How to separate text preparation, verifiable data and approval before sending a commercial proposal.
A commercial proposal combines several tasks: understanding the request, selecting the scope, inserting confirmed terms and explaining them clearly. If everything is delegated to one open-ended AI prompt, it becomes difficult to verify where items, numbers and promises came from. It is more useful to separate information sources, text preparation and the employee’s decision in advance.
For a company in Kazakhstan, the first result of such a project may be a verifiable draft based on an agreed template. Here we focus specifically on document preparation. Choosing a process for the first implementation is covered in the article about starting automation.
Define the scope of the draft
Choose a proposal type the team genuinely prepares regularly. Clarify which parts can follow a template and where an employee selects a solution or agrees individual terms. For the first launch, limiting the set of situations makes it possible to test the entire journey using clear examples.
Define the outcome of the automated step: the document receives draft status and goes to the assigned reviewer. Sending it to the client is a separate action with an agreed basis. If the company also wants to automate sending, describe and test that separately. Do not confuse preparing persuasive text with authority to grant a discount, change the scope or promise a deadline that is not supported by verified data.
Map the fields and sources
Break the template into elements: customer task, items, quantity, specifications, price, terms and next step. For each field, record where its value comes from and who confirms it. Some come from the request, some from a current company source, and others are determined only after clarification.
A field specification card may contain:
- Name and meaning: exactly what the proposal recipient will see.
- The source and how to check that it is current.
- Whether it is required for the selected document type.
- Permitted transformation: transfer, selection from a list, rule-based calculation, or text preparation.
- Action when information is missing or contradictory: ask for clarification, flag it for an employee or stop preparation.
- The person responsible for confirming the value.
Keep a way to trace the source values alongside the final figures. If they come from different document versions, resolve the discrepancy before approving the proposal.
Separate numbers from free text
Perform calculations using agreed company rules and verifiable source data. AI can prepare a task description or coherent text based on the already selected scope. However, a request to “make the proposal more attractive” must not authorise adding new properties, bonuses or terms.
A strictly defined response structure helps organise fields but does not replace checking their content. In OpenAI’s Structured Outputs documentation it explicitly states that these responses can still contain errors. Acceptance checks therefore examine the origin of values and the meaning of the text, even if the document looks tidy and all required fields are filled in. Choose a specific service after defining the task and the available connections.
Work through one sample request
Imagine a supplier receiving a request to furnish a meeting room. The enquiry specifies the type of room, but product options and final quantities have not been agreed. This is an illustrative example, not a result of GIFS client work.
The first draft can restate the task, list verified starting information and collect clarification questions. It must not present assumed items as an agreed scope. After the customer replies, an employee selects options from the current source and document preparation continues using that data.
Now change the condition: the customer replaces one option after the draft has been created. It should be clear which items, calculations and descriptions are affected. The reviewer receives an updated version and sees the reason for the change. A previously prepared file should not automatically be considered current simply because it is already in the shared folder.
Agree on the review and version to send
Ask an employee to review the document in decision-making order: was the task understood correctly, does the scope match it, are the figures and terms verified, and are there any unnecessary promises? Check names, units of measurement, and the chosen language separately. A Russian or Kazakh version needs someone who can assess the meaning of that text.
Record the reviewing employee, status and exact version approved for sending in the workflow. If a significant value changes after approval, define which checks must be repeated. Give process participants access to sources and results within the agreed scope. Before using an external service, establish which data will be transferred and what processing capabilities it offers; check the specific terms for the chosen solution.
Evaluate the pilot using comparable examples
Prepare several anonymised requests covering different situations: a complete scope, missing information, a conflicting condition and a change after the draft. Define an acceptable outcome for each in advance. Check not only that a file appears, but also that the employee knows what to do when preparation stops or needs clarification.
Compare the whole process, including finding source material, reviewing and making corrections. Track the time until a proposal is ready to send, the number of substantive edits and the reasons for manual intervention. Do not attribute savings to automation based on one successful example. If measurements are not yet available, record only observed results and leave the conclusion about impact open until there is enough evidence to compare.
Create a proposal evaluation card
The reviewer needs the same process for a manually prepared document and an AI draft. Decide in advance which errors prevent sending and which are editorial. An error in price, quantity or commitment matters more than elegant wording.
- Completeness: all required fields are present; unknown values are clearly marked for clarification.
- Facts: every item, specification, price and deadline matches the stated current source.
- Boundaries: no invented discounts, promises or independent agreement of terms.
- Employee work: substantive corrections and the total time to an approved version are recorded, including repeat review.
- Failures and changes: conflicting data blocks approval, while a new source version triggers rechecking of the affected fields.
For each test request, save the source materials, template version, draft and reviewer’s decision. A case with a missing deadline is handled correctly if the system asks for clarification instead of guessing a value. Compare similar cases before and after the pilot, and include service and review costs when evaluating the result.
This test helps decide whether to proceed to implementing AI in a workflow. It does not prove sales growth: that requires separate observation of sent proposals and other changes in the team's work.
Submit the task using a short template
To discuss the project, prepare one anonymised request, an agreed proposal template and a field map. Add sources, calculation rules, limits on free-form text and the approval process. Separately identify actions the system must not perform and the person who handles disputed cases.
The first launch can be scoped as follows: “We prepare a draft of this document type from these data; show gaps to the reviewer; and allow only an approved version to be sent.” This makes it possible to discuss a specific process rather than a general promise to automate a department’s work. If it needs to be connected to enquiries, a useful resource is the article on transferring enquiries to CRM.
In the section GIFS AI and automation you can discuss this scenario. Send an anonymized template, the current workflow, and examples of questions an employee resolves before sending a proposal.