Creating quotes with AI: Where do the prices come from?

A customer wants two rooms repainted. The enquiry includes photographs and a preferred date, so an AI system can quickly produce something that looks like a quotation. The wall area is still missing. Nobody knows whether the old coating needs preparation, and the furniture has to go somewhere. Those details decide the work and the price. Good automation keeps them visible instead of filling the gaps with confident prose.

For a small business, automated quote preparation can mean something quite practical: collect the facts, match approved services, fetch current prices and present a draft that someone can check. That removes repetitive assembly work. It also gives the estimator a clearer view of what remains uncertain. Asking a language model to invent a plausible total is a different activity, and a poor foundation for a customer commitment.

Choose a service you already know how to price

Start with a recurring job whose components your team can describe consistently. A defined maintenance visit or a standard installation has a better starting point than a renovation estimated from two dark photographs. You need a reliable link between the requested work, the quantity and the conditions under which your price applies. Polished wording will not establish that link for you.

Look at a handful of completed quotations. Mark every decision that required specialist judgement. Did someone have to measure the site? Did access change the required equipment? Was an existing customer entitled to a special price? If nearly every line depends on unresolved questions, the first useful tool may collect information and draft follow-up questions. Producing the full commercial document can wait.

Check the software already in use before buying another layer. For example, bexio documents separate manual and catalogue-linked positions. Existing structures may support a controlled drafting process. That is not evidence that your account already includes a working AI integration. Confirm the actual edition, available interfaces and permissions with the people responsible for your installation.

Decide what a quote line must contain

Give the automation a small, explicit input contract. An estimator should be able to read it without understanding the implementation. For each line that can reach a customer, specify the information that must be available and where it comes from. Our suggested starting fields are:

• An approved service identifier and description.
• A quantity, its unit and the source of that quantity.
• The price source, its version and the period it applies to.
• Permitted surcharges, discounts and customer-specific conditions.
• Included work, exclusions and unresolved prerequisites.
• The person or role allowed to decide an exception.

Information quality matters as much as its presence. A measurement confirmed on site differs from a customer estimate or a quantity copied from an old job. Keep that distinction in the internal draft. A system should not silently upgrade an estimate into a confirmed measurement just because it has found a number that fits a required field.

Units deserve their own check. Twelve square metres of wall and twelve metres of skirting board can both appear naturally in a painting enquiry. The calculation needs firm rules about which unit belongs to each service and which conversions are permitted. A language model may help interpret the wording. The permitted calculation should remain explicit and reproducible.

Put the arithmetic where the reviewer can see it

Here is a deliberately invented maintenance example. These figures are teaching assumptions, not AlpenAgent prices or a claim about market rates: three labour hours at CHF 95, two material units at CHF 42, and a CHF 35 travel charge. The subtotal is CHF 404 before tax. Your accounting rules must determine the tax treatment and rounding that actually apply to the transaction.

Suppose an authorised ten per cent discount applies to labour alone. It removes CHF 28.50, giving a new subtotal of CHF 375.50. Applying that percentage to the entire document would produce a different amount. Show the discount basis in the review screen. A reassuring sentence about a loyalty discount tells the estimator much less than the underlying line and calculation.

Now remove the confirmed material quantity. The draft should raise a question or present an explicitly labelled assumption for approval. It should not preserve two units because the last job used two, or enter zero and make the document look complete. This small variation makes a useful acceptance case: the automation must distinguish a valid calculation from insufficient information to calculate.

Give prices an owner and an effective date

Choose one authoritative price source for the pilot. It might be your ERP, a maintained item catalogue or a controlled spreadsheet used for a limited trial. Previous PDFs can supply examples of wording. They are a risky substitute for current commercial data because they may contain negotiated discounts, unusual scope or buying costs that have since changed.

A new catalogue version also needs a rule for unfinished drafts. Should an unapproved quote be recalculated before sending? Does an explicitly promised price hold? Who can override the standard conditions? These are commercial decisions for your business. A model should not infer them from a customer writing “same arrangement as last time”.

Display the price version in the internal review area. As a concrete software example, the bexio project-to-quote guide describes transferring estimated time and a configured project rate into quote positions. Whether your implementation has dependable values in those fields requires checking. A documented field is only useful when someone maintains the information it contains.

Approval belongs to a particular version

An estimator approves a draft. Later, a quantity changes or a refreshed material price enters the document. If that change affects what was approved, the revised version needs another review. Otherwise approval of one document becomes permission to send a different one. An email saying “approved” provides a weak safeguard when it cannot be tied to the exact content.

Connect the decision to a document identifier and version. Present the recipient, service lines, total, unresolved issues and important conditions together. Keep the reviewer close to the commercial facts. A button buried among technical messages encourages an administrative click rather than a meaningful check. Corrections should remain visible, along with the reason for an exception.

Our article on drafting, approval and execution explains the general pattern. Quotations add specific questions: who may approve a discount, confirm a delivery date or change an exclusion? Decide those permissions before connecting the send action. An instruction that “a human will look at it” leaves too much of the actual work undefined.

Use awkward enquiries as your acceptance set

A complete standard request demonstrates the happy path. Add requests that must remain incomplete: missing quantity, contradictory dimensions, expired price data, an unauthorised discount and a service outside the catalogue. Include the wrong customer record and a line changed after approval. Write down the expected result before running them. The tool should not earn a pass merely because every test produces a PDF.

Replay the same enquiry as well. A retried form delivery must not quietly create two quotations and two outgoing emails. Use a stable business reference and inspect the target system after an ambiguous result before trying the operation again. The reasoning is covered in our guide to preventing duplicate agent actions.

Ask the person who normally estimates the work to assess each result. Was the right service suggested? Were missing measurements obvious? Could the reviewer trace the price and reject an unsuitable position quickly? Measure the total handling time, including corrections and checks. A draft created in seconds can still cost more work if every line has to be reverse-engineered before anyone trusts it.

A useful pilot has a maintenance plan

Keep the first trial to one service family, one price source and a named approval role. Let drafts run alongside the existing process. Send through the agreed controlled route, with the normal manual method available for exceptions. This gives you room to improve matching without putting unreviewed commercial documents in front of customers.

Record changes to descriptions, price lists and required inputs during the trial. That helps you tell whether a better result came from a model change or simply from cleaner source data. Assign responsibility for keeping the catalogue current after the trial ends. An unattended price source becomes less reliable every time purchasing conditions or the scope of a service changes.

Once a quote has been sent, reminders and customer questions become another workflow. We cover that in following up outstanding quotations. The preparation pilot should answer a narrower question: can your actual enquiries produce a reviewable commercial draft with less total effort? Bring anonymised examples and your present price structure to a discussion about process automation for small businesses. Include the troublesome jobs that still make your estimator pick up the phone.