Compare quotes with AI: contractor, car repair, and insurance

Three contractors walk through the same kitchen. Three quotes arrive. One is eight neat lines and a total. One is a single paragraph and a bigger total. One is six lines, two hourly rates, and a smaller total with an asterisk. You lay them side by side on the table and realize you cannot compare them, because they do not describe the same job.

That is the normal state of quotes, and it is the problem this post solves. Paste all three into an AI model, ask for one table with the same rows for every vendor, and the documents finally answer the same question. Then ask what each one leaves out, and you get the questions to bring back to each vendor before you sign. The routine takes ten minutes, cost us about two cents when we ran it below, and works the same way for a kitchen, a transmission, or a home insurance renewal.

Quotes are written to be hard to compare

Nobody sets out to confuse you. Each vendor writes the estimate the way their business thinks: a cabinet shop itemizes cabinets, a design-build firm sells a finished room, a budget installer prices the parts it controls and bills the rest by the hour. The result is three honest documents that share almost no structure. The Federal Trade Commission’s contractor guidance anticipates exactly this: get multiple estimates, don’t automatically pick the lowest bidder, and “ask for an explanation if there’s a big difference among the estimates.”

The explanation is usually hiding in what a quote does not say. A total that excludes permits is not lower than one that includes them; it is incomplete. An hourly rate for electrical work is not a price; it is a promise to send a bill. A lump sum that includes cleanup is carrying a cost the other two left on your driveway. Until every quote is forced to speak the same rows, the totals at the bottom are not comparable numbers. They are three different sentences that happen to end in dollar signs.

People have always solved this with a legal pad: draw a grid, copy each line into the right box, write “?” where a vendor said nothing. It works, and it takes an evening. The reason to hand it to a model is not that the model knows kitchens. It is that normalizing three documents into one grid is a reading-and-sorting task, which is the thing language models do most reliably, and the “?” marks are the whole value.

Ask for a table, not a verdict

The prompt matters more than the model, and the most important word in it is what you leave out. Do not ask which quote to choose. A model asked for a verdict will give you one, confidently, on the basis of three documents that you have just established are not comparable. Ask for structure instead. Here is the prompt we used, trimmed only for length:

Below are three quotes for the same kitchen remodel. I am not asking which one to pick.

1. Put them into one comparison table with identical rows for every quote: Demolition and disposal, Cabinets, Countertops, Backsplash, Plumbing, Electrical, Installation labor, Painting, Permits, Cleanup, Warranty, Schedule, Deposit, Total. Where a quote gives no number for a row, write "not listed" or "included in lump sum" or "hourly", whichever is true. Do not invent or estimate any number.
2. Under the table, list for each quote what is missing or unclear that I would need before the three prices are comparable.
3. Give me exactly five questions to ask each contractor, specific to that quote.

[the three quotes, pasted in full]

Three details do the work. Naming the rows yourself keeps the model from inventing a taxonomy that flatters one vendor. Telling it which words to use for a gap (“not listed,” “included in lump sum,” “hourly”) turns silence into a visible cell instead of a blank. And “do not invent or estimate any number” is the instruction that keeps the table honest, because a model left to its own devices will helpfully fill an empty cell with a plausible figure. We wrote three sample quotes for a mid-sized kitchen with invented company names, ran the prompt once, and kept the first answer.

Read the amber cells before the totals. The cheapest quote has four of them, plus two hourly lines with no hours attached, and a 14-day expiry that the model flagged on its own as “creating time pressure.” The most expensive quote has almost none, because the lump sum absorbed everything, which is a different problem: the model noted it is “impossible to verify what share of the $29,150 covers each trade.” The middle quote looked the most complete and still hides a 50% deposit, which the model computed as $12,400 up front. None of that is a verdict. All of it is what you needed to know before forming one.

The missing line is the expensive line

Every consumer-protection agency that writes about quotes ends up writing about the same missing lines, because those are where the money moves after signing. The FTC’s list of what a written estimate should carry is short: a description of the work, materials, a completion date, and the price. Its list for the contract is longer, and the additions are the usual escape hatches: the contractor’s license number, an estimated start and completion date, and how changes will be handled.

California writes the same rules into law, which makes them a useful checklist even outside the state. The Contractors State License Board’s contract rules require a written contract for any home improvement job over $500, cap the down payment at $1,000 or 10% of the price, whichever is less, forbid progress payments that run ahead of completed work, and require every change in price or scope to be a written change order signed by both sides before the change happens. Hold those four rules against the sample quotes: one wanted 50% down, one wanted 40% down, and only one described a payment schedule tied to work actually done.

The practical move is to paste a list like that into the prompt as a second pass: “Check each quote against these items and tell me which are addressed, which are missing, and which are ambiguous.” The model did most of this unprompted in our run. It caught that one quote’s backsplash line was a materials allowance with no install labor, that “brought to code” described no actual scope, and that “unforeseen conditions” had no cap and no notification process. Those are the lines that turn a $15,350 quote into a $22,000 invoice, and they were all visible before a single call was made.

Turn the gaps into five questions per vendor

The last part of the prompt is the one people skip and the one that changes the conversation with the vendor. A gap in a quote is not an accusation; it is a question that has not been asked yet. Asking the model for exactly five per vendor, specific to that quote, produces a short list you can read over the phone, and it keeps the vendor answering about their own document rather than about the competition.

The questions our run produced were the kind a careful friend in the trade would ask. For the itemized quote: who schedules inspections and corrects failures if the owner pulls permits, and does that change the warranty? For the lump sum: an itemized breakdown, the exact progress milestones, and a spec sheet for the cabinets. For the budget quote, the model asked the question the whole exercise exists for: “What does ‘unforeseen conditions’ cover, what is your process for notifying me before proceeding, and is there a cap on additional charges?”

Good vendors welcome these calls. A firm that has priced the job properly can answer in a minute, and the answer often makes their quote look better, not worse: the lump sum that seemed high may turn out to include the permits, the cleanup, and the warranty the others left out. A vendor who bristles at “is there a cap on additional charges” has told you something the quote did not. Write the answers into the table as new rows, and the three columns finally describe the same job.

Car repair and insurance follow the same routine

The kitchen is the clearest example because the quotes are longest, but the routine transfers to anything with an estimate, and the missing lines simply change names. For a car, the FTC’s auto repair guidance says a written estimate should identify the condition, the parts, and the anticipated labor charge, and that the shop should have to contact you before exceeding a stated amount. It also notes there is “no standard warranty on repairs,” and that parts come in three kinds, new, rebuilt, and salvage, each with its own price and guarantee.

So the rows for three repair quotes are different, but the idea is identical: parts by type, labor hours and rate, diagnostic or tear-down fee, the approval threshold, warranty in months and miles. California’s Bureau of Automotive Repair turns most of these into rights: the written estimate before work, your authorization before any additional repairs, the return of replaced parts if you ask before authorizing, and an invoice that lists every part as new, used, or rebuilt. The underlying statute, section 9884.9, puts it in one sentence: no charge “in excess of the estimated price” without the customer’s consent. Those are worth remembering because repair is not a small line. AAA’s 2025 driving-cost study puts maintenance, repair, and tires at 11.04 cents per mile over five years and 75,000 miles, about $1,650 a year on top of $1,694 for insurance.

Insurance is the case where the table earns its keep most, because the totals are designed to be the only thing you compare. The National Association of Insurance Commissioners’ homeowners shopping worksheet is, literally, a three-column grid: dwelling limit, personal property limit, replacement cost or actual cash value for each, liability, loss-of-use, ordinance-or-law coverage, and then the deductibles, flat, percentage, and wind or hail, on separate lines. It recommends at least three quotes. Paste three declarations pages into the prompt with those rows and the model builds the worksheet for you, and it will catch the thing people miss: a quote that is $300 cheaper because it pays actual cash value on the contents, which the NAIC’s consumer guide describes as paying repair costs “after the insurer subtracts an amount based on the home’s or personal property’s age and the wear and tear.”

The stakes for reading insurance quotes properly are rising. The NAIC’s first national analysis of seven years of homeowners data, published August 5, 2026, found inflation-adjusted premiums up between 18.3% and 43.3% depending on region from 2018 to 2024, and insurer-initiated non-renewals up between 96% and 216%. More people are shopping under pressure, which is exactly when a cheaper total with a worse deductible looks attractive.

Keep the account numbers out, and keep the decision

Two cautions, both practical. The first is privacy. A contractor quote carries your address; an insurance quote carries your policy number, date of birth, and sometimes a Social Security number; a repair estimate has your plate and VIN. None of that helps the comparison, so strip it before pasting. Replace names with “Quote 1, 2, 3,” black out identifiers, and keep the numbers. If you would rather not paste the documents into a cloud service at all, this is a task a local model handles comfortably: the input is a few pages of text and the output is a table, well within what a model running offline on a laptop does well. We have covered what cloud chat services keep before; for documents with account numbers on them, the safest policy is not to send them.

The second caution is about what the model does not know. It has never seen your kitchen, does not know local labor rates, and cannot tell whether $9,400 for semi-custom cabinets is fair in your city this year. If you ask it to, it will guess, fluently. That is why the prompt bans invented numbers and why the output should be read as a map of the documents, not an appraisal of the job. The same goes for licenses: the model can tell you a quote lists no license number; it cannot tell you whether the one that does is valid. Check that with your state licensing board, which takes two minutes and is the one step in this routine no software should do for you.

Kept inside those lines, the tool is doing something modest and genuinely useful: reading three documents more carefully than a tired person does at 10 p.m., and refusing to let a total stand in for a scope. The decision stays with you, better informed and about twenty minutes sooner.

Three quotes, one table, five questions

The routine fits on an index card. Get three quotes, because two gives you a coin flip and one gives you nothing. Strip the identifiers. Paste all three into a model with the rows named, the gap words defined, and invented numbers forbidden. Read the amber cells before the totals. Then take the five questions per vendor back to each of them, write the answers into new rows, and only then look at the bottom line.

When we ran it on the three sample kitchen quotes above, the whole pass took 29 seconds and cost $0.023, and the cheapest quote came out of it as the one with the most unknowns rather than the best price. That is the honest result of a comparison, and it is the one the quotes were shaped to prevent you from seeing. If you have ever chosen a vendor by the number at the bottom of the page, the table is worth the two cents. The next time three envelopes land on the table, you will know what to ask before you open them.

Disclaimer: This article is general information, not legal advice, and reading it creates no attorney-client relationship. Laws, regulations, and court rulings summarized here reflect sources available as of September 2026 and may have changed. Consult counsel licensed in your jurisdiction before acting on any of it.