Sixty-eight percent of construction leaders deployed AI and did not get what they expected. That is the finding buried in Autodesk's 2026 State of AI in Construction report. Only 32% said they met their AI goals.
If you are a VP of Pre-Construction or Chief Estimator reading that number, it probably does not surprise you. You have seen the demos. You have heard the pitches. You may have even run a pilot that went nowhere.
But here is the thing: the problem is not AI. The problem is where and how construction firms are deploying it.
Preconstruction is the one area where AI has a documented, measurable case for ROI. Firms that deploy purpose-built tools on pre-con workflows are not in the 68% club. This article explains why.
Generic AI tools are built for general knowledge work. They summarize documents, draft emails, and answer broad questions. They were not built to read a 2,000-page project specification, cross-reference it against 400 drawing sheets, and produce a trade-by-trade scope package.
When GCs use general-purpose AI for construction document work, they run into the same wall every time:
These are not edge cases. They are the core of pre-con document work. And generic AI cannot do any of them reliably.
So when a GC deploys Copilot or ChatGPT on bid documents and gets inconsistent results, they write off "AI" as a category. That is the wrong conclusion. They deployed the wrong tool.
Pre-con is not like project delivery. The document volume is concentrated. A single pursuit can involve thousands of pages of drawings, specifications, contracts, addenda, and RFIs. All of it needs to be read, cross-referenced, and acted on in days, not weeks.
The manual work is also quantifiable. A single bid requires 30–40 hours of estimator time just to build a scope-of-work package. Contract and spec review adds another layer. For general contractors running 20, 30, or 50 pursuits a year, the math compounds fast.
And the cost of getting it wrong is documented. According to the Arcadis Global Construction Disputes Report, the average North American construction dispute in 2024 was worth $60.1 million. Errors and omissions in contract documents have repeatedly ranked as the number one dispute cause in North America.
That is the environment preconstruction teams operate in. High document volume, tight timelines, real financial consequences for errors, and not enough people to cover it all manually.
AI built for this environment does not look like a chatbot. It looks like a system that can ingest a full project set, understand the relationships between documents, and produce structured outputs that are actually usable on bid day.
One of the clearest signals of a pre-con process failure is a scope gap that shows up after award. A line item nobody priced. A trade boundary nobody defined. Work the sub says is not in their number.
These gaps are not random. They follow a predictable pattern. The Scope Gap Playbook, built from 200+ interviews with GC preconstruction teams across North America, found that the same habits produce the same gaps, project after project.
The most common anti-pattern: writing "as per plans and specs" in a scope sheet and moving on. That phrase does not define scope. It defers the conflict to later, usually after the sub has already underbid and is looking for a way out.
As one Pre-Construction Lead at a Top-ENR Canadian GC put it: "It's descriptive: bread, put it on a plate, use the open jar. You have to get to that level of detail or else they'll just be like, 'you didn't tell us that.'"
The dollar examples from the playbook tell the same story:
None of these were typos. They were systematic failures in the scope-writing process. And they are exactly the kind of problem that purpose-built AI is designed to catch before bid day.
The gap between "AI tools we bought" and "AI goals we met" is a capability gap. Construction AI needs to do things that general AI simply does not.
Here is where purpose-built tools are delivering measurable results for GC preconstruction teams in 2026:
Provision's Scope Agent reads the full project set: drawings, specs, and contracts together. It generates a complete scope-of-work package in under 60 minutes. The same work takes a senior estimator 30–40 hours manually.
In internal validation, Scope Agent achieves 97% match accuracy on scope item extraction. The human estimator baseline on the same exercise is 91.3%, at roughly four days per project. That gap is not about replacing estimators. It is about giving them a better starting point.
See the EllisDon case study for what scope generation looks like inside a large GC's pre-con workflow.
Provision's Risk Review runs a risk checklist across contracts and specs with 99% accuracy on risk identification. Every risk is cited to the exact clause, section, or page. It delivers an 80% reduction in contract and spec review time.
Compare that to the five-minutes-before-bid review that happens on most pursuits. That is not a process. That is hoping nothing gets missed.
EllisDon reports $1.8M CAD per year from contract risk review and reduced manual work. See the EllisDon case study for the detail behind that annual figure.
The Arcadis data on $60.1M dispute values exists for a reason. Contract risk that is not caught before award does not go away. It becomes a dispute.
Provision's Chat Agent answers questions about drawings, specs, contracts, RFIs, and addenda in under 20 seconds. Every answer is cited back to the source document. Accuracy on real estimator questions is 95%.
The alternative is an estimator spending 20 minutes hunting through a 1,400-page spec book for a single detail. Multiply that by the number of questions per pursuit and by the number of pursuits per year. The time cost is not trivial.
Provision has processed 66,000 documents and answered more than 50,000 queries across $100 billion in project value. Those are not pilot numbers.
The Autodesk 2026 report does not just surface a stat. It surfaces a structural problem. Generic AI is being deployed on workflows it was not designed for, and construction leaders are absorbing the disappointment as a signal that AI does not work.
The architecture of a purpose-built construction AI is fundamentally different from a general-purpose model:
These are not feature differences. They are capability differences. A tool that cannot read drawings cannot help you catch a scope gap between the civil drawings and the architectural finish schedule. No prompt engineering fixes that.
The firms meeting their AI goals in 2026 are not necessarily running the most sophisticated technology stacks. They are applying AI to the right problems with the right tools.
The pattern across the GCs seeing results:
The NAC case study and the Cleveland Construction case study both follow this pattern: specific workflow, specific tool, specific result.
The 2018 PlanGrid and FMI study "Construction Disconnected" found that miscommunication and bad project data drove $31.3 billion in U.S. rework costs. That figure is from 2018. The document complexity has only increased since then.
Pre-con is where bad data enters the project. A scope sheet that says "as per plans and specs" does not define work. It defers a future argument. A risk clause nobody flagged before execution becomes a dispute after. A missed trade boundary on bid day becomes a change order in the field.
As one Estimating Manager at a Canadian ICI GC said during research for the Scope Gap Playbook: "We have less subs who just kind of a gentleman's agreement. They've become more quick to clarify that we're not including that one piece of scope."
The margin pressure is real. The sub behavior is real. The document volume is real. AI that was built to handle this environment can help. AI that was not built for it will add noise, not signal.
If your firm is in the 68% that has not met its AI goals, the first question is not "should we invest more in AI?" The first question is: "Are we deploying AI on the right workflow with the right tool?"
For most GCs, the answer to that question starts in preconstruction. If you want to see what that looks like in practice, book a demo with Provision and bring a live pursuit. That is the fastest way to know if it fits your workflow.
Autodesk's 2026 State of AI in Construction report reports this figure across the construction leaders it surveyed, general contractors among them. Provision's reading of the cause, drawn from pre-con work with GCs rather than from the report itself, is that general-purpose AI is being deployed on construction-specific workflows. Generic tools lack the ability to read drawings, cross-reference full project sets, handle addenda, and produce structured trade-by-trade outputs. Without those capabilities, results are inconsistent and ROI is hard to prove.
Pre-con has concentrated, document-heavy tasks with clear outputs: scope packages, contract reviews, spec searches. The manual time cost is measurable (30–40 hours per bid scope package), and the error cost is documented. That makes ROI easier to identify and faster to prove than in project delivery workflows.
Generic AI cannot ingest full project sets (drawings plus specs plus contracts together), cross-reference documents, handle addenda without restarting, or produce structured scope outputs. Purpose-built tools like Provision are architected for these exact tasks. That architecture difference determines whether AI is useful on bid day or just impressive in a demo.
Scope Agent achieves 97% match accuracy on scope item extraction, compared to a 91.3% human estimator baseline on the same exercise. Risk Review identifies risks with 99% accuracy, with every risk cited to the exact clause, section, or page. Chat Agent answers real estimator questions with 95% verified accuracy. These stats are product-specific and measured separately.
The Scope Gap Playbook, based on 200+ GC interviews, documents recurring gaps: missed roof cover boards, lead-lined glass omitted from hospital imaging suites, wood-flooring conflicts between drawings and finish schedules, and trade boundary ambiguities written off as "as per plans and specs." Purpose-built AI catches these by reading the full project set and cross-referencing conflicting information before bid day.
Yes. Provision processes addenda as part of the live project set without requiring users to restart the ingestion. Generic AI tools typically require users to re-upload documents and re-run queries manually when addenda are issued, which is impractical on a real bid timeline.
The fastest way is to bring a live pursuit to a demo. Provision's team runs the tool on your actual documents so you can see scope extraction, risk flagging, and document Q&A on your project, not a generic sample set. Book a demo here to get started.
Bring a live pursuit. We run Provision on your documents so you can judge the output yourself.
Book a DemoMore Articles
