Quotations with AI: The Agent Does the Groundwork, You Set the Price
A request for quotation arrives by email, with an Excel list and drawings attached. Before anyone can price it, someone has to establish which line items the company has made before, which drawing revision applies and which certificates the customer demands. This groundwork takes time, and often only the person who knows the earlier orders can do it. There are two ways to shorten it: fixed rules or an AI agent.
Which one fits is decided step by step.
We know both approaches from our work for MAFU-SHERPA, though at different points in its sales process. MAFU-SHERPA develops AI-controlled CNC automation for small and medium batch sizes. Its website, video series and search campaigns brought in hundreds of sales inquiries, and every quotation rests on product logic with many technical dependencies. For the quotations, we built a configurator directly inside Business Central. Sales enters the customer requirements in a structured form, and bills of materials, routings and sales quotations follow from fixed rules; hundreds of quotations have now been produced automatically through it. MAFU-SHERPA uses an AI agent elsewhere, in finding new customers. For that, we run an AI sales assistant there. It reads the websites of potential customers, scores their technical fit on a scale from 0 to 100 and drafts an outreach message, which a salesperson reviews and sends from Outlook. We have not built an agent for incoming requests there. The section further down describes how we would set one up.
Which part is a rule
For MAFU-SHERPA's products, what counts as correct can be written down in advance. Structured requirements lead to specific bills of materials, routings and quotation lines, and a rule that has been tested once calculates every quotation the same way. The central question in the project was which logic belonged in the Business Central standard and which needed a targeted extension.
The websites of potential customers are a different matter. Whether a manufacturer is a candidate for the SherpaLoader is recorded in no database field; it shows on machine-park pages, in job ads and in investment news, and every website is built differently. That is where the assistant reads and judges in context. Where the same project only had to find the right LinkedIn company pages for a list of company names, a small application with fixed rules was enough, scoring search results in a traceable way.
For quotations, this suggests a simple split. Whatever can be written down as a rule, such as set-up costs per machine, minimum quantities or the surcharge for a material certificate, belongs in a spreadsheet, a configurator or the ERP system. The agent gets the part nobody can describe in advance, meaning the freely worded email, the note in the margin of a drawing and the parts list that every customer structures differently. If you sell standard products in variants, you may therefore never need an agent. If you manufacture to customer drawings, every request arrives in a different form, and that is where an agent earns its place by reading and comparing.
Where an agent takes over the groundwork
Even the best configurator starts only once someone has entered the requirements in a structured form. At MAFU-SHERPA, sales does that. In a company that manufactures to customer drawings, this is exactly the laborious step, and an agent can do the groundwork there. For that, it may read the inbox for requests, the item and order data in the ERP system and the archive of earlier quotations. It may write only drafts. With a request for quotation, it then works like this.
- It reads the request, the drawings and the purchasing terms, and records drawing revision, material, quantity and required certificates for each line item.
- It looks up every line item in your earlier orders. Where a part was made before at the same drawing revision, it adds the earlier price, with a reference to the old order.
- It flags what is missing or does not fit, such as a newer drawing revision than the one in the list, or a demand for an EN 10204 type 3.1 inspection certificate, which must be ordered with the material.
- It writes two drafts, a question to the customer and the quotation with every line item. Where there is no reference value, the price field stays empty and flagged.
The gain lies in the order of events. The question needs no costing to check, so it can go out early, long before anyone has time for the costing. When the responsible person sits down with the quotation, the customer's answers may already be in, and the real work starts with the new and changed line items.
Public tenders with procurement documents and formal eligibility evidence follow their own rules and are beyond the scope of this article. It deals with an industrial customer's request for quotation, even when the customer's purchasing department calls it a tender.
Price, date and commitment stay with a person
The price depends on things no file records. How busy will the machines be next year, how much does this customer matter to you, and will the material price hold over the term? An agent can pull a reference value from earlier orders. It does not set the price, because your company answers to the customer for that price. Months of sound drafts do not change that.
OpenAI describes a similar boundary. In its practical guide to building agents from April 2025, the company recommends rating an agent's tools by whether they only read or also write, whether their effect can be reversed and what financial impact they have. For high-risk tools, the guide describes additional checks or escalation to a person. A sent quotation changes something, is hard to take back and concerns money. In our delegation map, this step therefore stays permanently at the level "AI drafts, a human releases".
MAFU-SHERPA's sales assistant already works at this level. A draft goes out only after a salesperson has reviewed it and sent it from Outlook.
The draft quotation has a second boundary. The agent invents no value that someone will later have to stand behind, and we prefer an empty, flagged price field to a plausible number that slips through on a quick read. We do not yet know whether an agent can derive a usable machining time for a new part from similar earlier parts. We have not tested it on real orders, and until we have, such a value would at most sit beside the line item as a flagged estimate.
Five questions before the first attempt
Whether quotation preparation in your company suits an agent depends less on the tool than on your documents. You can answer these five questions in one meeting without us.
- Where are your earlier quotations and orders, and would a new employee find last year's similar part in them?
- Which price source applies when the price list, the costing spreadsheet and the last quotation disagree?
- Can you write down how you recognise a complete request?
- Which parts of your costing are fixed rules that belong in a spreadsheet, a configurator or your ERP system?
- Who releases questions and quotations, and who stands in when that person is at a customer?
The answer to the third question becomes the agent's first working instruction, and it grows with every request. If the agent misses a required heat treatment, the note belongs in that list so that the next run already knows it. No tool helps with the second question, however. Where three sources give three prices, a person first decides which one applies; only then is it worth letting an agent work with them.
The first attempt with the tool you already have
The first attempt needs no new software; as with using AI in your business generally, the existing tool comes first. In the AI tool you already use, such as a project in Claude or ChatGPT, store your list for complete requests and the current price list. Then give it three real requests from recent weeks, including one that later went wrong, and compare the gaps it flags with what you noticed at the time. If the tool finds the gaps but not the earlier orders, the next question is a read-only connection to your ERP system; in contract manufacturing, that connection is often the larger part of the work. If your costing follows fixed rules, a configurator comes first. And if such requests arrive only a few times a year, that project in the existing tool is probably already the smallest useful solution.
Before you start, check whether customer drawings may go into your AI tool at all. That depends on your plan and its settings, and often on a non-disclosure agreement with the customer.
At MAFU-SHERPA, the configurator came with the ERP system, once the company was ready for it. Whether a configurator, an agent or your existing tool comes first in your company depends on one question: which line items a rule can calculate, which an agent has to read and which only a person may decide.
The free AI assessment answers which of your workflows suits a first attempt best, whether quotation preparation or another. In 60 minutes by video we go through one or two recent cases, and within two working days you receive a written recommendation for the first step, which may be that your existing tool is enough.
If you want to try this with one of your own requests: Book a free AI assessment.