changes to a higher-risk building project, updated 1 October 2026
Photographs by Tsuyoshi Kozu and Scott Blake on Unsplash, used under the Unsplash Licence.