In mid-2026, Suffolk Construction made headlines for its internal AI strategy. The takeaway that landed hardest: data quality matters more than model selection.
That's a blunt statement from one of the largest GCs in North America. And it cuts against how most construction firms are currently approaching AI. Most teams are asking "which model should we use?" Suffolk is asking "what data are we feeding it?"
That reframe has real implications for pre-construction. Specifically, it explains why generic AI tools keep disappointing estimators, and why purpose-built construction AI performs differently.
When Suffolk talks about data quality, they're not talking about clean spreadsheets. They're talking about whether the information an AI model receives is structured, complete, and contextualized enough to produce useful outputs.
In pre-construction, that problem is acute. A typical commercial bid package includes:
These documents don't just contain data. They contradict each other. Drawings conflict with specs. Addenda supersede prior revisions. Scope responsibility sits in three different places depending on which document controls.
A generic AI model can read those documents. But it doesn't know what to look for, how to resolve conflicts, or what a complete scope package is supposed to contain. It lacks the construction context to turn raw document data into a usable output.
This is not a knock on large language models. ChatGPT and Copilot are genuinely capable tools. But "capable" and "construction-ready" are different things.
Here's where the gap shows up in practice:
When you ask a generic AI to summarize a spec section, you get a summary. When you ask Scope Agent to generate a scope package, you get a structured, trade-by-trade breakdown with document references attached. One is a paragraph. The other is a bid-ready deliverable.
Generic AI tools typically handle one document at a time. Pre-construction doesn't work that way. A scope gap between the structural drawings and the mechanical spec won't show up if you're only reading one file. Purpose-built tools ingest drawings, specs, and contracts together and flag conflicts across the full project set.
Addenda issued 48 hours before bid day change everything. A generic AI session has no memory of what came before. Purpose-built tools update the project record as addenda come in. You don't restart. You just review the delta.
Accuracy matters differently in construction than in general knowledge tasks. Getting 90% of a scope package right sounds good. In practice, a missed trade or a misclassified item is a change order waiting to happen. Scope Agent achieves 97% match accuracy on scope item extraction, validated against a dataset of real estimator work. Human estimators working the same exercise scored 91.3% on average, at approximately four days per project. Scope Agent completes the same work in under 60 minutes.
Suffolk's insight goes deeper than "use better data." They're pointing at something structural. Construction documents are not designed for machine reading. They're designed for humans who already know what to look for.
Spec divisions assume you know the CSI hierarchy. Drawing notes assume you know standard abbreviations. Contract language assumes you know what "readily inferable" means and how courts have interpreted it.
A generic AI has none of that background. It processes tokens. It doesn't carry the 20 years of construction context that a senior estimator brings to a document review.
Purpose-built construction AI is trained on that context. It knows that Division 03 governs concrete, that "furnish and install" is different from "install only," and that a scope gap between the civil and structural drawings on a slab detail can mean a $45,000 cost swing on a single line item. That last example is real. It comes from a project documented in Provision's Scope Gap Playbook, drawn from interviews with 200+ general contractors.
Suffolk's data-quality-first argument suggests a two-part framework for evaluating AI in pre-construction:
A tool that reads only specs misses drawing conflicts. A tool that reads only contracts misses scope gaps in the field instruction chain. The full project set, drawings plus specs plus contracts, is the minimum data footprint for pre-con AI to be reliable.
Provision's Chat Agent is built on that principle. It reads the complete project set and returns cited answers in under 20 seconds. Every response links back to the exact page and section it drew from. That's not a UX feature. It's the only way a skeptical VP of Pre-Construction will trust an AI answer on a $40M pursuit.
Data quality isn't just about inputs. It's about whether the output matches how construction teams actually work.
A pre-construction team doesn't need a text summary of a spec. They need a scope package organized by trade, with inclusions and exclusions spelled out, and with every line traceable to a document reference. That's what Scope Agent produces. And it's the output structure that makes the difference between "interesting AI output" and "something I can actually use on bid day."
There's a risk angle here that doesn't get enough attention. When a generic AI misses a scope item, nobody flags it. There's no checklist. There's no audit trail. You find out when the sub calls to say that line item wasn't in their number.
A 2018 study by PlanGrid and FMI found that miscommunication and bad project data drove $31.3 billion in U.S. rework costs annually. That figure is seven years old. The underlying problem hasn't shrunk. Scope gaps remain one of the most consistent drivers of construction disputes in North America. According to the 2024 Arcadis Global Construction Disputes Report, the average North America construction dispute was worth $60.1 million. "Errors and omissions in contract documents" has repeatedly ranked as the top dispute cause in North America across the years Arcadis has tracked.
Generic AI that produces plausible-sounding but unverified scope outputs doesn't reduce that risk. It can add to it. If an estimator trusts a ChatGPT-generated scope summary without checking citations, they're adding a new failure point to a process that already has too many.
Purpose-built risk AI handles this differently. Risk Review identifies risks with 99% accuracy and cites every finding to the exact clause, section, or page. It cuts contract and spec review time by 80%. That's not a claim about which AI model is "smarter." It's about what the tool is built to do and whether its outputs are auditable.
Most GCs evaluating AI in 2026 are still at the demo stage. They've seen a few tools. They've run a test. They're trying to figure out whether AI is worth the investment.
Suffolk's data-quality frame gives you a sharper evaluation lens. Instead of asking "can this tool read a spec?" ask these questions:
These aren't questions a generic AI can answer well. They're the questions that separate tools built for construction from tools adapted to it.
Provision has processed over $100 billion in project value across more than 100,000 documents. That's not a marketing number. It's the data foundation that makes construction-specific accuracy possible. General contractors who want to see how that translates into daily pre-con workflows can request a demo and run it against a live project set.
Suffolk is right. Data quality beats model selection. But for most GCs, the practical implication isn't "build better data pipelines." It's "choose tools that already have construction-quality data and outputs built in."
Generic AI is a starting point. Purpose-built construction AI is what you use when accuracy on a $30M bid actually matters.
If you're building out your pre-construction AI stack, the for general contractors overview lays out how Provision fits into the workflows your team already runs.
In construction AI, data quality refers to whether the information fed to an AI model is structured, complete, and contextualized for construction workflows. A generic AI can read a spec page. A purpose-built tool understands where that spec sits in the full project hierarchy, how it interacts with the drawings, and what a complete scope output should look like.
ChatGPT can read and summarize construction documents. But it doesn't ingest full project sets, doesn't produce structured trade-by-trade scope packages, doesn't handle addenda natively, and has no published accuracy benchmark for construction-specific tasks. For a bid where a missed line item costs $200K or more, that's a significant gap.
Provision's Scope Agent achieves 97% match accuracy on scope item extraction, validated against real estimator work. Human estimators working the same task scored 91.3% on average. Generic AI tools have no published construction-specific accuracy benchmark to compare against. The comparison has to be made on capability and architecture, not a single number.
Yes. Provision's tools update the project record as addenda are issued. You review the changes without restarting your session or reingesting the full document set. Generic AI has no project memory between sessions, so addenda handling requires manual re-input every time.
Suffolk's 2026 AI approach emphasized data quality as a more important variable than model selection. The argument is that feeding high-quality, well-structured data to a capable model produces better results than choosing the most powerful model and feeding it poor data. For GC pre-construction, this validates the purpose-built AI argument over generic tools.
Provision reads the full project set, including drawings, specs, and contracts together. This matters for pre-con workflows like scope package generation, where conflicts often live across document types. DocumentCrunch handles contracts, specs, and document chat but does not read drawings. For end-to-end pre-construction workflows, full project-set ingestion is a requirement, not a feature.
Provision's Scope Gap Playbook is built from 200+ GC interviews and covers the root causes of scope gaps, the habits that prevent them, and trade-specific examples with real dollar figures. You can find the full resource at provision.com/ebooks/scope-gap-playbook.
Provision reads drawings, specs, and contracts together. Generic AI does not.
Book a demoMore Articles
