People imagine AI as the new senior estimator who reads the contract for you. That's the marketing version. The actual version is closer to a smart intern with a fast highlighter and no judgment about what the highlights mean.
That distinction is the whole article.
What the Marketing Says
A model reads the RFP. It finds the indemnity clause that's too broad. It flags the warranty extension. It spots the O&M obligations buried in a spec section nobody opens. It writes up a clean summary. The estimator saves four hours.
True. All of that happens.
What the Marketing Skips
The model doesn't know your bonding capacity. It doesn't know whether the GC is a hard-line client or a flexible one. It doesn't know which of the four risks it flagged is actually a dealbreaker for your firm and which is standard for the market you're working in.
The model also doesn't know which jobs you're trying to win. A flagged risk on a pursuit you'd walk away from anyway is a non-event. The same flagged risk on a pursuit your project executive has been chasing for two years is a strategic decision that involves three people and a phone call.
What AI Should Be Asked to Do
Sweep the RFP, the specs, the supplementary conditions, and any attached owner standards. Pull out anything that deviates from your company baseline. Categorize it.
The categories are not vague. They're specific:
- Indemnity language broader than your standard.
- Warranty periods longer than your defaults.
- O&M and lifecycle obligations that aren't in the base spec.
- Safety requirements above industry norm—full-time site safety, additional certifications, extra PPE.
- Insurance limits or endorsements that don't match what you typically carry.
- Liquidated damages amounts and triggers.
- Schedule constraints that imply acceleration cost.
The output is a one-page Risk Snapshot for the estimator. What's flagged. Why it matters. What questions to ask the GC. Who internally needs to weigh in—legal on indemnity, ops on safety staffing, the PX on schedule.
Then a draft qualifications list anchored to the flags. Not a final qualifications list. A draft.
What AI Should Not Be Asked to Do
The commercial impact call. The model can tell you a warranty is two years longer than your standard. Only the estimator can decide whether your margin can absorb it, whether the warranty cost goes in the number, or whether you push back on the language.
The contract strategy. The model can identify risky indemnity. The PX or the BD lead chooses what to do about it—qualify, redline, accept, or walk away. None of those are pattern recognition.
The relationship read. The model has no memory of the GC's last six bids with you. It doesn't know that this project executive softened on warranty language two years ago when you walked them through it. It doesn't know that this owner's rep is new and not yet trusted by the GC's PX. Humans carry that.
The final language. AI drafts. A human signs. Anything that touches price, scope, or schedule needs the signature. There is no shortcut for this and there is no model that should write the qualification that ends up on the executed contract.
The Compromise That Actually Works
Use the model first. It runs in minutes. It reads more carefully than a tired senior estimator at 9 p.m. It doesn't get bored on page 70. The Risk Snapshot lands in the estimator's inbox before the kickoff meeting.
Then the human reads it with project context. The PX weighs in. Legal weighs in if needed. The qualifications list gets shaped by people, not by a draft.
The estimator stops spending the first three hours of every pursuit highlighting clauses by hand. Those three hours go into the actual work—pricing, sub conversations, scope strategy, the parts of the job that win or lose work.
That's the trade. Speed up the scan. Slow down the decisions. The teams that get this balance right will move faster and take on better-qualified risk than the teams that either ignore the tools or trust them too far.