AI for Underwriting in Construction

The groundworks package is going to four subcontractors and the file on the cheapest one is open in the commercial office: a completed prequalification questionnaire, abridged accounts filed fourteen months ago, two insurance certificates, a safety policy signed by a director who has left, and an accreditation certificate that expired in March. Someone decides by Thursday whether this firm can carry a package worth more than the turnover it last reported.

Nobody calls this underwriting. It is underwriting: sizing an unsecured exposure against a counterparty you cannot see inside, on evidence of varying quality, to a programme deadline. The same business also buys formal underwriting from sureties, insurers and warranty providers, assembling a submission pack for each by hand. Neither half is bottlenecked on the decision but on assembling and cross checking the evidence.

The questionnaire comes back complete and tells you almost nothing

A prequalification return arrives finished: every box ticked, every policy attached, the answers the same ones this subcontractor sent three other contractors this quarter. Completeness is not the signal people treat it as. What matters is the gap between an answer and its attachment, and the gaps are routine. A procedure with another company's name in the footer. A competent person who left eighteen months ago. A turnover that does not reconcile to the accounts behind it. An entity on the form that is not the entity on the certificate. Matching each claim against the proof offered is agent work, and it returns a dated discrepancy list rather than nine attachments nobody reads end to end.

Reading a contractor's accounts for a job that has not started

The accounts you are handed are old by design. A small company files months after its year end, so you read a position that closed eighteen months ago, from a business whose order book has turned over twice since, with the profit and loss stripped out by exemption. They still carry something: net assets and their direction across three filings, cash against creditors falling due within the year, trade creditor days as a read on whether their own supply chain is stretched, directors' loans, charges registered and to whom, late filings, a changed accounting reference date. Then the comparison that matters, package value against anything this firm has ever turned over. An agent assembles that for four bidders at once, traced to source. Whether thin net assets matter in a groundworks firm with twenty years behind it stays judgement.

Who actually stands behind the company you approved

The trading name on the tender, the registered entity on the subcontract, the company on the insurance certificate and the name on the invoices often resolve to two or three legal persons, and the one you contracted with can be a thinly capitalised operating company while everything of value sits a level up. Worse, three firms on your approved list, in different trades under different names, roll up to one ultimate parent, and the combined exposure is a number no system holds. An agent can keep that mapping current rather than rebuilding it at audit time, matching on registered number, registered office, common directors, group filings and remittance bank details, and flag both the newly formed company at the address of a business that failed last year and the firm wrongly refused because a namesake failed.

A parent company guarantee is worth what the parent is worth

A guarantee gets filed and treated as solved. Worth checking: whether the guarantor trades at all, what its filings show, whether liability is capped below the package value, whether it expires at practical completion or runs through the defects period, and whether the signatory could bind that company. Comparing caps against subcontract sums across every guarantee you hold is mechanical.

A reference about a comparable project, and what comparable means

Referees are chosen by the applicant, so they describe the jobs that went well. The question is not whether the referee was happy but whether the project resembles yours in value, duration, sector, contract form and the thing you are worried about, usually programme or interface rather than workmanship. An agent can issue the requests, chase the silent ones and normalise replies into the same fields: value, dates, scope, finished on programme or not, delivered by this entity rather than an associate. It should also state what the set does not cover. No reference near your package value, nothing under your contract form, two referees sharing a registered office with the applicant: all findings.

Safety and quality history is evidence, not a score

Accident frequency rates on a questionnaire are self reported and self calculated, which is why two firms with identical records present differently. Harder evidence sits in reportable incidents, enforcement notices and prosecutions, and in what your own projects recorded the last time this subcontractor worked for you. That internal record is the most relevant source and the least used: close call reporting, open and closed nonconformances, how long corrective actions ran, inductions against operatives actually on site. Set beside the declared figures, the disagreements show. A firm claiming an immaculate record while carrying four open nonconformances on your current site is a conversation to have before the order goes out.

Every set of payment terms you agree is an unsecured loan

Procurement does not use the word credit, but it grants credit constantly. An advance payment for mobilisation is credit. Payment for materials stored off site is credit, secured by a vesting certificate and an advance payment bond if you are careful and unsecured if you are not. Front loaded rates are credit, and so is an over valued interim application, which is money paid for work that does not exist yet, recoverable only from a company that still trades. The exposure at any moment is cash out against work genuinely in place, plus materials paid for and not delivered, less retention held. An agent can compute that per subcontractor per period from the valuations, measured progress and the payment ledger, so over valuation becomes a weekly number somebody owns.

What it costs to replace a trade at week thirty

Prequalification thresholds usually scale with package value, which is the wrong variable. A default costs the premium to bring a replacement in at short notice, the time lost while they price and mobilise, extended preliminaries, damages if the critical path moves, rectification of half finished work, materials paid for and never delivered, and warranties that now cover nothing. A small package can sit on the critical path with no float and no alternative supplier while a large one is substitutable in a fortnight. Package value, programme float, payment position and the number of credible alternatives give an exposure figure per package, so vetting depth matches the damage rather than the order value.

Onboarding is a gate when the risk runs for two more years

Approval is a moment; the risk lasts through the package and the defects period behind it. In practice the file is built once, the company joins the approved list, and nothing is examined again until the trade is visibly in trouble, by which point the options are all bad. Monitoring is tedious, which is why it suits an agent: new and late filings, fresh charges, judgments, director resignations, insurance renewals and accreditation expiries, alongside the signals already in your systems, which are labour short of the planned resource three weeks running, applications rising while measured progress flattens, a request to shorten payment terms, hard chasing of retention, changed bank details on a remittance. Each is weak alone. Ranked weekly with the evidence attached, they arrive while you can still act.

Every bond and policy wants the same facts in a different form

A surety considering a performance bond wants filed and management accounts, a contracts in progress schedule, work in hand, bank facilities, bond history and usually directors' indemnities. Contractors all risks cover wants contract value and period, scope, existing structures, piling, hot works, flood exposure and the conditions requiring joint names. Renewal comes round annually and asks again. The facts barely move; the forms do, so the pack is rebuilt by hand, usually by someone senior, and versions drift until a broker notices that the work in hand figure contradicts the one sent nine weeks earlier. An agent can hold one maintained evidence set, generate each submission in the shape requested, list what is genuinely missing and flag any answer contradicting something already sent. The same set catches certificates that expire on their own anniversaries in the middle of a fourteen month programme, carry limits below what the subcontract requires, or end up in the wrong name after a restructure.

Latent defects cover is underwritten long before anyone claims

Structural warranty and latent defects policies are priced before work starts, with a technical audit running through construction: the surveyor inspects at defined stages, the design is reviewed, and sign off at completion depends on evidence that accrued months earlier. The failure mode is familiar, a stage inspection missed and an element covered up without the record, then a dispute at practical completion about something nobody can re-open. An agent can track the inspection sequence against site progress and hold the evidence each stage needs as it is produced.

Design and build moves a professional risk onto the contractor

Design liability flows down to trades that design their portion, and their professional indemnity cover is where it lands or does not. The mismatches are specific and findable: an aggregate limit where the subcontract required each and every claim, a limit below the requirement, a fitness for purpose obligation imposed by the subcontract against a policy excluding exactly that, six years of run off required and a certificate silent on it, collateral warranties promised to a funder on terms the insurer will not accept. Comparing the insurance clause against the certificate and policy summary for every design portion is repetitive reading an agent can finish with clause references attached.

Where AI fails here

Appetite is not an analytical output. Using a subcontractor with weak accounts because the managing director has delivered twice for you under pressure is a commercial judgement no model holds the inputs for. An agent can show the exposure, the comparable packages and the policy position. It cannot tell you whether this firm is worth the risk on this job.

Documents built to deceive will pass. A doctored insurance certificate, references from associated companies, accounts for a company that never traded as the questionnaire claims: these are structurally correct and will be read as correct. Verifying against the issuing source is a much harder problem than reading what you were sent.

The duty of fair presentation stays with you. An agent that drafts a submission and omits something material has transferred no responsibility; the signature is human, and the non disclosure is yours.

Your supplier master will defeat group aggregation before any model does, and fixing duplicate records, trading names held as separate suppliers and missing registered numbers is your work. Nor can an agent anticipate the failure that starts elsewhere: a subcontractor made insolvent because a different contractor did not pay a different account looks healthy in your data until the week it stops.

Where to start

Take the prequalification returns for one trade on one project and have an agent assemble them rather than score them. Each bidder comes back as one page: entity and group position, spread accounts with sources, the discrepancies between answers and attachments, insurance measured against the subcontract requirement with renewal dates set against the programme, safety and quality history including what your own projects recorded, and references normalised into the same fields. Approval authority moves nowhere.

Leave the submission packs alone until the evidence underneath them is clean, and leave the declaration and the award with the people who hold them now. The second build is usually monitoring across the approved list, because it tends to find exposure the business was already carrying.

Adopt AI Partners builds these agents inside the systems you already run, so the prepared file lands on the package rather than in another portal. If this sounds like your prequalification process, describe how a subcontractor gets approved today to hello@adoptaipartners.com.