Most CSMs know AI could save them time, but "use AI" is not a workflow. You need specific prompts, a clear idea of which data is safe to share, and a habit of reviewing the output before it reaches a customer. This guide gives you six copyable prompts for the jobs that eat the most hours: account research, call prep, call follow-up, risk review, renewal prep and QBR drafts.

Where AI actually helps a CSM

AI is strongest on work that is repetitive, text-heavy and easy to check. It is weakest where judgment and relationships matter.

Planhat's guide for CSMs lists the same short set of use cases: pre-call briefs, post-call summaries and follow-up emails, QBR preparation, account prioritization, and re-engagement or handover notes. Planhat sells a CS platform, and its time estimates (for example, pre-call research dropping from 20-30 minutes to a 3-minute review) are its own and vary by call type, so treat them as indicative rather than measured. The pattern is still useful: AI gathers and drafts, you decide and send.

ChurnZero's 2026 customer success trends panel is similarly forward-looking. Its CEO predicts the average CSM will have 25-50% more bandwidth by the end of 2026, which is a prediction, not a result. Another panelist, Anika Zubair of The Customer Success Pro, put the stakes plainly: AI will not replace CSMs, but will replace the teams that never learn how to use it.

Before you paste anything: a data checklist

Every prompt that contains a customer name, email, call transcript or contract term is likely to involve personal or confidential data. Planhat's guidance is to check your company's data policy first and avoid putting customer PII, contract terms or confidential pricing into prompts sent to external AI services. The UK ICO's guidance on AI and data protection covers data minimisation, accountability and DPIAs (data protection impact assessments), and is a good starting point for your privacy team. This is not legal advice; ask your DPO or counsel about your own setup.

A practical checklist:

  • Your company has approved the AI tool you are using for customer data.
  • You share only what the task needs (a health summary, not the whole contact list).
  • Individual names and emails are replaced with roles ("the VP Operations") where possible.
  • Contract terms and pricing stay out unless the tool is approved for them.
  • A human reads every output before it is sent to a customer.
  • You can tell where each claim in the output came from.

Six prompts you can copy

Replace the bracketed parts. Paste in only the data the prompt asks for.

1. Account research before a first call

You are helping a customer success manager prepare for a first call.
Account: [company name], [industry], about [size] employees.
Goal of our product for them: [one sentence].
Task: List 5 recent public facts about this company that could affect
how they use our product (hiring, funding, product launches, leadership
changes). For each, give the source URL. If you cannot find a source,
say so instead of guessing. Then suggest 3 discovery questions.

The "give the source or say so" line matters. Assistants can state plausible but wrong facts, so verify anything you plan to say out loud.

2. A pre-call brief

Prepare a one-page brief for my call with [customer role] at [account].
Context I am pasting below: last call summary, open tasks, open support
tickets, usage trend (last 90 days), renewal date.
Output: (1) Where the relationship stands in 3 sentences. (2) Three
things to raise. (3) One risk to probe gently. (4) One question to
ask about their goals.
Do not invent facts. Mark anything missing as "unknown".

3. Call notes and follow-up email

Here are my notes from the call: [notes, with names removed].
Produce: (1) Decisions made. (2) Action items with owner and date,
as a checklist. (3) A follow-up email under 150 words, plain tone,
confirming the actions. Do not add commitments that are not in the notes.

The last line prevents the most common failure: the assistant promising something you never agreed to.

4. A 30-minute risk review

I manage [N] accounts. Below is a table with: account, ARR, renewal date,
health score, usage change vs. prior 30 days, days since last contact,
open tickets.
Task: Rank the 5 accounts most at risk in the next 90 days. For each,
give the two data points that drove the ranking and one specific next
action. Flag any account where the data conflicts (for example,
healthy score but no contact in 60 days).

5. Renewal prep

Renewal is in [N] days for [account], ARR [amount].
Known: goals we agreed, outcomes delivered, open issues, stakeholders
and their roles, any budget signals.
Task: Draft a renewal brief with: value delivered (only from the facts
above), risks, the likely objections, and a recommended renewal
conversation plan. List what information is still missing.

Pair this with a timeline such as our 120-day renewal checklist.

6. A QBR first draft

Draft a QBR narrative for [account] using these facts: goals set last
quarter, results against each goal, usage highlights, support summary,
and next-quarter priorities.
Structure: what we set out to do, what happened, what it means for
their business, what we recommend next.
Keep claims tied to the facts. Mark any gap as "needs input".

Planhat warns that executives "will notice if it's generic," and describes QBR output as a first draft, not a final. Add your own judgment on what matters. Our QBR agenda template shows what the finished meeting should cover.

A worked example

This is an illustrative example, not a real customer. Imagine a CSM with 40 accounts who runs a Monday risk review. They export one row per account (health score, usage change, days since last contact, renewal date, open tickets), remove contact names, and use prompt 4.

The assistant returns five accounts. For the first, it notes a 35% usage drop and no contact in 52 days, and suggests booking a call with the sponsor before the renewal enters its 90-day window. The CSM checks the numbers against the source system, sees one is off because a user migration skewed usage, and corrects the prompt. The assistant did the sorting; the CSM did the verifying. That division of labor is the whole point.

Common mistakes

  1. Pasting everything. More data is not better. It raises privacy risk and buries the signal.
  2. Trusting numbers the model calculated. Do arithmetic (GRR, NRR, ARR at risk) in your system of record and give the assistant the results.
  3. Sending unreviewed drafts. An email that tells an at-risk customer you "noticed low engagement" can sound surveillant, as Planhat points out. Rewrite it.
  4. No shared prompts. If every CSM invents their own, quality varies. Keep a shared prompt library and refine it monthly.
  5. Skipping the privacy conversation. Decide as a team which tools and data are approved before habits form.

Getting the data to the assistant

Copy-pasting works, but it is slow and easy to get wrong. A better setup is an assistant that reads your CRM directly, with the same permissions you have. XPI CRM has an AI connector (MCP) that works with Claude, ChatGPT, Gemini, Grok, Microsoft Copilot, Cursor and VS Code. It acts as the signed-in user with their permissions, cannot delete anything, and every change shows in History. Setup is covered in the AI connector docs, and we have a step-by-step guide to connecting ChatGPT. XPI CRM stores data in the EU (Ireland) and offers a GDPR Data Processing Agreement at xpicrm.com/legal/dpa, which may help your privacy review.

Start small

Pick one prompt from above and use it on every account for two weeks. Track how long the task takes before and after, and how many corrections you make to the output. Keep what works, and drop what does not.

Try it in XPI CRM

If you want your health scores, renewals, tickets and meeting history in one place that an AI assistant can read, XPI CRM is built for that. Start free — every feature included, no credit card.