AI Can Draft Your Claim. It Can't Prove It.
What artificial intelligence actually changes about construction claims — and what it doesn't
Every few months, a new tool arrives promising to read your contract, sort your correspondence, and hand you a claim narrative before lunch. Some of them are genuinely impressive. We've watched systems pull a coherent chronology out of eighteen months of site diaries and RFIs faster than a graduate engineer could do in a week.
So the question we keep getting from contractors, developers and project teams isn't unreasonable: if AI can read the contract, sort the records and draft the claim, what's actually left for the claims professional to do?
The honest answer sits between the two extremes people usually reach for — "AI will replace claims consultants" on one side, "it changes nothing" on the other. AI is genuinely reshaping how a claim gets built. It has had no effect whatsoever on what a claim has to prove. And that gap between building a claim and proving one is exactly where this conversation needs to start.
A claim is not a writing exercise
We've seen beautifully produced claim submissions — properly bound, cross-referenced, hundreds of pages of correspondence, programmes and calculations — collapse under fifteen minutes of cross-examination, because nobody had actually established the causal chain the document was supposed to represent.
Presentation was never the hard part. A construction claim, however it's drafted, still has to walk a straight line from contract, to event, to entitlement, to cause, to effect, to time, to cost, to evidence. FIDIC's own guidance on claims puts the same elements at the centre of any properly structured submission: the contractual basis, the cause, the effect, entitlement and quantum. AI can help you organise and analyse those elements. It cannot conjure them into existence where the underlying facts don't support them. That's the difference between drafting a claim and proving one — and it's a difference that no amount of polish closes.
Where AI genuinely earns its keep
We don't think it's useful to be instinctively sceptical about this technology, because on a typical major project there is a genuinely enormous amount of material to get through: thousands of emails and letters, hundreds of RFIs, multiple drawing revisions, monthly progress reports, baseline and updated programmes, minutes, payment applications, variation instructions, site diaries, photographs, inspection records, and subcontract correspondence. Historically, a claims team burned weeks just locating and organising this before the actual analysis could begin.
That's the part AI is genuinely good at compressing:
Document discovery — pulling every piece of correspondence tied to a specific instruction, drawing revision, subcontractor or date, in seconds rather than days.
Contract analysis — surfacing the relevant clauses, notice mechanisms and obligations buried across a lengthy conditions of contract and its amendments.
Chronology building — assembling a first-pass sequence of what happened, when, and who was involved.
Pattern recognition — flagging recurring references to delay, access restrictions, approvals, or disruption across correspondence that no single person could read in full.
Classification — sorting thousands of records by event, package, discipline or contractual issue.
Early signal detection — identifying commercial issues while they're still emerging, rather than after they've hardened into disputes.
This isn't a fringe conversation anymore. FIDIC has run dedicated sessions on AI's role in contract administration, project delivery and the balance between AI output and professional judgement, and its 2026 Global Infrastructure Conference gave real airtime to agentic AI across the infrastructure lifecycle. The industry has moved past asking whether this matters.
The problem AI can't route around
Here's the limitation that doesn't get discussed enough: AI is only as good as the project information it's fed, and construction projects are not exactly famous for clean, complete, well-indexed records. Information is missing, contradictory, duplicated across platforms, scanned in badly, recorded well after the fact, or built on assumptions that were never formally confirmed by anyone.
Which leads to a principle we would want any project team to adopt before they get excited about AI-assisted claims: AI can process evidence. It cannot create evidence that was never generated.
If the required notice was never issued, no tool will manufacture one after the fact. If your programme doesn't actually demonstrate the delay you're alleging, AI cannot turn a weak programme into reliable delay analysis — it will just describe the weak programme more efficiently. If additional costs were never properly recorded at the time, there's nothing for the model to retrieve. And if an instruction exists only as a passing reference in an email thread with no formal authority behind it, a well-formatted AI-generated submission doesn't cure that gap. It just makes the gap look more presentable.
Correlation is not causation — and AI doesn't know the difference
This, to us, is the sharpest risk in the whole conversation. Feed a system enough correspondence and it will happily surface a tidy sequence: drawing delayed, activity delayed, completion date moved, additional cost incurred. It looks convincing precisely because it's coherent.
But coherence isn't causation, and this is where judgment still has to do the work a model can't. Was that activity actually on the critical path, or was there float to absorb it? Was another event already delaying the same activity independently? Did the contractor mitigate, and could it have progressed something else in the meantime? Was the drawing genuinely required at that point in the sequence, or is that an assumption baked into the analysis? And under the contract's risk allocation, does this event even sit with the party bringing the claim?
None of that is a document-retrieval question. It's contractual interpretation, programme logic, and commercial judgment — exactly the territory a claims professional is trained for and a language model isn't.
The "black box" problem
Imagine a system reviews a hundred thousand project documents and returns a clean conclusion: an entitlement to 84 additional days and $4.2 million. The instinct might be to ask how the tool got there. That's the wrong first question.
The right one is: can we actually reconstruct and defend that number? Which contractual clause was applied, and by what reasoning? Which specific documents establish the event and the causal link? Which programme version was used, and why that one? What methodology drove the quantum? What assumptions were made, and what was left out? Can another professional independently reproduce the analysis? Would it survive scrutiny in front of an Engineer, a dispute board, an adjudicator, or an arbitral tribunal?
A claim is a professional and contractual proposition before it's anything else. If it can't be traced back, step by step, to something real, it isn't ready — regardless of how it was produced.
No technology replaces the underlying project record
This is where we would steer the conversation away from the tools entirely. The strongest claims position I've ever seen on a project didn't come from the most sophisticated software — it came from a commercial team that had simply kept its house in order: timely notices, contemporaneous records, an accurate and regularly updated programme, clearly documented instructions, a properly maintained variation register, reliable cost records, and evidence of mitigation as it happened rather than reconstructed afterward.
Those fundamentals predate generative AI by decades, and they'll outlast it. A recent RICS programme for construction professionals across the GCC still centres on exactly this — identifying entitlement, structuring claims around cause and quantum, and supporting every submission with proper evidence. Technology changes how fast you can analyse a position. It has never changed what that position actually requires to hold up.
The real shift: professional plus AI plus better data
We, at VersaMinds, don't think the future here is claims professionals versus AI. It's more accurate to describe it as claims professional, plus AI, plus meaningfully better project data — three things reinforcing each other rather than one replacing the others.
Used well, AI takes the multi-day document hunt off a professional's plate and returns that time to the questions that actually decide a claim's outcome: what does the contract provide for here, was the notice valid, what genuinely caused the delay, what was the critical-path impact, what mitigation was realistically available, what costs were actually incurred, and where is the opposing position weakest? The technology absorbs more of the information burden. The professional spends more of their time on judgment. That trade is worth having.
What this means for how claims professionals need to work
The core skill set hasn't changed — contracts, programmes, costs, evidence, still the foundation. What's being added is a layer of literacy around how technology interacts with those disciplines.
That means learning to frame a problem precisely enough that an AI system can actually help with it, rather than accepting a fluent-sounding answer at face value. It means treating any AI output as an analytical input to be tested, not a conclusion to be filed. It means being able to trace every material conclusion back to a genuine source document, every time, without exception. And it means taking data risk seriously — construction projects carry commercially sensitive pricing, designs and contractual material, and Autodesk's own guidance to the AEC industry has flagged real concerns around uploading that kind of information into uncontrolled public AI environments. Knowing when not to automate a decision — a formal notice, a legal conclusion, a dispute strategy — is as much a professional skill now as knowing when to use the tool.
FIDIC's own discussions on this point are consistent: reliable inputs and human oversight aren't a final check bolted onto the end of the process, they're part of the system throughout. The workflow isn't documents in, AI out, claim done. It's closer to contract and project records and professional review feeding AI-assisted analysis, which then gets validated, assessed commercially and contractually, and only then shaped into claim strategy. AI should accelerate professional work. It was never going to be able to absorb professional accountability, and it still can't.
The more interesting opportunity is upstream of the claim
The application of this technology we find genuinely valuable isn't drafting claims at all — it's catching the issues before they become claims. A system watching project data can flag repeated delayed approvals, a rising volume of RFIs, recurring design revisions, instructions drifting from original scope, productivity anomalies, or programme slippage long before any of it is formally captured as a commercial event.
The point of that early signal isn't to generate another claim faster. It's to put a genuinely useful question in front of management while there's still time to act on it: is this becoming a contractual issue that needs addressing now, rather than in six months' time when the evidence trail has gone cold?
That reframes the whole function. Instead of a contracts team asking, after the fact, "what claims do we have," it becomes a team that can ask, in near real time, "where is our contractual and commercial exposure actually developing" — by pulling contract data, project data, programme data, cost data and correspondence into one connected picture. For major capital projects, that's less an incremental improvement and more a shift from a largely administrative function toward genuine project-control discipline.
So — can AI draft your claim?
Yes. It can help structure the narrative, search the records, identify the relevant clauses, build the chronology, classify the documents, flag the gaps, and accelerate the analysis considerably. All of that is real and worth using.
But there's a line we don't think should move: AI can help present a case. It cannot manufacture one. A claim still needs entitlement, causation, evidence, and a defensible position on time and cost — and it still needs someone who understands what the contract actually means against what genuinely happened on the project.
The real competitive advantage
The question worth asking isn't "do we have AI." It's whether you have the contractual data, project records, processes and professional capability to use AI properly in the first place. Applied to poor project information, AI just produces confident-sounding bad answers faster than a person could. Applied to a genuinely disciplined contract-management environment, it can give commercial teams something they've historically struggled to get: a near real-time view of contractual risk, entitlement and exposure across the project.
That's where the real opportunity sits. This was never going to come down to humans versus machines. It comes down to professionals who know how to use intelligent technology properly, against everyone still trying to manage increasingly complex projects with fragmented information and entirely manual processes.
AI can draft your claim. It can't prove it. The professionals who understand both halves of that sentence are the ones who'll still be standing when the dust settles.
A few questions we get asked often
Can AI replace a claims consultant on a construction project? No. AI can accelerate document review, chronology-building and drafting, but it cannot establish contractual entitlement, causation or the credibility of evidence — that still requires professional judgment and contractual interpretation.
What's the biggest risk of using AI to prepare a construction claim? Treating a coherent-looking output as proof. AI can surface a plausible sequence of events, but confirming genuine causation, critical-path impact and entitlement under the contract still requires a professional review.
Is it safe to upload project data into AI tools? Not into uncontrolled public tools. Construction projects contain sensitive pricing, design and contractual information, and industry guidance — including from Autodesk — has flagged real data-privacy and confidentiality risks in AEC's adoption of AI.
What should contractors actually do to get value from AI in claims management? Fix the underlying project data first — timely notices, contemporaneous records, accurate programmes, proper variation registers. AI applied to disorganised project information simply produces unreliable conclusions faster.
How we think about this at VersaMinds
Technology is an enabler of better contract and commercial management here — not a substitute for professional judgment. Our work covers contract management, commercial management, claims, EOT, variations, procurement, project controls and dispute support, always anchored in connecting what the contract requires with what's actually happening on the ground.
For any organisation looking at AI-enabled contract and claims workflows, the starting point isn't the software. It's the contract, the data, the processes and the commercial objective. Get those right first, and the technology can do a great deal more than draft a claim.