A senior estimating exec at a multifamily GC told us something that stuck. Some of what AI does now is borderline scary, he said, because people get dependent on generated output without knowing what they're looking at.
He wasn't talking about our product. He was talking about his own team, and the fact that nobody had written down what they were and weren't allowed to do with these tools.
Most GCs are in that position right now. Estimators are already pasting spec sections into chat tools, because it works and nobody told them not to. The policy question isn't whether to allow it. That decision got made for you months ago. The question is what the rules are.
Four rules. They fit on one page, which is the point.
A director of preconstruction at a GC put the bright line better than we could: he'd never want a PM to ask an AI system for a scope of work, print it, and issue it.
Make that the first rule and make it about the signature, not the tool. Whoever's name goes on the document read the document. AI can produce the draft. It can't be the reason nobody checked.
This one rule handles most of the risk, because it puts accountability back where your contract already puts it.
Any AI output that feeds a bid-facing deliverable has to be traceable to a source in the project documents. If a line can't be tied to a drawing detail or a spec paragraph, it doesn't go in the package until someone verifies it by hand.
This is the rule that separates usable tools from unusable ones, and it's worth writing down as a procurement standard rather than a preference. It also gives your team a clean way to reject output: not "I don't trust the AI," but "there's no citation, so I can't check it."
Our Chat Agent answers estimator questions with the page and section attached, at 95% verified accuracy on real estimator questions, in under 20 seconds. We built it that way because the alternative fails this rule.
One person accountable for each place AI touches your process. Not a champion who likes the tool. An owner who answers for whether it was used correctly on the last pursuit.
Without this you get the worst version: two estimators using different tools with different prompts, producing scope packages that don't look alike. A director of preconstruction at a GC described the underlying split in the playbook interviews: "Pre-con is working in the scope sheet world and project management is working in the scopes of work." Ungoverned AI widens that gap instead of closing it.
If your company splits supply and install for rebar, that belongs in a configured ruleset, not in whatever somebody typed into a prompt box that morning.
This is the rule most policies miss, and the one with the longest tail.
If a junior estimator's first two years are spent reviewing generated output, they never build the pattern recognition that makes a senior estimator valuable. They'll be fast and they won't know when the output is wrong. That's a worse position than slow.
So: new estimators do a defined number of scope packages the hard way before they're allowed to work from AI drafts. Pick the number, write it down. Some firms use three projects, some use a year. The number matters less than having one.
McKinsey found 98% of megaprojects run more than 30% over budget. Those overruns aren't caused by people who know their drawings too well.
Two things we'd keep out of the policy.
A tool whitelist. It'll be stale in a quarter and it invites shadow use. Write rules about outputs and accountability, which survive tool churn.
A blanket ban on public chat tools. People will use them anyway and stop telling you. Better to say what can't be uploaded, which is usually anything under an NDA or an owner confidentiality clause, and leave the rest alone.
That's a policy your team will follow, because none of it stops them doing their job. If you want to see what citation-backed output looks like before you write your version, Scope Agent traces every scope line to its source page, and this is how GCs run it.
Yes. Your team is likely using public chat tools already. The policy governs behavior, not procurement.
A blanket ban tends to push usage underground. Restricting what can be uploaded, particularly anything under NDA or owner confidentiality, is more enforceable.
Require a set number of scope packages produced unassisted before they work from AI drafts. The specific number matters less than writing one down.
A citation to the exact page or section it came from, so a qualified estimator can verify it in seconds rather than re-reading the set.
Whoever owns preconstruction quality. It's a scope-accuracy document, not an IT document.
Every scope line traced to the page and section it came from.
See Scope AgentMore Articles
