AI for salespeople: prepare for a meeting from CRM, website, and emails in 15 minutes

Discovery calls often go wrong because of one single thing: the salesperson goes into the meeting with a few notes in the CRM, a quickly skimmed client website, and an incomplete idea of what has already been said in emails. The result is predictable. The first minutes are spent looking up context, the questions are too general, and the client feels they have to explain the basics all over again. AI is not a replacement for the salesperson here, but a very practical layer for quickly gathering materials, summarizing them, and turning them into a concrete call plan.

Claude

In this article, I describe a workflow that combines three sources: CRM, the company’s publicly available website, and email communication. The goal is simple: in about 15 minutes, create a concise pre-meeting briefing, a list of hypotheses, risks, and a set of questions for the discovery call. This is not a futuristic concept. It is a combination of commonly available services that sales teams already use today: for example HubSpot AI, Salesforce Einstein, Microsoft Copilot for Outlook and Excel, ChatGPT, Claude, or Google Gemini. If you are still getting oriented in these tools, overviews on AIVýběr and topic-specific articles in the AI tools section are also useful.

An important note at the start: not all data should be sent to just any model. If you work with personal data, non-public pricing, contractual terms, or health or financial data, you must first address the legal framework, retention policy, and access permissions. The prices mentioned are indicative and may vary depending on licenses, region, and scope of use.

What the output of a 15-minute preparation should be

The most common mistake is not that the salesperson uses the wrong model. The mistake is that they do not know what the end result should be. If you only have AI generate a generic “company summary,” you will get text that does not help in the call. The output must be strictly structured and short.

What to do: introduce a fixed one-page brief template. It should contain at minimum: 1) the profile of the call participants, 2) a summary of the latest interactions from CRM and email, 3) the company’s likely priorities based on the website and public sources, 4) three working hypotheses about the problem, 5) five discovery questions, 6) two risks or sensitive topics, 7) the proposed goal of the call. This structure keeps AI within bounds and increases the chance that you get usable materials instead of a long text.

Who it is for: this is most beneficial for account executives, SDRs, and B2B salespeople with multiple meetings a day, typically in SaaS, IT services, consulting, and enterprise sales.

When not to use it: do not use it as the only basis for highly sensitive meetings, such as complaint handling, legal disputes, price escalations, or contract renewals after an incident. In those cases, every claim and the tone of communication must be checked manually.

The practical standard is this: the briefing must be readable in two minutes. If it is longer than 400 to 500 words, it is too long. AI should save the salesperson mental capacity before the call, not fill it with more reading.

Which data sources to involve and in what order

Not all materials have the same value. The highest value comes from data that captures the specific relationship with the client. That is why the correct order is almost always the same: first CRM, then emails, and finally the company website. Many teams do it the other way around and start by browsing the website. But the website tells you what the company wants to say about itself. CRM and email are more likely to show what it is actually dealing with.

What to do: prepare a three-step input. From CRM, pull open opportunities, deal value, pipeline stage, last activity, saved notes, and records of past calls. From email, take the latest relevant thread, ideally from the last 7 to 30 days. From the website, load the homepage, product pages, pricing, case studies, careers, and news. For most companies, these pages best reveal what the company sells, which segment it targets, and what is currently changing inside it.

Who it is for: for salespeople taking over a lead from marketing or a colleague and needing to quickly understand the context without a long handover.

When not to use it: it is not worth it for small transactional sales with a very short cycle, where the preparation takes longer than the call itself. Typically, for example, simple inbound inquiries with clearly defined pricing.

In CRM, prioritize fields that have business impact: contact role, number of users, estimated budget, reason for past loss, renewal date, connection to competing solutions. On the website, focus mainly on signals of change: new hiring positions, pricing changes, launch of a new service, regional expansion, mentions of partners, or acquisitions. These are better inputs for discovery than the company’s general mission and vision.

Practical workflow: 15 minutes from data to brief

A good workflow is short, repeatable, and divided into steps. Below is a version that can be done manually with common tools even without complex automation.

Minute 0 to 3: extracting CRM data

Open the company, deal, and contact records. Copy only the relevant parts: the last two to three notes, deal status, expected use case, known blockers, and the date of the next step. If your CRM supports its own summaries, use them as the first layer. HubSpot, for example, offers AI assistance for writing and summarization in various parts of the platform; Salesforce has Einstein features on top of CRM data; Microsoft has Copilot across Dynamics 365 and Microsoft 365. The scope of features depends on the license.

What to do: manually remove outdated notes and duplicates before pasting them into the model. AI cannot reliably distinguish which note was a mistake and which one is still valid.

Who it is for: for teams where multiple people write into one account and there is a lot of clutter in the CRM.

When not to use it: if the CRM is significantly neglected and the last relevant entry is half a year old. In that case, AI is more likely to reinforce an incorrect picture of the situation.

Minute 3 to 6: summarizing the email thread

In Outlook or Gmail, open the latest conversation and process only the thread related to the current meeting. Microsoft Copilot in Outlook can summarize threads and suggest replies; Google Gemini offers similar functions in Google Workspace depending on the plan. If you do not have this integration, paste the text manually into the model and ask for a structured summary: what the client explicitly wants, what they rejected, what deadlines were mentioned, who decides, and what remains unclear.

What to do: have the model label statements by certainty: “explicitly stated,” “inferred,” “unconfirmed.” This reduces the risk that you take an assumption into the call as if it were a fact.

Who it is for: for salespeople dealing with longer threads with multiple participants and losing track of who promised what.

When not to use it: do not use automatic summarization without manual review for emails containing contract attachments, non-public pricing, or personal data, unless you have an approved enterprise environment and clear processing rules.

Minute 6 to 9: quick website research

Only now should you go to the client’s website. The goal is not to do complete desk research. The goal is to supplement the picture of the company with publicly available signals that may not be in the CRM. You are interested in four things: who they sell to, what they are trying to emphasize, how they roughly monetize, and what is changing for them.

What to do: load the homepage, product page, pricing, case studies, and careers. Ask the model to extract from these pages: target segment, main product promises, possible operational priorities, and triggers of change. If the company is hiring, for example, RevOps, data engineer, or customer success roles, that may be a signal of growth or process change. If it publishes case studies from regulated industries, security and compliance may be important.

Who it is for: for B2B salespeople in segments where decision-making depends on processes, integrations, and internal approvals.

When not to use it: do not draw hard conclusions from the website for companies with weak marketing sites or sites that have not been updated for a long time. For many traditional companies, the website reflects reality only partially.

Minute 9 to 12: synthesis into a brief

Now combine CRM, email, and website into a single prompt with an instruction for a maximum of 300 to 400 words. The key is to forbid the model from making things up and require it to work with evidence.

What to do: use an instruction along the lines of: “Create a briefing for a discovery call. Distinguish between confirmed fact, probable hypothesis, and open question. If something does not follow from the materials, write that we do not know.” This single sentence will improve output quality more than long prompt wizardry.

Who it is for: for anyone who tends to overestimate AI summaries and adopt them without verification.

When not to use it: when you need a detailed account plan for multiple stakeholders and a longer sales cycle. That is no longer a 15-minute task, but a separate preparation effort.

Minute 12 to 15: turning it into questions and a call goal

The last step is the most important. A briefing alone will not make the meeting successful. You need to extract questions and a concrete goal from it.

What to do: have the model suggest five questions at three levels: verification, diagnostic, and prioritization. Verification questions confirm basic facts. Diagnostic questions uncover process, impact, and obstacles. Prioritization questions determine what must be fulfilled for the deal to move forward. Finally, manually write one sentence: “By the end of the call, I want to know X and agree on Y.”

Who it is for: for salespeople who tend to lead a pleasant conversation without a clear move toward the next step.

When not to use it: do not use generated questions without adjustment in industries with very specific terminology, such as healthcare, law, or cybersecurity. In those cases, a human must check the terminology.

Which tools make sense in practice

It is not necessary to deploy ten platforms. For most teams, one layer in the CRM or office suite and one universal model for synthesis are enough. What matters is where you already have data today and what the conditions are for safe use.

Claude

What to do: choose the tool based on where the salesperson spends the most time. If you work in Microsoft, it makes sense to try Microsoft Copilot for Outlook, Teams, Excel, and possibly Dynamics 365. If you are heavily invested in HubSpot, start with the features directly in HubSpot. If you need universal text work and custom instructions, ChatGPT, Claude, or Google Gemini in paid versions are suitable.

Who it is for: for companies that want a quick pilot without major integration.

When not to use it: do not buy a new tool just for meeting summaries if you already have the same function in limited form in your CRM or office suite. First measure where the real bottleneck actually is.

As a rough guide: ChatGPT Plus has long been around USD 20 per month per user, Claude Pro around USD 20 per month, Google One AI Premium with Gemini for individuals is often around USD 20 per month, and Microsoft Copilot for Microsoft 365 in enterprise licenses typically comes to around USD 30 per month per user. For CRM platforms, pricing is more complex and AI features are often tied to a specific plan or add-on license. These are indicative figures; always verify current terms in the vendor’s official pricing.

If you are choosing between general models, focus mainly on three things: handling long context, the ability to disable or limit data retention according to the service terms, and the quality of work with structured instructions. Comparisons of specific models and their strengths are also regularly covered in editorial content on AIVýběr – ChatGPT and other topic pages on the website.

How to write prompts that do not end in generic phrases

A prompt for sales preparation should not be a literary assignment. It must force the model to distinguish between fact, interpretation, and gaps in the information. This is crucial, because this is exactly where the most expensive mistakes arise: AI creates a smooth summary that sounds convincing, but in reality fills in what is not in the materials.

What to do: use a fixed prompt with four blocks: role, sources, output format, rules. For example: “You are a sales assistant preparing a discovery call. Use only the attached materials from CRM, email, and website. Divide the output into Confirmed facts, Hypotheses, Open questions, Risks, Recommended call goal. Do not cite anything that is not in the materials. If a piece of information is unclear, write We do not know.”

Who it is for: for teams that want multiple salespeople to get comparable output regardless of individual work style.

When not to use it: do not use one universal prompt for everything. A prompt for a discovery call is different from a prompt for a QBR, pricing call, or renewals.

A good practice is to add one more part: “Rank the questions by likely impact on deal qualification.” This limits the production of unnecessary questions. It is equally important to set the length. If you do not limit the length, the model will fill the space. If, on the other hand, you ask for 5 bullet points of one sentence each, you will get a clear output usable right before the meeting.

Three practical scenarios from sales practice

1. Inbound inquiry from a mid-sized company

In CRM, you have a new lead from a form, in email a confirmed meeting time, and on the client’s website several product pages. Here, AI mainly serves to quickly determine whether the company has an obvious use case and what questions to ask so the call does not get stuck on general expectations.

What to do: have the model extract the likely business model from the website and formulate the first problem hypothesis from the form plus the email. Then prepare three questions about impact, urgency, and the decision-making process.

Who it is for: for SDRs or AEs who need to qualify a lead quickly without lengthy research.

When not to use it: if the client stated a very specific request in the inquiry and wants a demo of a specific feature right away. In that case, product preparation takes priority over general discovery.

2. Contract renewal and churn risk

In CRM, you see an upcoming renewal, in emails there are complaints about support, and meanwhile the client is hiring its own specialists on the website. That may indicate an internal effort to take over part of the agenda or pressure to reduce costs.

What to do: use AI to create a list of warning signals from the last 90 days: delayed responses, unused licenses, repeated complaints, team changes. From that, derive questions about adoption, value, and the future plan.

Who it is for: for account managers and customer success teams.

When not to use it: do not use automatic conclusions such as “the client is leaving” if you only have weak indications. In renewals, false conclusions are especially expensive.

3. Enterprise meeting with multiple participants

From CRM, you know that a business sponsor, an operations manager, and someone from IT will join the call. The email thread is long and each person is dealing with something different. Here, the main value of AI is in sorting stakeholders and their priorities.

What to do: have the model assign each participant a likely interest: business impact, process, security, integrations, budget. Then prepare one question for each person and one shared prioritization question.

Who it is for: for enterprise AEs and solution sales roles.

When not to use it: do not treat such mapping as a finished stakeholder analysis if some people are not on the website or in the CRM and you only know them from an email signature.

Limits, risks, and rules worth introducing

AI meeting preparation is useful only to the extent that you can manage its limits. The first limit is the quality of the input data. Bad CRM produces a bad briefing. The second limit is hallucination: the model can smoothly fill in missing links. The third limit is security and compliance.

What to do: introduce three mandatory rules. First, every briefing must visually separate facts from hypotheses. Second, no salesperson may put into an unapproved tool personal data beyond the necessary minimum, non-public contract attachments, or sensitive pricing terms. Third, before the meeting, at least three key points must be manually verified: the goal of the call, the last agreed step, and the current roles of the participants.

Who it is for: for all teams that want to use AI at a larger scale and not just ad hoc.

When not to use it: if the company still does not have even basic rules for working with data and access. Without governance, a quick pilot is often more expensive than a slow start.

A practical limit is also time. If the preparation grows to 25 minutes and three tools, it loses its point. The goal is not to maximize the amount of text, but to improve the quality of the first part of the call. If after a month of use you do not see a shorter time to the first relevant question, less repetition of information, and better follow-up steps, adjust or shorten the workflow.

How to measure whether the workflow really works

Without measurement, AI preparation quickly becomes just another routine the team does “because they are asked to.” But you do not need to measure in a complicated way. What matters are indicators that are tied to the actual course of the call and the business outcome.

What to do: track four simple metrics: 1) meeting preparation time, 2) the share of calls where a specific next step was agreed by the end, 3) the number of situations where the client had to explain the basic context again, 4) the salesperson’s subjective rating of the brief’s usefulness after the meeting on a scale of 1 to 5. After two weeks, compare calls with a brief and without a brief.

Who it is for: for sales team leaders who need to decide whether to roll the workflow out to the whole team.

When not to use it: do not attribute improvement only to AI if you are simultaneously changing the call script, lead segmentation, or pricing. In that case, changes need to be evaluated more carefully.

A good sign is when the opening of the call gets shorter and the salesperson gets to specific questions within a few minutes. An even better sign is when the briefing helps reveal a mismatch between what is in the CRM and what the client is actually dealing with. That is exactly the moment when AI is not writing text for effect, but improving the quality of sales work.

FAQ

Is ChatGPT or another general model alone enough for this?

Yes, for a manual workflow it is often enough. But you need well-prepared inputs and a fixed output format. If you already have a CRM or email suite with built-in AI, it is usually more practical to start there and use a general model only for the final synthesis.

How long an email thread is worth summarizing?

Typically the latest relevant thread from the last 7 to 30 days. For longer sales cycles, it makes sense to add only a few key older messages, not the entire archive. The more clutter you put in, the worse the output tends to be.

Can a client’s priorities be reliably identified from their website?

Only partially. A website is good for signals, not definitive conclusions. Treat it as a supplement to CRM and emails, not as the main source of truth.

How much does it cost per user?

For general models, paid plans for individuals often roughly come to around USD 20 per month. Enterprise licenses in the Microsoft ecosystem or CRM platforms tend to be more expensive and more dependent on the specific plan. These are indicative figures.

Where is the biggest risk of errors?

In mixing fact and hypothesis. If the model does not structure the output into separate blocks and the salesperson does not check it, it is very easy to take an assumption as a certainty.

How do you get started without integrations?

Very simply: manually copy the relevant parts from CRM, email, and website into one template and have the model create a briefing. First verify that the output actually helps in calls. Only then deal with automation.

Conclusion

AI preparation for a discovery call is not about the model “knowing the client better than the salesperson.” It is about assembling scattered information within 15 minutes into a form that enables a better first part of the meeting. The biggest benefit does not arise when you get a longer summary. It arises when the salesperson comes to the call with three verified facts, two reasonable hypotheses, and five questions that have a clear purpose.

Start with a small pilot. Take five to ten sales calls, use a fixed brief template, separate facts from hypotheses, and after two weeks evaluate whether preparation time has shortened and discovery quality has improved. If yes, only then deal with integrations and broader deployment. In sales, the winner is not the one who deploys the most AI features. The winner is the one who gets to the core of the call faster because of them.

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