Construction Dive ran a piece in September 2026 that cut through most of the AI noise in one sentence: the question leaders should be asking is not "does this tool use AI?" It is "what data can it actually see?"
That framing is exactly right. And for GCs running preconstruction, it has a specific, testable answer.
The data that matters in pre-con is not stored in one place. It lives across drawings, specifications, contracts, geotechnical reports, and addenda. Those documents refer to each other constantly. A scope gap, a conflicting requirement, or a commercial exposure almost never announces itself in a single file. It hides in the gap between two files.
An AI tool that reads one document at a time cannot see that gap. It will give you an answer. The answer may even look credible. But it is built on incomplete information, and in a $50M project, incomplete information is not a small problem.
Connected data means the AI can see your full project set as a single body of evidence. Not "here is a contract, tell me the risk clauses." Not "here is a spec section, write me a scope." Connected means: drawings, specifications, contracts, and addenda, all ingested together, all cross-referenced against each other when it generates an output.
That matters because GC preconstruction is a cross-document discipline. Consider three common problems:
None of these are edge cases. They are routine. The $300K lead-lined glass omitted from a hospital imaging suite, later absorbed by the GC under "readily inferable" language, is exactly this kind of gap. It did not hide in a vacuum. It hid in the space between a drawing note and a spec callout. No single-document AI would have caught it.
ChatGPT and general-purpose AI tools are not built for a 10,000-page project set. They are built for a conversation. Upload a contract and ask a question: they can help. Upload your full drawing set, your specifications, your supplementary conditions, and your addenda, then ask a question that depends on all four: they cannot hold all of that in context at once.
This is not a criticism of those tools. They are good at what they are designed for. The problem is deploying them in a workflow they were not designed for and trusting the output.
A Chief Estimator at a mid-market GC described the risk plainly: "If you miss anything, they'll bill it." That quote comes from Provision's Scope Gap Playbook, which pulled it from interviews with more than 200 general contractors. The subs know their scope. The question is whether your AI tool knows it too, or whether it only knows what was in the one file you uploaded.
General AI tools also have a fabrication problem on technical documents. In a head-to-head test on a live hospital project's plumbing drawings, Claude produced 18 fabricated scope items. Scope Agent, Provision's purpose-built scope tool, produced zero. That is not a minor accuracy difference. A fabricated scope item goes into a bid package. It goes to a sub. The sub prices it, or excludes it, or prices it wrong. Any outcome creates exposure.
If you are evaluating AI tools for preconstruction, these three questions will separate the tools that can help from the tools that will create liability.
This is the connected data question. A tool that reads one file at a time cannot surface cross-document conflicts. Ask the vendor to show you a workflow where a drawing conflicts with a spec section. If they cannot demonstrate that, the tool is not designed for GC preconstruction.
Provision's Scope Agent reviews the full document set together. It does not ask you to upload one file, get an output, upload the next file, and reconcile the results manually. That reconciliation step is exactly what the tool should be doing for you.
Addenda land at the worst possible time. Three days before bid. Sometimes one day before. A tool that requires you to restart your review every time an addendum arrives is not useful on a live pursuit. It adds work instead of removing it.
Scope Agent handles addenda automatically. The review updates without starting over. That matters most when time matters most.
An AI output without a citation is an opinion. In preconstruction, you need to know exactly where a finding comes from so you can verify it, act on it, or push back on it. "The AI flagged a risk" is not useful. "The AI flagged a risk on page 47, clause 12.3 of the supplementary conditions" is actionable.
Provision's Risk Review ties every risk to the exact clause, section, or page where it appears. Chat Agent returns answers with source citations in under 20 seconds. RFI Agent ties each potential issue to the exact specification page or drawing where it found it. No finding floats without a source.
When an AI tool can read the full project set together, it changes what you can do before bid day.
A scope of work built from spec sections alone misses what the drawings say. The drawings often contain the most specific requirements: material callouts, installation notes, coordination requirements. A scope package that ignores the drawing set is incomplete by definition.
Scope Agent generates trade-specific scope packages from both. It reads drawings, specs, tables, and notes together and generates a coordinated output by trade. The result is a scope that reflects what the project actually requires, not just what the spec sections say in isolation.
ProWest Constructors cut estimating time by 50% on the first project scoped with Provision, saving three to four days. The Director of Project Development put it plainly: on a progressive design-build, early scope definition narrows GMP risk down to what is truly unforeseen. That is the value of a scope package built on connected data.
Contract risk does not live only in the contract. It lives in the relationship between the contract and the specifications. A contract that says "work includes all items readily inferable from the drawings" is only as clear as the drawings themselves. If the drawings show something the specs do not mention, and the contract says it is included, that is a risk. You cannot find it by reading the contract alone.
Risk Review reads drawings, specs, and contracts together. It surfaces conflicts that only appear across documents. EllisDon used it to save six hours per contract review, across 40+ contract analyses. That is 240+ hours recovered, on work that required reading across multiple document types every time.
An estimator on a live pursuit does not ask generic questions. They ask questions like: "Does the structural spec conflict with the architectural drawings on the slab assembly at grid line C4?" That question requires the AI to hold both documents in context and reason across them.
Chat Agent answers construction document questions across drawings, specs, contracts, RFIs, tables, and addenda. It handles up to 100,000 pages per project. Every answer links to the exact page or section where it found the information. EllisDon saved 200 hours per month on document queries using the chat function. That is not a rounding error. That is a full-time role's worth of time recovered each month.
A 2018 study by PlanGrid and FMI found that miscommunication and bad project data cost the U.S. construction industry $31.3 billion in rework in a single year. Twenty-six percent of that rework traced back to communication breakdowns. Twenty-two percent traced back to inaccurate project data. Together, nearly half of all rework came from information that was missing or misread.
That study is eight years old. The document sets have gotten larger, not smaller. The coordination requirements have gotten more complex. The window between addendum and bid has not gotten wider.
The only way AI helps with this problem is if it can read the full picture. A tool that reads one file at a time gives you a partial picture and no way to know what it missed. A tool that reads the connected document set gives you a complete picture and a source citation for every finding.
That is the difference between useful AI and AI hype. It is not about the model. It is not about the interface. It is about what data the tool can actually see when it generates an answer.
If you are running preconstruction at a GC and evaluating AI tools, that is the question to put to every vendor in the room. Show me a cross-document conflict. Show me how you handle an addendum. Show me where this finding came from in the documents.
If they cannot answer all three, the tool is not ready for your workflow. If they can, it is worth looking at how it fits into your bid process.
Provision was built for GC preconstruction from the start. It reads the full project set, handles addenda without restarting, and cites every output to its source. See how it works on a real project: request a demo or explore how Scope Agent builds trade-specific scope packages from your documents.
Connected document data means an AI tool can read your full project set together: drawings, specifications, contracts, and addenda, and reason across all of them at once. Most general AI tools read one file at a time. GC preconstruction requires cross-document context because scope gaps and risk exposures almost always live between documents, not inside a single one.
ChatGPT and similar general tools are not built to hold a full construction project set in context. They handle individual documents well but cannot cross-reference a drawing against a spec section against a contract clause in a single query. On a live hospital project, ChatGPT's underlying model produced 18 fabricated scope items on the same drawing set where Provision's Scope Agent produced zero.
Purpose-built AI for GC preconstruction cuts scope admin by up to 75% and reduces contract review time by 80%. EllisDon saved six hours per contract review and 200 hours per month on document queries. ProWest Constructors cut estimating time by 50% on the first project scoped with Provision, saving three to four days on a single bid.
Provision's Scope Agent ingests addenda automatically without requiring you to restart the full document review. When an addendum modifies a spec section or drawing, the review updates to reflect the change. This matters most in the final days before bid, when addenda land late and review time is shortest.
General AI quality is measured on comprehension of a single document. Preconstruction AI quality depends on accuracy across a full project set, including cross-document conflicts, addenda, drawing notes, and table data. A tool can score well on a single-document test and still miss a $300K scope gap that only appears when you read the drawing and the spec section together.
Yes. Every output from Provision's products links to its exact source. Risk Review cites the clause, section, or page. Chat Agent returns source links in under 20 seconds. RFI Agent identifies the exact specification page or drawing where each potential issue was found. No finding is generated without a traceable source in the document set.
Ask three questions: Can it read the full project set at once, not one file at a time? Does it handle addenda without restarting the review? Does it cite every finding to an exact page or clause? A tool that cannot answer yes to all three is not designed for GC preconstruction workflows. It is a general tool applied to a specialized problem.
Provision reads drawings, specs, and contracts as one project set. Every finding is cited to its exact source.
See it in actionMore Articles
