NewThe Scope Gap Playbook - 8 Habits Behind Subcontractor Scopes That Survive the ProjectRead the Playbook →

Contract Risk Analysis: What Generic AI Misses That Purpose-Built Tools Catch

By Provision·June 23, 2026

TL;DR

  • Generic AI tools like ChatGPT miss supplementary conditions, Division 1 requirements, and trade-specific risk clauses on real construction documents.
  • Errors and omissions in contract documents have been the #1 dispute cause for 6 of the last 9 years, per Arcadis.
  • Purpose-built contract risk analysis AI is trained on construction-specific language and structured to catch the gaps that generic tools skip.
  • Provision's Risk Review delivers 99.5% accuracy on pre-built risk checklists — versus far lower accuracy on the same documents from general-purpose AI.
  • The difference isn't marketing. It shows up in your change order log.

Every GC using ChatGPT for contract review in 2026 is running the same experiment. The question is whether you find the gaps before bid day or after mobilization.

Generic AI is fast. It summarizes. It answers questions. But contract risk analysis on real construction documents requires more than summarization. It requires knowing what to look for, where to find it, and how clauses interact across a 2,000-page project set.

That's where general-purpose tools fall short — and where the difference between "AI-assisted" and "purpose-built" becomes a dollar figure in your project closeout.

Why Contract Risk Analysis Is a Hard Problem

Construction contracts aren't standalone documents. A $50M ICI project might include the prime contract, supplementary conditions, Division 1 general requirements, Division 2–49 trade specs, addenda, and owner-furnished schedules. Each document can modify, override, or add to the others.

The risk isn't always in the obvious clause. It's in the interaction between a supplementary condition that shifts "readily inferable" scope to the GC and a trade spec that's silent on material protection. Miss the connection — and you've absorbed a loss the sub won't cover.

A $300K lead-lined glass scope gap at a hospital imaging suite is a documented example of this. The GC absorbed the cost under "readily inferable" language. The spec was clear to anyone who read it in context. No one did. That example comes from the Scope Gap Playbook's chapter on subcontract language — it's one of dozens of real project losses tied directly to contract language that looked clean on the surface.

Generic AI doesn't read documents in context. It processes text. That's a meaningful difference on a construction contract.

What Generic AI Actually Gets Wrong

1. Supplementary Conditions Get Summarized, Not Interrogated

Supplementary conditions are where owners move risk. Liquidated damages, indemnity expansions, insurance requirements, notice periods — these are rarely in the main body of the contract. They're in the SCs, which look like boilerplate but aren't.

Generic AI reads supplementary conditions as text. It doesn't flag that a specific SC clause eliminates the notice requirement for back-charges, or that an insurance provision requires coverage levels that change your sub-tier requirements. It summarizes the section. That's not the same as catching the risk.

2. Division 1 Is Systematically Under-Read

Division 01 — General Requirements — sets the rules for every other division. Submittals, substitutions, quality control, temporary facilities, schedule of values format, and coordination requirements all live here. So do clauses that affect your cost: owner-supplied material allowances, testing responsibilities, and phasing requirements.

Early tests of ChatGPT's Construction Mode on real project sets show consistent gaps in Division 1 interpretation. The tool reads the text. It doesn't synthesize how a submittal requirement in Division 01 30 00 affects the MEP coordination workflow in Division 23.

3. Trade Assignments Are Guessed, Not Confirmed

One of the most common pre-construction failures is scope that falls between trades. Who supplies the concrete pump? Who owns the motor starters? Who installs the hollow metal frames — drywall or doors?

Generic AI, when asked to assign trade responsibility from a spec book, relies on general construction knowledge. That works when specs are explicit. It fails when specs are silent, ambiguous, or use non-standard terminology. A generic tool doesn't know that "gypcrete" and "gypsum underlayment" are sometimes used interchangeably in specs but mean different products with different subs — a naming collision that creates gaps.

A $45K stone-depth mismatch between civil, structural, and architectural drawings on a single slab is exactly this type of gap. It wasn't a missing clause. It was a conflict nobody caught because no tool was looking for it systematically.

4. No Memory Across the Document Set

Ask ChatGPT a question about a specific clause. It answers based on what it can read in its context window. Ask it to cross-reference that clause against a drawing callout, an addendum revision, and a supplementary condition — and it loses coherence.

Real contract risk lives across documents. An addendum that changes the spec section number without updating the drawing callout. A prime contract clause that contradicts the subcontract form. A phasing drawing that conflicts with the Division 01 schedule requirement.

Purpose-built tools ingest the full project set and hold it in memory. Answers are cited back to the specific document, section, and page. That's not a feature — it's the baseline requirement for construction contract review.

What Purpose-Built Contract Risk Analysis Actually Does

Provision's Risk Review was built specifically for GC pre-construction workflows. It doesn't summarize documents. It checks them against a structured risk framework — one developed from reviewing over $100 billion in project value and processing more than 66,000 construction documents.

Here's what that looks like in practice:

Pre-Built Risk Checklists Tuned to Construction Language

Risk Review runs a structured checklist across your contract documents. The checklist items aren't generic legal risk categories. They're construction-specific: notice requirements for concealed conditions, indemnity scope relative to sub-tier coverage, liquidated damage caps, schedule of values approval rights, and dozens more.

The pre-built checklists achieve 99.5% accuracy on real construction documents. Custom checklists built for specific project types or owner requirements come in at 97%+. These aren't internal benchmarks — they're verified against real project sets.

Cross-Document Synthesis

Provision's Chat Agent reads drawings, specs, contracts, RFIs, and addenda as a unified project set. Ask a question about trade responsibility for a scope item, and the answer cites the spec section, the relevant drawing callout, and any addendum that modified either. Answers come back in under 20 seconds.

That citation chain is what makes the answer usable. An answer without a source is just a guess. On a $50M project, guesses become change orders.

Scope Package Generation That Reflects Actual Contract Requirements

Missing contract risk doesn't just affect your prime contract exposure. It affects your subcontract scope packages. A GC that misses a Division 01 requirement for sub-tier insurance doesn't just have a risk in their prime — they have a gap in every subcontract scope package they issued.

Provision's Scope Agent generates complete scope-of-work packages from construction documents. It reads the drawings and specs together — not just the spec book — and produces packages that reflect what the documents actually require. That matters because, as the Scope Gap Playbook documents, the most common anti-pattern in scope writing is "as per plans and specs." That phrase looks like coverage. It isn't.

"If you miss anything, they'll bill it," as one Pre-Construction Lead at a Top-ENR Canadian GC put it. The sentiment is universal. The prevention is in the detail of the scope package, not in the breadth of the incorporation clause.

The Accuracy Gap Is Not Hypothetical

Arcadis tracked $60.1M as the average U.S. construction dispute value in 2026. Errors and omissions in contract documents have been the #1 dispute cause for 6 of the last 9 years. That's not a technology problem. It's a review problem.

The FMI Construction Disconnected report puts annual U.S. rework costs from miscommunication and bad project data at $31 billion. Twenty-six percent of that rework comes from communication breakdowns. Twenty-two percent from bad project data. Both are preventable with better document review.

Generic AI reduces the time you spend reading. Purpose-built AI reduces the risk you carry forward. Those are different outcomes.

A Practical Comparison: Generic AI vs. Purpose-Built Tools

Capability Generic AI (ChatGPT, Copilot) Provision Risk Review
Reads drawings + specs together No — text-only, no drawing interpretation Yes — full project set ingestion
Structured risk checklist No — freeform Q&A only Yes — 99.5% accuracy on pre-built checklists
Division 01 synthesis Partial — reads text, misses cross-document implications Yes — flags how Division 01 affects trade packages
Trade assignment from specs General knowledge only — misses naming collisions Document-cited — based on what the spec actually says
Supplementary conditions flagging Summarizes — does not flag risk interactions Checks against structured risk framework
Cross-document citations No — answers not tied to source documents Yes — every answer cites document, section, page
Custom checklists No Yes — 97%+ accuracy on custom checklists
Scope package output No Yes — via Scope Agent, complete bid-ready packages

Who Should Be Using Purpose-Built Contract Review AI

If your team is reviewing one major pursuit per quarter, a manual review process might be manageable. If you're running 15–20 pursuits at once — which most $100M+ GCs are — the review gap compounds fast.

The GCs using Provision are for general contractors who can't afford to staff every bid with a senior estimator doing a full contract read. Provision reviewed more than 50,000 queries in 2026 and found over 1,000,000 risks across real project documents. That volume represents the review capacity of a large preconstruction team — running continuously.

The EllisDon team used Provision on a complex ICI project and documented $1.8M in savings. See the full EllisDon case study.

Cleveland Construction used Provision to speed up their bid review workflow. See the Cleveland Construction case study.

The Real Risk of Using Generic AI for Contract Review

Generic AI gives you the feeling of coverage without the substance. You asked the tool about the contract. It answered. You feel like you reviewed it.

But a summary of the indemnity clause is not a risk analysis. A list of "key terms" is not a flagged list of exposure items. And an answer that's not cited to a source document is not something you can take to a subcontract negotiation or use to defend a back-charge.

Change orders as a share of project cost run 8–14% on commercial work, per Navigant. On projects with weak scope and contract review, that number exceeds 25%. Every point of change order exposure above your baseline is a margin problem — and most of it starts in preconstruction.

The answer isn't to stop using AI. It's to use AI that was built to catch what construction documents actually hide.

Provision's Risk Review is purpose-built for GC contract review. It covers the risks that matter — supplementary conditions, Division 01, trade assignments, notice requirements — and it delivers cited answers you can act on. Book a demo and see it run against your documents.


Frequently Asked Questions

What is contract risk analysis AI in construction?

Contract risk analysis AI reviews construction contract documents — prime contracts, supplementary conditions, specs, and subcontract forms — to identify clauses that expose the GC to financial or schedule risk. Purpose-built tools use structured checklists tuned to construction language. Generic AI summarizes documents without applying a construction-specific risk framework.

Why does generic AI miss risks in construction contracts?

Generic AI reads text but doesn't synthesize across documents, interpret drawing callouts, or apply construction-specific risk logic. It misses cross-document clause interactions — for example, a supplementary condition that modifies a Division 01 requirement in a way that shifts scope responsibility to the GC. Purpose-built tools are trained to flag these interactions.

How accurate is Provision's Risk Review compared to ChatGPT?

Provision's Risk Review achieves 99.5% accuracy on pre-built risk checklists and 97%+ on custom checklists, verified against real construction documents. ChatGPT's accuracy on the same documents is significantly lower, particularly on supplementary conditions, Division 01, and trade-specific scope assignments. The gap grows on complex project sets.

Can Provision read drawings and specs together?

Yes. Provision ingests the full project set — drawings, specs, contracts, RFIs, and addenda — and treats them as a unified document environment. Answers are cited back to the specific document, section, and page. Generic AI tools process text only and cannot interpret drawing content.

What contract risks do GCs most commonly miss?

The most common missed risks include supplementary conditions that expand indemnity or shift "readily inferable" scope, Division 01 requirements that affect trade coordination and cost, trade assignment gaps where specs are silent or use conflicting terminology, and notice requirements that are modified or eliminated by SC clauses.

Does Provision integrate with our existing bid workflow?

Provision is designed for GC preconstruction workflows — bid review, scope package generation, and subcontract buyout. It works alongside your existing estimating tools, not as a replacement. Upload your project set, run the risk checklist, and get cited results your team can act on. See the full workflow in a demo.

Is purpose-built construction AI worth the cost over free tools?

One missed supplementary condition can cost more than a year of software licensing. The $300K lead-lined glass gap and the $400K missed roof cover board documented in the Scope Gap Playbook were both recoverable at review — and both slipped through manual processes. Purpose-built accuracy pays for itself when it catches one item per project.

See what generic AI misses on your contract.

Risk Review checks supplementary conditions, Division 01, and trade assignments at 99.5% accuracy.

See Risk Review

Share

More Articles

AI in Construction

Healthcare Construction Specs Are 2,400 Pages Long — How GCs Are Keeping Up in 2026

By Provision·June 18, 2026
AI in Construction

Plumbing Scope of Work Review: How AI Finds Requirements Estimators Miss

By Provision·June 2, 2026
AI in Construction

How to Evaluate AI Pre-Construction Tools: AGC's 2026 Procurement Framework Explained

By Provision·June 1, 2026