AI quote drafting in the trades: from dictation to quote

Dictation costs 1.34 cents, an office hour costs 39.60 euros

Inference is cheap; retyping afterwards is not. Turning a four-minute voice note from a building site into a quote draft costs 0.012 US dollars for transcription, since OpenAI lists gpt-4o-mini-transcribe at 0.003 US dollars per minute. Add roughly 3,600 input and 1,200 output tokens for the draft: 0.0033 US dollars on gpt-5-mini, priced at 0.25 against 2.00 US dollars per million tokens.

That is 0.0153 US dollars in total. At the European Central Bank reference rate of 1.1403 US dollars per euro, dated 25 September 2026, you pay 1.34 cents per quote. Going through Gemini 2.5 Flash, which handles audio directly at a documented 32 tokens per second of sound, works out at 1.02 cents.

The other side of the ledger sits in payroll. One hour worked in German construction cost an average of 39.60 euros in 2025, non-wage labour costs included. Swap 25 minutes of typing per quote for 9 minutes of checking: 16 minutes saved, 10.56 euros saved. At 13 quotes a month, that is 137.28 euros.

Four steps, with the checkpoint built into step three

The workflow has four steps and one fixed checkpoint. plancraft states as vendor that notes, voice recordings, photos, measurements or price lists turn into a structured quote draft. You can build the same thing without trade software, as long as you paste your own price list into the prompt.

The checkpoint lives in a column called Source, so make it mandatory. OpenAI's Structured Outputs guarantees a response matching the JSON schema you supply, so required fields cannot go missing. One glance then tells you which line item the model invented.

• Step 1, dictate instead of narrating: item by item, each with room, task, quantity, unit of measure, anything unusual. Four minutes covers eight items.
• Step 2, hand over master data: task text, unit of measure, unit price belong in the prompt as plain text. Leave them out, so the model substitutes market prices that are not yours.
• Step 3, ask for a table: one row per item with quantity, unit of measure, unit price, total, and a Source column pointing at a price-list row. Empty Source means invented item.
• Step 4, check three lines: unit of measure, material markup, incidental work. Only then does the draft reach your template.

Three lines decide the price, never the prose

The wording lands; three numbers do not. Phrasing a task description is easy for a language model, matching quantity, unit, markup is not. The quote below is arithmetic, no customer case: painting work, 8,140 euros net, eight line items.

Those three lines add up to 563.80 euros, so 6.9 percent of the quote. Nine minutes of checking cost 5.94 euros, a ratio of 95 to 1.

Unit of measure is the expensive one, because it breaks both ways. Too high, so the customer compares you out of contention. Too low, so section 145 of the German Civil Code binds you to your own offer the moment the customer accepts.

• Unit of measure: the note says 24 running metres of skirting board, the draft writes 24 pieces. At 4.80 euros per metre against 11.50 euros per board, the quote shows 276.00 instead of 115.20 euros, overstated by 160.80 euros.
• Material markup: the model copies your purchase price straight from the list. On 1,240.00 euros of material at a 15 percent in-house markup, 186.00 euros go missing.
• Incidental work: masking, covering, disposal never reach a voice note, because they go without saying. At 3.5 hours times a 62.00 euro charge-out rate, nobody billed 217.00 euros.

A prompt that refuses to invent prices

Four rules decide whether a draft is usable. Without them the model fills every gap, because complete answers are what it was trained to produce. Copy the text below, pasting your transcript and price list after the output line.

Rule 4 is the part packaged tools rarely offer. It hands you the points to settle with the customer before a draft becomes binding. Skip it, so those points return as a variation request under section 650b of the German Civil Code, by which time the money is already spent.

• Role: you are an estimator at a painting firm, turning a dictation into a quote draft, adding nothing of your own.
• Rule 1: use unit prices from my price list only. Find no matching row, write MISSING into Source, leave the unit price blank.
• Rule 2: copy the unit of measure verbatim. Never convert metres into pieces, nor square metres into metres.
• Rule 3: calculate no markups, no discounts, no value added tax.
• Rule 4: at the end, under a heading OPEN, list every task the dictation hints at without a measurement: covering, disposal, scaffolding.
• Output: a table with item, task text, quantity, unit of measure, unit price, total, source. Then the OPEN section. No cover letter, no sign-off.

Trade software or your own workflow: where things stand

The integrated route costs more, and it is unfinished in places. plancraft lists AI quote drafting as a shipping feature. HERO Software labels HERO Command, its planned quote builder, coming soon on its own product page; what ships there today is the phone assistant HERO Voice.

The model costs 1.34 cents, the subscription 343 times as much. You are paying for everything downstream, never for inference: customer file, numbering, invoice, post-costing. Already run a trade suite, so the AI feature is one more row on the bill.

Both vendors describe their own features; no independent measurement of draft quality exists for either.

• plancraft Business: 59.90 euros a month on annual prepayment, before value added tax, or 47.92 euros on two-year prepayment, one office seat.
• HERO OS: 109 euros a month under a time-limited launch offer, 119 euros normally, AI features included per its pricing page, the quote module not yet shipped.
• Your own workflow over the APIs: 1.34 cents per quote, no subscription. In exchange you maintain the price list yourself, and moving the draft into your template stays manual.

What I would do in your position

Six quotes a month pays for the subscription. The threshold falls out of the arithmetic above: 59.90 euros divided by 10.56 euros saved per quote gives 5.7 quotes. Below that, take the prompt plus your phone's dictation app; above it, take the trade suite.

Yes to AI quote drafting when your price list is current and your quotes repeat the same items: painting, plumbing maintenance, electrical retrofits. No when you work without a price list, or mostly price one-off fabrication. There the model substitutes somebody else's prices, so checking outlasts typing.

No as well when nobody in the firm wants to check three lines. An unchecked draft buys no time; it produces a cost estimate whose material overrun you must report at once under section 649 of the German Civil Code.

Frequently asked questions

• How long may the voice note run? Gemini accepts a documented 9.5 hours of audio per request. In practice stay under five minutes, or you will never find the passage again while checking.
• May I dictate customer names and addresses? Not on a first run. Redact name, address, phone number using our free tool for redacting prompts; where redaction fails is spelled out in four patterns and one limit.
• Which fields may AI never fill alone? Price, deadline, liability. The six-check sign-off appears in creating quotes with AI.
• Does AI do the post-costing too? No. It delivers a draft, never an actual. Comparing booked hours against material receipts needs reporting in your system.
• What about incoming supplier quotes? Reverse direction, different workflow, described in comparing supplier quotes in Excel.

Your next step

Take the last quote you wrote. Re-dictate it in four minutes, push it through a model with the prompt from section four and your price list, then compare row by row against the original. One example tells you whether your price list is current enough.

Should the Source column come back blank more than twice, the model is fine, your list is stale. Anyone wanting this workflow wired into their own trade software finds the matching services under automating internal processes.