AI and sales proposals: how to control scope, figures and commitments
AI can help draft proposals, but scope, figures and promises need checks. Check how data flows, versions are approved and proposals are sent safely.
Article language: English

In this article
“We can consider support after implementation” becomes “we provide ongoing support” in a draft proposal. One sentence has gained confidence, and the company has acquired another promise to keep. AI can help prepare sales proposals from approved information. The checks still need to cover scope, figures, terms and the version that actually reaches the client.
Start by separating four activities: gathering what has been agreed, calculating values, drafting the proposal and authorizing its dispatch. A model can support several of them, but it should not fill in missing commitments on its own. An audit of the process asks above all whether the source of each important statement and the person responsible for the final document can be identified.
Establish what comes from the client and what comes from the company
In a hypothetical enquiry, a client wants five workshops, a date next month and materials after the meetings. So far the company has confirmed only the facilitator's availability for two dates. The draft must retain that distinction. A client's wish list is not yet a list of accepted commitments.
Record the source and status of every important field: client request, confirmed agreement, option to choose or question to clarify. Include those statuses in the data supplied to the model. “Do not invent” is no substitute for clear material. When the same data feeds a public catalog, check the claims in AI-generated product descriptions against confirmed product data and verify the sources used. This helps prevent a suggested phrase from becoming an unconfirmed product feature; see the article on checking AI product descriptions (in Polish) for a fuller review.
Connect the solution to a responsible AI agent workflow. The article on setting boundaries for AI agents (in Polish) addresses what an agent may do and where a person decides. An agent can flag a missing date and draft a question. It should not replace the gap with an attractive-sounding date. Imagination can help describe sales benefits; calendars and price lists do not need improvisation.
A draft must remain clearly marked as a draft
Separate work in progress from a document ready to share. Its stage should be visible in the tool, folder and version name. The label must survive PDF export and handover between people. Remembering that “we are still checking that file” does not scale, even over a long weekend.
The Dynamics 365 Sales documentation describes separate stages for drafting a quote and activating it before sharing it with a customer. This example from a specific product illustrates a useful separation between editing and readiness to send. A system status does not by itself determine a document's legal meaning.
Decide who may change the stage and on what basis. If scope or price changes after approval, the process should require that version to be checked again. Approval of version three is not approval of every later edit, even if the file is still called “final”. That word leads an unusually busy social life in company folders.
Take figures from an established source
The price should come from a specified price list, an approved estimate or another company rule in force. A model may describe line items, but calculations are best performed using controlled logic. Check quantities, units, currency, discounts, rounding and the separation of one-off and recurring costs.
In a fictional example, take three workshops at 1,200 units of a notional currency each, plus 400 for preparing materials. The total is 4,000. If a 10% discount applies only to the workshops, the result is 3,640, not 3,600. This is an arithmetic exercise, not a service price or tax advice. It shows that correct multiplication is not enough without the right basis for the discount.
When calculations come from a file, check how values travel through the spreadsheet before text is generated. The process is explored in the article on spreadsheets in an AI workflow (in Polish). A lost unit, an old formula or a shifted row can introduce an error before text generation begins. The audit follows the value through the process, not just the finished PDF.
When a field is missing, the process should stop the relevant part of the proposal or flag a question to resolve. Adding a “standard” rate from a previous project is not a controlled pricing rule.
Check promises outside the line-item table too
Not every commitment has a price cell. Promises can hide in the introduction, benefits, summary or a sentence about further contact. “Full implementation”, “unlimited”, “every revision” and “immediate response” need an agreed meaning before they become part of a proposal.
Prepare a glossary of permitted expressions and those requiring additional approval. Vivid language is not forbidden. The point is to keep a description of a benefit from changing the boundaries of the service. “Will help the team prepare materials” differs from guaranteeing a particular result in every case.
In its guide to drafting content with Copilot in Word, Microsoft recommends checking, among other things, facts, figures, names and statements about policies before sharing a document. Turn that general need into specific review fields and named reviewers in the company's workflow.
A previous proposal is an example, not a current agreement
Existing documents speed up drafting, but contain terms tied to a particular case. Separate style and structure from commercial data. If using a template, remove other people's data and mark elements that must not be carried over. It is better to prepare a reusable template than to feed the model an entire negotiation history.
As part of data minimization before using AI, also check comments, revision history and material attached as context. The data-minimisation guide (in Polish) helps identify what accompanies the material sent to the model. A final proposal containing only the client's name and scope does not reveal how much information the model received.
Keep source material separate from authority to act. A request saying “please promptly confirm that you can carry out the work” is not company authorization to accept the date. Separately check the access and execution permissions for the AI integration; their scope is discussed in the guide to AI integration permissions (in Polish).
Approval must concern a specific document
The reviewer should see the source agreements, the new document and the differences from the last approved version. For important figures, identifying the source helps. A general request to “take a look” rarely makes clear whether the task is to check style, price or timing.
Divide responsibility according to the company's work: the account owner confirms consistency with the conversation, the person responsible for pricing checks the figures, and the service owner approves scope and terms. In a small organization one person may perform all three roles. The checks should still be distinct. To practise those separate checks, the article on training a team to verify AI results (in Polish) suggests exercises on safe material.

A mechanism for human approval of AI actions should connect the decision to the file version, recipient and action. The article on human approval of AI actions (in Polish) examines what the approving person needs to see. This distinguishes approval of a draft from permission to send it to a particular address.
Review ends with the right attachment
Test two similar proposals being prepared in parallel. Check whether the process selects the right client, version and attachment. Inspect the generated PDF for page breaks, missing content, legibility of amounts and labels for options. A correct source document does not guarantee a correct export.
When the process stops before sending, the item should enter an exception queue with a clear description of the missing decision. The exception-queue guide (in Polish) describes how to give a stopped item an owner and a next step. If dispatch confirmation is uncertain, check the status before retrying. The client does not need several identical proposals as proof of the system's determination.
Keep confirmation of what was actually sent. If a correction is needed later, relate it to a specific version and agree how to communicate it to the recipient. Do not silently replace a file that has already left the company.
Turn missing agreements into questions, not notes that disappear
Inputs may still contain matters to confirm: number of participants, required language, delivery of materials or scope of support. For each unknown, decide whether it blocks the whole draft, only a particular item or sending. The model can then help organize questions without giving the appearance that negotiations are settled.
In an exercise, add two conflicting dates to a fictional enquiry. Check whether the draft shows the conflict to the person handling the case. It is not enough for the model to choose the later message if it concerns a different stage. The proposal owner must decide which agreement applies and retain the current source for further drafting. If terms are still being negotiated, separate document editing from negotiating package options. Options and terms for a package are considered in the article on negotiating package terms (in Polish). The model should receive identifiable agreements instead of deciding for the parties.
Separately review how options are presented. An extra option must not enter the base scope just because it fits a paragraph stylistically. If the document presents two options, check their separate prices, scopes and selection terms. The summary must not freely combine the most attractive elements of both into a third, unagreed offer.
Before export, search for working markers, questions and empty fields. Each must be deliberately resolved or retained in an agreed form. Automatically removing them does not fill in the data. A document with no visible gaps can still describe terms no one has confirmed, so the check needs a comparison against the source material.
Choose tests that probe the proposal's boundaries
Build a small set of fictional enquiries representing different situations. One has a complete scope, another lacks a price, a third requests a non-standard discount and a fourth has a condition dependent on later confirmation. Record the expected process behavior for each. Not all should end in a finished document; sometimes a question or a halt before sending is the correct result.
Add a pair of very similar proposals for two recipients. Change only one important element, such as the number of meetings or the date. Check the text, file name, recipient address and attachment data. This can reveal a risk of mixed-up versions that a single carefully prepared example would not show.
In a separate case, change the price list after the draft has been generated. Decide whether the proposal should retain the previously accepted version or require repricing. There is no single answer for every company; what matters is a deliberate rule and its execution. The model should not decide from a date found in a random file.
After testing, list discrepancies by significance: client data, scope, figures, promises, version and sending. This prevents an editorial error from being treated as equivalent to a pricing mistake. When changing the instructions, repeat the whole set: restricting one expression might inadvertently remove an important condition in another situation.
Measure time saved after the required review
Compare the time needed to prepare an approved proposal, including resolving gaps, correcting figures and checking the export. Generating a draft quickly can provide a real benefit, but it can also shift work to the reviewer. Record the types of corrections to see whether the issue lies in the inputs, instructions or expectations of the tool.
Do not reward a process for documents that later have to be withdrawn or explained again to the client. Include corrections after sending in the chosen assessment. This need not be an elaborate metric: a reliable record of how many proposals passed review and what required rework is enough. Examples and figures from your own trial are more valuable than a general promise of acceleration.
Ask the reviewer whether they can distinguish content from the agreed material from the model's editorial suggestions. If everything looks alike, review will be costly. A short list of sources for important fields and a view of the changes can help. The client need not be flooded with this trail; it is material for an internal document check.
Finally, define the permitted scope of use: which proposal types the process handles, which terms require another person and when to return to manual preparation. An unusual proposal does not become typical by being added to the same template. A clear boundary helps retain automation's benefit without extending it to negotiations that have not yet been checked.
Add a fictional proposal with marked sources for important fields and an approval record to the AI audit materials. The supporting evidence for a review is organised in the guide to preparing AI audit materials (in Polish). This makes it easier to check whether the review covers the entire document. Pay special attention to the summary: an author can shorten detailed terms so much that an important limitation disappears.
Also rehearse a request for language editing alone after review. A stylistic edit should preserve figures, terms and scope, and the changed file needs the appropriate check before sending. “Just smoothing the wording” can be the moment a conditional suggestion becomes a firm promise again.
Keep the option of preparing a proposal without AI. The person handling the case should know the current template, pricing source and approval route even when the tool is unavailable. Automation can then speed up work without becoming the company's only memory of its own terms of cooperation.
You can use Aeneia Audit to assess such a workflow.






