An AI agent that writes sales proposals
From meeting notes, emails and the salesperson's notes to a PDF ready for review. A full case, with the final document at the bottom.
The case
Valdiserra Logistica S.r.l. is an invented third-party logistics company: 140 people, three warehouses, 900 shipments a day. After a first meeting it asks for a proposal for an AI platform with 20 users, an index over contracts and procedures, an agent that prepares transport quotes, support and two training sessions. As an option, an agent that answers customers about shipment status by reading the TMS.
The material the salesperson has is what every deal produces:
- the meeting notes generated automatically from the recording, with transcription errors and figures quoted out loud (“you told me sixty-five euros an hour”);
- the customer’s email with the scope, the constraints (data in Europe, login with company credentials) and the deadline;
- the salesperson’s internal notes: what is agreed, what is uncertain, what to watch out for.
Three traps were planted on purpose: a wrong hourly rate, a misremembered discount and a “5% fee” mentioned in an email that nobody can place any more.
The agent
It is called Offerta Commerciale and it is an AIDeskPro Configuration. Its text, in plain language, says what to consult, how an Open Gate proposal is structured, the style rules (“the Customer”, no personal names, prices excluding VAT), the rules on numbers and a checklist to pass before delivering.
It has three Resources of type Index:
- Product sheets: the products and services Open Gate sells, with fees, rates and SLAs. The only source allowed for prices.
- Past proposals: proposals already issued, for structure, wording and payment terms.
- Proposal material: the documents of the deal. This index is chosen at launch, because it changes with every customer.
Plus one attached file, the rendering kit: logo, fonts, layout script and PDF verification script.
What happened, step by step
1. The command
You create a new Execution, pick the index with the customer’s material and write the mandate:
Prepare the proposal for Valdiserra Logistica S.r.l. based on the Proposal material. If anything is missing, ask me before writing. At the end, also produce the PDF with the rendering kit.
2. The reading
The agent lists the documents of the three indexes and reads them in full: the customer’s three files, the eight product sheets, the past proposals. Every read shows up in the Console.
3. The three questions
Before writing a line, the agent stops. It found three points where the sources are not enough or contradict each other, and turns them into closed questions, with options and a recommendation.


The third question was about the support fee, which differed between the product sheet’s table, a side note and last year’s proposal: the agent asked which one applies and proposed the one in the sheet. Three clicks and the agent resumes.
4. The writing
The proposal is born as Markdown in the workspace, together with two files the customer will never see:
- fonti.md: for every statement in the proposal, where it comes from. Three labels: SOURCE (written in a document), DEDUCTION (derived by a calculation), MY CHOICE (not in the documents, to be confirmed).
- stime.md: for every estimated figure, the breakdown into days and rates, the assumptions and a confidence level.
5. The PDF
On request, the agent unzips the kit, runs the layout script and then the verification script: pages, table of contents with real page numbers, embedded fonts, unbroken tables, two-column signatures. If a check fails it corrects and repeats. The PDF lands in the workspace next to the proposal.

6. The delivery notes
The agent closes with a message to the salesperson: where the files are, what the proposal contains and, above all, what to verify before sending: the estimates in days, the conditions marked MY CHOICE, the ambiguities in the spoken sources, the placeholders for the customer’s administrative data.
From the delivery notes of this execution, translated:
Estimates in days (judgement, not price list data): the 5-8 days of specifications come from the technical lead, but the estimates for setup/SSO/SharePoint, for the development of the quoting agent and for the TMS option are my own engineering estimates and must be validated.
The PDF still contains the placeholders for the Customer’s data and is not ready to be sent.
One detail we had not planned: the internal notes said “Treviglio, about 30 km from Gorgonzola”, beyond the 25 km threshold of the travel price list. The agent did not take it on trust, computed the road route, found 21 km and applied the right rate, flagging the inconsistency among the points to verify.
The result
21 pages, six chapters, Open Gate’s layout: cover, recipient, table of contents with page numbers, introduction, details of the proposal, estimates, fees, payments, validity with the signature and administrative data tables.
The introduction separates the fixed part (AIDeskPro plan for 20 users, support, training: 27,688 € a year) from the part to be estimated (setup and SharePoint connector, specifications and development of the quoting agent, in a stated range) and from the option for the second phase. Then, since the customer had said out loud “around 35,000 euros”, a paragraph explains why the full estimate is higher and proposes three levers to get back to it.
The three traps: the “65 euros an hour” the customer remembered became the list rates; the “10% discount on ten days” became the real rule of prepaid packages; the “5% fee” was clarified as a line of a different plan, not the one offered. All three are also in the list of things to verify before sending.
Reading and questions were done by Claude Sonnet 5.5; writing and PDF by Gemini 4 Argon: the model can be switched even while the execution is running.
The document below is exactly what came out of the workspace, untouched. It is in Italian, like the whole case. The placeholders for the customer’s data are the ones the agent leaves to be filled in.

