A quoting desk at a Tier Two supplier outside Grand Rapids can burn a whole morning on one RFQ that looks simple on paper. The drawing is incomplete. The customer wants a price by Thursday. The estimator has three similar jobs in history and none of them match cleanly. That is the job AI is supposed to help with. It is also the job where a confident first draft can quietly push a shop into a bad price.
We see this pattern often enough that it is worth saying plainly. AI is useful in quoting. Trusting the first draft is not. The shops that get value treat the model the way they treat a new estimator on day thirty: useful for pulling history, dangerous if left alone with the customer.
What the first draft is good for
Used well, an AI quoting assistant does three things quickly. It pulls features from the RFQ and drawing package. It finds comparable past jobs from your own history. It drafts a recommended price with a short trail of reasoning the estimator can argue with.
That last part matters more than the number. A draft that says "$18.40 each at 500, based on Job 4417 and Job 4622, plus twenty minutes of secondary ops" is usable. A draft that says "$17.90" with no trail is a coin flip dressed up as software.
The time save is real even when you rewrite half the draft. Searching the ERP for similar jobs, chasing the last change order on a near-match part, and reconstructing setup assumptions from old travelers eats the day. If the tool puts those pieces on the desk in ten minutes, the estimator spends the remaining time on judgment instead of archaeology.
For high-volume part families, that alone can change the queue. A stamped bracket family with dozens of variants, or a machined housing that returns every quarter with a small revision, is fertile ground. The history is thick. The differences are local. The AI is strongest when the past actually looks like the present.
Where the first draft lies
The draft fails in familiar Midwestern ways.
It will treat a "similar" past job as equivalent when the material is different, the tolerance stack is tighter, or the secondary operation that used to be outside is now in-house. It will miss a note buried in a customer PDF that adds a coating, a serialization requirement, or a PPAP package. It will underweight setup time on a short run because the training examples were mostly production volumes. It will sound certain while doing all of that.
We watched a West Michigan machine shop nearly send a quote that was eight percent light because the draft matched on geometry and ignored a callout that forced a slower finish pass. The estimator caught it because he still opens the drawing himself. Another shop in the Metro Detroit supply chain caught a draft that priced aluminum history against a steel revision. The part number had stayed the same. The material had not. The AI had no malice and no sense. It pattern-matched.
These are not exotic failure modes. They are the same mistakes a rushed junior estimator makes. The difference is that the AI makes them at volume, with clean formatting, and without the sheepish look that makes a supervisor check twice.
If your quoting process treats the draft as "probably right unless something looks weird," you will ship some of those mistakes. Weirdness is not a reliable detector. The dangerous drafts look ordinary.
A review habit that actually works
The shops that keep the benefit and cut the risk use a short review ritual. It is boring on purpose.
First, the estimator names the three past jobs the draft leaned on and opens them. Not the summary cards. The actual travelers, scrap notes, and final cost. If the draft cited jobs the estimator would never have used, the draft is already on probation.
Second, they check the assumptions that move money: material, volume break, setup hours, secondary ops, scrap allowance, and freight or packaging when those sit in the quote. Write the assumed numbers next to the draft numbers. Where they disagree, the human wins unless the AI can point to a specific historical reason.
Third, they price the risk of being wrong. A five percent miss on a $2,000 prototype job is a learning expense. A five percent miss on a blanket order that runs for two years is a slow leak. The review depth should track the downside, not the elegance of the draft.
Fourth, they keep a reject log. When a draft is wrong, record why in one sentence: wrong material family, missed heat treat, ignored fixture complexity, over-trusted a 2022 job before a process change. After a month, the log tells you whether the tool is improving your quoting or just accelerating your errors.
This ritual takes fifteen to twenty-five minutes on a normal RFQ. That still beats a half day of hunting. It does not beat clicking "accept" and moving on, which is why people skip it until a bad win shows up in the monthly margin review.
Data the tool needs before you trust it more
AI quoting quality is mostly a data problem wearing a model costume.
If your job history has inconsistent part descriptions, missing secondary ops, or costs that were never closed properly, the draft will inherit that mess and present it politely. Clean the last two years of the part families you quote most before you ask the tool to get smarter. That work is unglamorous and usually higher return than buying another feature module.
Also decide what the tool is allowed to see. Customer drawings and RFQs often contain information your shop should not paste into a random consumer chatbot. Use a setup your IT vendor can stand behind: access controls, retention rules, and a clear answer to where the data lives. Michigan suppliers already live under customer audit pressure. Do not invent a new exposure because the demo was convenient.
Start narrow. Pick one or two part families with clean history and stable processes. Run the assistant there for a quarter. Compare quoted margin to actual margin the old way and the new way. Expand only when the reject log shrinks and the estimator stops treating the draft as a stranger.
What "good" looks like after ninety days
A healthy AI quoting pilot does not mean the first draft ships untouched. It means the first draft is usually directionally useful, the review catches the misses, quote turnaround drops, and win rate on the jobs you actually want either holds or improves.
You should also see fewer surprises when the job hits the floor. If quotes get faster but travelers keep revealing missed ops, you did not improve quoting. You accelerated underpricing.
Owners sometimes ask whether the AI will replace the estimator. In a mid-sized Michigan manufacturer, that is the wrong question. The scarce resource is experienced people who can smell a bad print. The useful outcome is giving those people more of their day back and a tighter second set of eyes on history. The draft is a starting point. The signature on the quote is still a human act, and it should stay that way until your reject log is boring for a long stretch.
If a vendor tells you the first draft is already good enough to send, ask them to sit with your estimators for a week of live RFQs from your actual customers. The conversation gets honest quickly. Quoting is where AI can pay for itself in a shop. It pays for itself through faster, more consistent human judgment, not through blind trust in a tidy paragraph that arrived first.