AI in a CRM: what it does and does not do

An AI assistant in a CRM is almost always a language model with tools. The model itself knows nothing about your customers. It was trained on a lot of text and is good at reading, summarizing and writing. What makes it useful inside a CRM are the tools: small, fixed functions the model is allowed to call. Find an organization. Fetch the open deals for this person. Draft an activity.

Ask "which deals in the proposal stage have had no activity for two weeks?" and roughly this happens:

• The model reads your question and picks a tool, for example "search deals" with a filter on stage and last activity.
• The CRM runs that search with your permissions and returns the result.
• The model reads the result and turns it into an answer in plain language.
• If you then ask "schedule a call for each of them tomorrow", the model drafts those activities and, if it is set up well, asks for your approval.

That covers three kinds of work. Read: look up data and explain it. Draft: prepare an email, a segment, a report or an automation without changing anything. Act: actually save, send or switch something on. That split is the core of this guide. When reading, the worst outcome is a wrong answer. When acting, an email can go to two thousand people.

Worth knowing: the model does not remember your data between conversations unless the vendor builds it that way. It fetches what it needs every time. That is good for freshness, because it reads the current state. It is also why an assistant is only as good as the data in your CRM. If a deal has sat in the wrong stage for three months, the assistant will report that wrong stage back to you, neatly and with confidence.

So an AI assistant is not a separate brain that takes over your sales. It is another way to operate your CRM: one sentence instead of ten clicks. That saves time on work you can describe precisely. It helps little with work that needs judgement.

It sometimes makes things up. A language model writes what sounds likely. Most of the time that matches the data it just fetched. Sometimes it fills a gap with something that is not there: a job title, a date, an amount. This is called hallucination. Better tools make it rarer, but never zero. Any number from an assistant that you pass on to a customer or to your board, check in the record first.

It only knows what is in the CRM. Yesterday's phone call that nobody logged does not exist for the assistant. Neither does the meeting a colleague put in a private calendar. Ask "how are things with this customer?" and you get a summary of what was recorded, not of what happened.

It has no judgement. An assistant can list the ten deals that have been quiet the longest. It does not know that deal number three is quiet because the buyer is on holiday and that this is fine. It does not know that this buyer prefers a call to an email. Prioritizing, negotiating and deciding whether a lead fits remain human work.

It cannot count what was never defined. "How are we doing this quarter?" is not a question an assistant can answer unless you agreed what doing well means. "What is the win rate this quarter, won divided by won plus lost?" is. The more precise the question, the more reliable the answer. The guide on sales KPIs shows how to write those definitions.

It does not fix bad data. Duplicate organizations, empty fields and outdated stages flow straight into every answer. An assistant on a messy CRM gives messy answers, only faster. If in doubt, start with data quality.

For each use below you get what it delivers, what can go wrong and which safeguard to build in. The level in brackets is read, draft or act.

1. Search in plain language (read) Delivers: "all organizations in Manchester with an open deal over ten thousand" without building filters. Risk: the assistant picks a different filter than you meant, such as the person's city instead of the organization's. Safeguard: have it show the filter it used and click through to the list.

2. Summarize a deal (read) Delivers: in thirty seconds you know what is going on before you call. Risk: key information sits in an attachment or in someone's head and is missing. Safeguard: treat the summary as a start and read the last two notes yourself.

3. Find stalled deals (read) Delivers: a list to pick up this week. Risk: deals that are waiting on purpose show up too. Safeguard: the rep decides per deal. The assistant only supplies the list.

4. Write a follow-up email (draft) Delivers: a first version based on the last note. Risk: a promise on price or timing that was never made. Safeguard: the email is always a draft and you send it yourself. The guide on following up covers what to write.

5. Clean up a note or meeting summary (draft) Delivers: rough notes become a readable summary with agreed actions. Risk: nuance gets lost. Safeguard: keep the original or read the summary back before saving.

6. Update fields (act) Delivers: "set the expected close date of these three deals to the end of March" in one sentence. Risk: the wrong deals or the wrong date. Safeguard: show what will change, ask for approval before saving, and keep a log.

7. Schedule activities (act) Delivers: a batch of calls planned at once. Risk: calls on a day off or duplicates. Safeguard: drafts first, then approval.

8. Build a segment (draft) Delivers: "customers who bought nothing last year and gave marketing consent" as a filter. Risk: people without consent slip through. Safeguard: the CRM checks consent itself at the moment of sending, not the assistant.

9. Send a campaign (act, outward) Delivers: a faster route from idea to sent email. Risk: the biggest mistake an assistant can make: the wrong text to the wrong list. Safeguard: always a test send, always explicit approval, ideally a second person who checks.

10. Create a report from a sentence (draft) Delivers: "won revenue per month this year, by owner" as a chart. Risk: a different definition than your meeting uses. Safeguard: check the measure and the period before you share the report.

11. Prepare an automation (draft, then act) Delivers: a rule such as "create a task when a deal has been quiet for fourteen days" without learning the builder. Risk: a rule that fires on hundreds of records at once. Safeguard: a dry run first, then switch it on.

12. Spot duplicates and gaps (read) Delivers: a list of organizations without an owner or people without an email address. Risk: low, as long as the assistant only reports. Safeguard: do the merging in the dedupe tool, not through the chat.

A good question to an assistant is precise about three things: which records, which period and what you want back. Below are templates to copy. Fill in the parts in brackets.

Preparing a call. "Summarize the deal [name]: stage, value, last three activities and open tasks. Only use what is in the CRM and say so if something is missing."

Following up. "Write a short follow-up email to [person] after our call on [date]. Use the last note. Make no promises on price or delivery time. At most 120 words, tone: businesslike and friendly."

Monday start. "Show my open deals in the stage [stage] with no activity in the last [number] days, sorted by value, with the owner."

Report. "Create a report: number of won deals per month this calendar year, split by pipeline. Use the date the deal was won."

Segment. "Build a segment of people at organizations in the [industry] industry who gave marketing consent and received no campaign in the last [number] months. Show the count first."

Clean-up. "Show organizations without an owner and people without an email address created in the last [number] months. Change nothing."

Note the last sentence of the clean-up template. "Change nothing" and "show the count first" are small phrases that keep a read question a read question. And ask for sources: "name the record for every number" makes checking much faster.

Do not measure an assistant by the number of messages. That only shows people open it. Pick two or three questions you can answer before and after the rollout:

• How long does preparing the weekly sales meeting take? Ask the manager in week 1 and in week 5.
• How many open deals have a next activity scheduled? You can see that in the CRM itself.
• How many drafts get used as is or with light edits, and how many are thrown away? Have the pilot users keep a tally for a week.
• How many mistakes did you find in the log? A mistake is an answer or action you had to undo.

Write down the starting point before you begin. Without a baseline every review becomes a feeling.

Pitfalls Switching everything on at once. Read, draft and send on day one leaves little time to learn where it goes wrong. Start with read.

Approving on autopilot. Someone who clicks approve twenty times a day stops reading at some point. Keep the actions that need approval limited to the ones that matter.

Treating the assistant as the source of truth. A chat answer is a summary of the CRM, not a replacement. If a number goes into a board report, it comes from the report.

Sensitive data on a free model. Some free plans with model providers may use your input for training. Pick a plan or key where that does not happen before you work with real customer data.

Expecting AI to score your leads. An assistant that reads does not produce a reliable score by itself. A score you want to steer on needs agreed criteria.