AI for salespeople: an assistant for preparing a discovery call and proposal draft

A discovery call often determines whether a salesperson gets a precise brief or walks away from the conversation with only a mix of general requirements. The second critical point comes immediately after: turning notes, company context, and signals from the call into a proposal that is specific, realistic, and understandable for both the buyer and the end requester. This is exactly where AI makes the most sense — not as a replacement for the salesperson, but as a working layer between research, question preparation, note-taking, summarization, and the first draft of the proposal.

In practice, this is not about “magically closing deals.” What works best is whatever shortens preparation, improves the structure of the call, and reduces errors when transferring information into the proposal. Typically: finding public information about the company, creating a discovery call outline by segment, transcribing and summarizing the call, extracting risks, proposing next steps, and drafting the first version of the sales proposal. If you want broader context on using AI in everyday business work, this topic is also followed up by content in the AI tools on AIVýběr section.

It makes sense to be strict about the output. AI should provide supporting materials, not improvise commitments to the client. That is why in this article I will distinguish what to automate, what to only assist with, and where it is safer to leave the process entirely to a human.

What AI should actually do in discovery call preparation

The best use of AI before a call is very specific: prepare a brief company briefing, suggest hypotheses about problems, compile questions, and point out gaps in the information. In practice, this means combining publicly available data from the client’s website, LinkedIn, press releases, and possibly CRM notes into one working outline. The output should not be longer than one page or 8 to 12 bullet points for the call.

OpenAI

What to do: Create a fixed pre-call brief template with sections such as “company and context,” “likely goals,” “risks,” “questions to uncover budget, process, and priorities,” and “what not to promise without verification.” AI should only fill in the supporting materials. Tools such as ChatGPT, Claude, or Google Gemini work well if you provide them with specific sources and a clear output format.

Who it is for: This is most beneficial for B2B salespeople, account executive roles, and founders of smaller agencies or SaaS companies who run several different types of meetings each week. For them, AI mainly saves context switching between segments and reduces the risk that generic questions unrelated to the client’s industry will come up on the call.

When not to use it: Do not use an AI briefing as the only source of truth for large enterprise accounts, where internal relationship history, previous tenders, or sensitive business terms are critical. If 70% of the decisive information is not in public sources but in the CRM and in the heads of people on the team, AI without quality internal inputs will produce only a convincing-sounding surface.

The practical decision rule is simple: if call preparation usually takes more than 20 minutes and you still use the same question framework, automating preparation is worth it. If every deal is so unique that you still manually assemble hypotheses from internal data and stakeholder maps, let AI only help summarize materials, not create the entire call strategy.

What a usable pre-call briefing should look like

A usable briefing is not an essay. It is enough to include: the company’s industry and business model, recent relevant changes, an estimate of the roles of the call participants, three likely pain points, three qualification questions, and two warnings. For example, warnings such as “the website has outdated pricing” or “the company is hiring internally, so it may be considering an in-house solution instead of a vendor.” That is more valuable to a salesperson than a long market description.

A good practice is to require AI to cite sources or at least list the URLs it used. If the tool cannot do this reliably, do not accept figures or claims about revenue, customers, or technologies without manual verification. In a discovery call, one confidently wrong sentence is worse than five unfilled bullet points.

AI during the discovery call: transcription, structure, and capturing signals

During the call, AI makes the most sense mainly as a note-taker and structurer. The goal is not to “put a bot into the meeting,” but to get a quality transcript, summary, list of open points, and a follow-up suggestion. This is where specialized services such as Fireflies.ai, Fathom, Avoma, or Otter.ai clearly have their place. In reality, they work by joining the call, creating a transcript, and after the meeting generating a summary, action items, and sometimes even topics by chapter.

What to do: After every discovery call, have five fields automatically listed: the client’s goals, the current problem, the decision-making process, the budget or its signals, and the next step with a deadline. Add a separate field for “what remained unverified.” This last section is often more important than the summary itself, because it prevents an assumption from the call from turning into a fixed premise in the proposal.

Who it is for: This works best for teams that run a larger volume of repeatable B2B meetings: software, IT services, marketing agencies, recruitment services, or consulting. If salespeople have 10 or more calls per week, the time savings on note-taking and looking up details are immediate.

When not to use it: Do not use recording without clear consent and without an internal policy for where the recording is stored and who has access to it. It is also unsuitable for highly sensitive topics, for example when sharing non-public financial data, security incidents, or contractual disputes, unless you have an approved tool and legal framework.

The decision rule: if the client refuses recording, do not try to “smuggle” AI in another way. Instead, use live note-taking in a prepared outline and after the call let AI generate a structured summary from your manual notes. The result will be weaker, but the process will remain safe and transparent.

Indicative prices vary by features and number of users. For similar meeting assistants, the basic paid plan is usually roughly in the range of 15 to 40 USD per user per month; more for advanced team plans. This is only an indicative figure, because pricing changes and differs depending on annual commitment, number of recorded minutes, and CRM integrations.

Which signals to focus on in the summary

Do not underestimate signals the client does not say directly. For example, statements like “we need to get this done by the quarter,” “it has to go through finance,” “we already tried another vendor,” or “IT is still involved” are more important for the proposal than a long description of features. AI should pull such statements into a separate “risks and constraints” section, not leave them buried in the transcript.

It is also useful to mark passages where the client uses their own terminology. Then carry that into the proposal. Not because of style, but because of accuracy. If the customer talks about “branches,” “dealers,” or “regional managers,” do not rewrite that in the draft as generic “users” and “operating units.”

How to turn the call into a first proposal draft without losing accuracy

The most common mistake is not poor wording, but poor proposal structure. Salespeople often hand the entire transcript to AI and expect a finished document. This creates a mix of assumptions, client wishes, and technical details without priorities. The correct approach is two-step: first extract verified facts and open points from the call, and only then generate a proposal draft according to a fixed outline.

OpenAI

What to do: Use a proposal template with mandatory sections: starting situation, project goals, proposed scope, what is not included, assumptions on the client side, timeline, price or pricing logic, risks, and next step. AI may fill in only what is documented from the call or from internal materials. Everything else must be marked as a question to be completed.

Who it is for: This approach is suitable for salespeople who sell a service or solution with multiple variables — for example software implementation, marketing services, custom development, training, or an audit. For simple price-list sales, it would be unnecessarily cumbersome.

When not to use it: Do not use AI for final proposal creation where one inaccurate phrase creates a legal or pricing commitment. Typically public tenders, complex enterprise contracts, security audits, or regulated services. There, AI can help with the outline and summary, but the final text must go through manual expert review.

As a generator of the first draft, general models such as ChatGPT or Claude work well, or the Microsoft 365 corporate environment with Microsoft Copilot, if you already have documents and notes inside the Microsoft ecosystem. What matters, however, is forcing a “ask first” mode. That is, the model should first list ambiguities and only after they are clarified assemble the proposal.

The practical rule: if more than three unverified points remain in the proposal draft regarding scope, deadline, or client responsibilities, do not send the PDF. Send a text summary with questions first. This prevents the client from starting negotiations over a document built on an incorrect brief.

What should be explicitly separated in the proposal

AI should strictly separate four layers: what the client said, what the salesperson proposes, what is an assumption, and what is out of scope. Mixing these layers is exactly what creates later disputes. If the client mentioned “we need to connect the CRM,” that is not the same as “the delivery includes integration with a specific CRM system including data migration.”

A good proposal therefore does not need a long marketing introduction. It needs accuracy. If AI generates flowery wording, shorten the prompt to instructions such as: “Write as a working draft, not a sales brochure. Every statement must be tied to a source or marked as a proposal.”

Integration with CRM and documents: where the real time savings come from

A chatbot alone is not enough. Real productivity only comes when outputs from the call flow into CRM, tasks, and documents. Otherwise, you only move time from writing to copying. In sales practice, it is therefore worth checking whether the tool can connect to systems such as Salesforce, HubSpot, or Microsoft Dynamics, or at least export to email, notes, and documents.

What to do: Set up a minimum data flow after the call: summary into CRM, next-step task into the task manager, and working proposal into the document template. For meeting assistants, verify whether they support CRM sync or automatic sending of summaries. For the document part, a combination with Google Workspace or Microsoft 365 makes sense, where comments can stay within the team without manual rewriting.

Who it is for: For teams with two or more salespeople, where presales, delivery, or management are involved in the opportunity. As soon as more than one person works on a deal, the quality of information handoff becomes just as important as the call itself.

When not to use it: If you do not have unified field names in CRM, a definition of sales stages, and rules for what gets recorded after the call, automation will only increase data chaos. First fix the process. Only then connect the AI layer.

The decision rule is practical: if after a meeting a salesperson performs the same three manual tasks — summary into CRM, follow-up email, and proposal skeleton — you have a suitable candidate process for automation. If every deal requires a completely different internal approval flow, start only with summaries and leave the rest manual.

For choosing specific services, the overviews on AIVýběr are also useful, especially content focused on AI assistants, where you can compare what types of work tasks different tools cover.

Practical deployment scenarios in a small team and for an experienced salesperson

The usefulness of AI is visible in a specific workflow, not in a list of features. Below are two scenarios that are realistic and deployable without a half-year implementation.

Scenario 1: a small B2B agency with 1 to 3 salespeople

Claude

The agency has several inbound inquiries per week and some leads from referrals. Before the call, the salesperson puts the client’s website, the company’s LinkedIn, and an internal note into a model such as ChatGPT or Claude and has it prepare a one-page pre-call brief. During the call, Fathom or Fireflies.ai runs for transcription and summary. After the call, AI creates a follow-up email, a list of open questions, and a proposal skeleton according to the agency template.

What to do: Introduce a unified proposal with the sections “goals, scope, exclusions, timeline, price, assumptions” and do not let AI change the order or names of the chapters. This shortens review time and makes it easier to compare projects with each other.

Who it is for: For agencies that sell repeatable services, but each project differs in scope. Typically marketing, web development, UX audit, analytics, or B2B content.

When not to use it: Not for projects where the brief is still being formed through a series of workshops and the discovery call is only introductory. There, an automatically generated proposal would create a false impression of precision.

Scenario 2: an experienced account executive in SaaS

The salesperson runs the discovery call, the follow-up demo, and internal coordination with the solution engineer. Here AI does not write the final proposal, but helps with consistency. From the call transcript, it extracts pain points, decision criteria, objections, and the stakeholder map. A brief MEDDICC or other qualification framework is entered into the CRM, according to which the salesperson can tell whether it makes sense to continue.

What to do: Have AI list two extra sections after every call: “why the deal may fall through” and “what must be confirmed before the demo/proposal.” This negative lens is often more valuable for experienced salespeople than a positive summary.

Who it is for: For SaaS salespeople with a longer cycle, where multiple stakeholders are involved and decision-making stretches across several meetings.

When not to use it: Not in situations where the client is testing multiple vendors and sharing sensitive internal materials, if your company does not have an approved enterprise environment and data terms. Here it is safer to do minimal work with the recording and maximum work with a manually verified summary.

Limits, risks, and rules that should be mandatory

The biggest limitation is not the quality of English or style. It is working with incomplete inputs. AI reliably generates text, but unreliably estimates business reality if it lacks context or internal constraints. That is why the sales team must have rules for what AI may and may not do. Without them, speed is easily traded for error rate.

What to do: Introduce four mandatory checks before sending a proposal: scope verification, price verification, verification of assumptions on the client side, and verification of sensitive claims. Each point must be confirmed by a human. AI may suggest, but it must not be the final authority.

Who it is for: For every team that works with multiple salespeople or external contractors. The more people touch the proposal, the higher the risk that an inaccuracy from one note will carry over into the final document.

When not to use it: Do not use an open public AI interface to enter non-public data if your company does not have approved processing terms and does not know how the data is handled. This applies especially to personal data, non-public financial information, and the client’s technical documentation.

There are three practical rules. First: record the call only with consent. Second: an AI-generated proposal must have visible unverified points. Third: no automatic promises of deadlines, integrations, or results without human confirmation. If the tool cannot enforce these guardrails, use it only for working drafts.

Another limitation is language. Czech is usable today in the main models, but errors still appear in industry terminology and legal wording. That is why it pays to keep proposals in a simple, working style and not let the model “polish” the wording. The more decorative the text, the harder it is to detect factual inaccuracies.

FAQ

Is one universal chatbot enough for a salesperson?

For preparing a brief and a first text draft, often yes. For call transcription and structured notes, usually no. In practice, a combination of one generative tool for text work and one meeting assistant for recording and summarization works well.

Does it make sense to generate a proposal directly from the call transcript?

Only if you insert a control step in between that separates verified facts from assumptions. Without this intermediate step, AI can easily mistake the client’s opinion, exaggeration, or preliminary idea for a binding requirement.

How do you know AI saved time and did not just add another layer of work?

Measure three things: call preparation time, post-call note-taking time, and the number of proposal revisions caused by incorrectly captured requirements. If after a month at least one of these metrics does not decrease, the workflow is poorly designed or unnecessarily complex.

Is it safe to record discovery calls?

Only under clear conditions: the client knows about it, agrees, the company has an approved tool, and it is defined who accesses the recording and how long it is retained. Without these conditions, it is better to work with manual notes and subsequent AI-assisted summarization.

How much does it roughly cost?

For common text-based AI tools, paid individual plans are often roughly around 20 to 30 USD per month. For meeting assistants, roughly 15 to 40 USD per user per month. These are indicative figures; the actual price depends on usage limits, team size, and business features.

What is the biggest mistake when deploying AI in sales?

Deploying a tool without a fixed output template. When every salesperson uses a different prompt, a different structure, and a different verification method, chaos quickly arises. AI then does not speed up the process, it only multiplies inconsistent documents.

Conclusion

AI is most valuable to a salesperson at two precise moments: just before the discovery call and immediately after it. Before the call, it helps shorten preparation and sharpen questions. After the call, it turns raw notes into a structure that can be further used in CRM and in the proposal draft. The biggest benefit does not come from “writing nicer texts,” but from better capturing the client’s reality and reducing the number of errors between the meeting and the sent proposal.

If you are just starting with deployment, do not automate everything at once. First introduce a unified pre-call brief, then call summaries, and only then the first proposal draft. Each step must have a clear format, mandatory verification, and a boundary where a human makes the decision. That is exactly what distinguishes a useful AI assistant from another layer of text that looks smart but does not help commercially.

Recommended AI stack for implementation

Choose tools according to your budget and level of automation. Below is a direct overview of services for implementing the project.

Service	Service description	Offer
NordVPN	VPN service for privacy protection and secure connections.	Open offer
Semrush	SEO and marketing platform for analysis and traffic growth.	Open offer
Make	Advanced visual automation for workflows and integrations.	Open offer
Hostinger	Web hosting and domains for fast website launch.	Open offer
Fiverr	Marketplace for freelancers and external specialists.	Open offer
Adobe	Creative tools for graphics, video, and digital content.	Open offer
Canva	Online design tool for graphics, presentations, and social media.	Open offer
Jasper	AI tool for marketing copy and content campaigns.	Open offer