English

AI in the AM Workflow: What Saves Time Without Breaking Trust

An AM I worked with walked into a quarterly business review last fall with a 22-slide deck her AI assistant had drafted overnight. It looked clean. The charts rendered. The narrative arc made sense in preview.

Halfway through slide 4, the customer's VP of Operations stopped her. "That number's not right. We were at 412 active seats in March, not 380. And that quote on the next slide, none of us said that."

The room went quiet. The AM had skimmed the deck on the train that morning, decided it looked fine, and trusted the model.

The renewal stalled a quarter. Not because the AM was bad. She was actually one of the better ones on the team. She just outsourced a piece of work to AI that AI is genuinely bad at, and she didn't catch it because the output looked confident.

This is the AM-AI problem in one sentence: trust is your product, and AI is uniquely good at producing things that look trustworthy and aren't.

Why This Matters Now

A developer using AI ships a bug. It gets caught in code review, or in staging, or by a test. There's a system between the AI and the customer.

An account manager using AI badly emails a customer the wrong renewal terms. Or quotes a stat that doesn't exist. Or sends a follow-up that uses phrases the customer's CFO has never heard from them before. There's no system between the AI and the customer. There's just you, and how carefully you read what it gave you.

Meanwhile, the AMs who refuse AI entirely are losing 5 to 10 hours a week to admin work their peers have automated. Meeting prep, note-taking, follow-up drafting, usage trend scanning: all things AI handles well when treated as a draft generator instead of a finished product.

So the question isn't whether you use AI. You'll use it or you'll fall behind. The question is where. This guide is about the line — where AI saves you time, and where it quietly destroys the relationship you spent two years building.

Where AI Helps (And Where It Hurts)

Before the prompts, here's the map. Tape it to your monitor if you have to.

AI helps with:

  • Meeting prep. Summarizing 90 days of email, Slack, and ticket history into a one-pager. This is pure synthesis from data you already have. AI is great at it.
  • Post-call notes. Turning a transcript into structured action items, owners, and dates. You'll always edit before sending, but the bones are right.
  • Pattern detection across usage data. Feeding 6 months of logins, feature usage, and ticket volume to AI and asking what looks unusual. Treat output as hypotheses, never verdicts.
  • First-draft writing. Getting past the blank page on a follow-up email or executive summary. Then you rewrite in your voice.

AI hurts with:

  • Customer-facing comms unedited. Copy-paste-send is the fastest way to sound like a stranger. Your customers can tell.
  • End-to-end QBR generation. AI hallucinates numbers, misattributes quotes, fabricates feature names, and produces decks that look polished but fall apart under questioning.
  • Escalation tone. Tone is judgment. AI defaults to corporate-soft when the moment needs directness, or to bluntness when the relationship needs care. It cannot read the room.
  • Hard conversations. When you write a tough renewal email through AI to soften it, the customer feels the distance. They can tell you're hiding behind something.

The difference between the help column and the hurt column is judgment. AI is good at synthesis and bad at judgment. Put it where synthesis is needed. Keep it out of where judgment is needed.

The AM Prompt Library

Seven prompts I've tested across roughly 200 customer conversations. They're long because real prompts are long. Vendor demos use 8-word prompts because those are what fit on a slide. Real prompts give the model context, constraints, and a clear output shape.

1. Meeting Prep One-Pager

Drop this before any customer call where you don't have time to re-read 90 days of history.

You are preparing me for a 30-minute customer call with [CUSTOMER NAME],
account owner [CONTACT NAME], on [DATE].

I'm pasting below: the last 90 days of email threads with this account,
the last 60 days of Slack messages from our shared channel, and the
last 30 days of support tickets they've opened.

Produce a one-pager with these sections:
1. Top 3 themes from their communication (not feature requests — themes,
   like "frustrated with onboarding speed" or "expanding to a new team")
2. Open commitments WE have made to THEM that aren't yet closed
3. Open commitments THEY have made to US (e.g., "we'll send the security
   review by April 1")
4. Anything that has changed in their tone in the last 30 days vs. the
   prior 60 (more positive, more terse, slower to respond)
5. Three questions I should ask in this call to advance the relationship

Do not invent feature names, contract terms, or numbers. If something
isn't in the source data, say so explicitly.

The "do not invent" line is non-negotiable. Without it the model will fill gaps confidently. With it, you'll get "the source data does not specify the contract end date" instead of a hallucinated date.

2. Post-Call Action Item Extractor

Run this on the transcript right after the call, while you still remember the meeting.

Below is a transcript from a customer call between me and [CUSTOMER NAME].
Extract action items into this exact format:

| Action | Owner (us / customer) | Due date if mentioned | Quote from transcript |

Rules:
- Only include items where someone explicitly committed to do something.
  "We should look into X" is not an action item. "I'll send you X by
  Friday" is.
- If the owner is ambiguous, mark it AMBIGUOUS and quote the line.
- Do not infer due dates that weren't said.
- After the table, list any decisions that were made (e.g., "agreed to
  pilot the analytics module for 30 days").

End with one section called "Things I should follow up on this week"
based only on what was actually said in the call.

The quote column is the safety net. If the action looks weird, you can verify it against the actual sentence in seconds.

3. Churn Signal Scan

Run this monthly per account, or quarterly across your full book.

Below is usage data for [CUSTOMER NAME] for the last 6 months. Columns
are: date, weekly active users, key feature adoption %, support ticket
count, NPS score (if collected that month), executive sponsor login
frequency.

Identify patterns that historically correlate with churn risk in B2B
SaaS:
- Drop in weekly active users >20% sustained for 3+ weeks
- Decline in executive sponsor engagement
- Rising support ticket volume on core features (not edge features)
- Feature adoption regression after an initial rise
- NPS decline of 2+ points

For each pattern you find, give me:
1. The pattern name
2. The specific weeks where it appears
3. A confidence score (low / medium / high) based on how clean the signal is
4. One question I could ask the customer to test the hypothesis

Be skeptical. If the data is too noisy to draw a conclusion, say so.
Do not generate a "churn risk score" — give me hypotheses, not verdicts.

The "do not generate a churn risk score" line is there because once a number is in the prompt output, AMs treat it like fact. Hypotheses force you to go talk to the customer. That's the goal.

4. Expansion Opportunity Finder

Use this before any account planning session.

Below is the account record for [CUSTOMER NAME]: current contract, seats
purchased, modules active, modules NOT active but available on their
tier, usage data for the last quarter, and notes from my last 4 calls
with them.

Find expansion opportunities and rank them by likelihood-to-close in the
next 90 days, not by deal size.

For each opportunity, give me:
1. The specific module or seat expansion
2. The signal in their behavior or comments that points to it (quote
   from notes if possible)
3. The likely objection they'd raise
4. Who internally at the customer would champion this

Do not include opportunities that would require them to switch tiers up
unless there's clear signal they're ready. Do not pad the list — three
real opportunities are better than seven speculative ones.

"Do not pad the list" matters. AI defaults to giving you the number of items the prompt suggests, even if only two are real.

5. Executive Update Draft

For the monthly note to the customer's exec sponsor.

Draft a 200-word executive update from me to [SPONSOR NAME] at
[CUSTOMER NAME]. Voice: confident peer, not vendor. They are a [TITLE]
who cares about [SPECIFIC BUSINESS OUTCOME they've mentioned in past
calls — paste below].

Source material below: their usage trends this month, the 2 wins their
team had, the 1 issue we're working on, and the 1 commitment we're
making for next month.

Format:
- Open with a specific outcome from their team (not a generic greeting)
- 2 sentences on momentum
- 1 sentence on the issue and our owner on it
- 1 sentence on what they should expect next month
- Sign-off

Do not use the phrases "I hope this finds you well," "circling back,"
"touching base," "leveraging," or "synergies." Do not start with
"Just wanted to..." Do not end with "Let me know if you have any
questions."

The blocked-phrases list is what saves this from sounding like AI. Every AM should keep their own list of phrases they don't say. Add them to the prompt.

6. Renewal Email v0

This is for hard renewals: pricing increase, terms changing, or a customer that's been quiet.

I need to draft a renewal conversation opener for [CUSTOMER NAME].
Context: their renewal is in 60 days. This year's contract is
$[AMOUNT]. The renewal pricing is going to be $[AMOUNT] (a [%] increase
because of [SPECIFIC REASON — paste]). The customer has had [SPECIFIC
WINS] this year and [SPECIFIC ISSUES].

Draft an email that:
- Opens by referencing a specific win from their team this year
- States the renewal timing clearly
- Names the price change directly without softening it ("the new pricing
  is $X" — not "we have some updates to share around investment levels")
- Asks for a 30-minute conversation to walk through it
- Does not justify the increase in the email — that's the conversation

This is a v0. I will rewrite it in my voice. Give me the bones.

Note the explicit "do not soften the price change." Left to itself, AI will write four paragraphs of justification before mentioning the new number. Customers see that as evasion. Be direct.

7. Internal Account Brief

For when you're handing the account to a teammate, briefing your manager, or prepping for a deal review.

Build an internal-only account brief for [CUSTOMER NAME] using the
sources below: contract, contacts and roles, last 6 months of usage,
last 4 call notes, support history, expansion history, and current open
risks/opportunities.

Format (max 1 page):
1. Account snapshot — ARR, contract dates, key contacts with role and
   our relationship temperature (warm / neutral / cool)
2. Health — green / yellow / red on: product usage, executive
   engagement, support load, payment history. Justify each color in one
   sentence.
3. Top 3 risks (with the specific signal, not generic concerns)
4. Top 3 opportunities (same standard)
5. What I would tell my replacement on day 1

Internal only — be blunt. If I think a contact is checked out, say
"appears checked out" not "engagement could be stronger."

The "be blunt, internal only" framing matters. Without it, AI defaults to vendor-speak even for internal docs.

The "AI Here, Not There" Decision Tree

When you're not sure whether to reach for AI, walk this in order. If any answer puts you in the "don't use AI" lane, stop there.

  1. Is the output going directly to a customer with no review? → Don't use AI. Period.
  2. Does the output contain numbers (usage, contract, pricing)? → Use AI for the draft, but verify every number against the source before sending. If you don't have time to verify, don't use AI.
  3. Does the moment require tone judgment (escalation, bad news, conflict)? → Don't use AI for the language. You can use it to think through the framing, but write the words yourself.
  4. Does the task involve summarizing data you already have access to? → Use AI. This is the sweet spot.
  5. Is the task a first draft of writing you'll edit? → Use AI for v0. Always rewrite in your voice. Never send v0.
  6. Are you reaching for AI because the conversation is hard? → Don't use AI. Have the conversation.

The sixth question catches more disasters than the other five combined.

The Output Review Checklist

Before any AI output leaves your machine, go through these seven items. It takes 90 seconds. It saves quarters.

  1. Numbers checked. Every number in the output exists in the source data, with the same value.
  2. Quotes verified. Every quoted line was actually said by the person it's attributed to.
  3. No hallucinated features. No reference to a product capability that doesn't exist in your stack.
  4. No fabricated dates. No "as we discussed on March 14th" if you didn't discuss it on March 14th.
  5. Tone matches relationship stage. Warm with warm accounts, direct with cool ones. Generic friendly with neither.
  6. Action items have owners and dates. No floating "we should look into this."
  7. You'd be comfortable if it leaked. Internal channel, customer's lawyer, your own VP. Would any of them flag something?

Print it. Tape it to your monitor next to the decision tree.

Common Pitfalls

Sending raw AI output to a customer. Not "lightly edited." Raw. Customers learn to spot it within two emails. They stop replying with substance because they're not sure they're talking to a person. The relationship goes flat without any one moment you can point to.

Generating a QBR deck end-to-end and showing it live. It will hold up in preview. It will break in the room. The fix is simple: never present an AI-generated deck without rebuilding the three most important slides by hand. Those are usually the executive summary, the ROI / outcome slide, and the renewal narrative.

Treating an AI churn risk score as gospel. AI says "high churn risk." The AM panics, calls the customer, over-discounts, and trains the customer that pressure works. Six months later the customer renews, and asks for the same discount again. The score wasn't wrong; the response to it was.

Skipping the review step "just this once." This is how every AI-in-AM disaster starts. There is no "just this once." You either review every AI output or you'll eventually send something you regret.

Using AI to dodge a hard conversation. Writing a tough renewal email through AI because it feels easier than calling. The customer feels the distance. AI is not for hiding behind. If the conversation is hard, have it.

Measuring Whether This Is Working

The numbers that matter:

  • Admin time saved per week. Target: 5-8 hours back, redirected to actual customer-facing time. If you're saving time and not putting it into customer conversations, you're just doing more admin.
  • QBR prep time. Target: from 6 hours to 2 hours, with same or better deck quality. If quality drops, you've crossed the line.
  • CSAT on AM-touch interactions. This should stay flat or rise after AI adoption. If it falls, customers are noticing.
  • Customer-facing comms walked back per quarter. This should trend toward zero. Every "actually, I want to correct what I sent yesterday" is a trust withdrawal.

For the full picture of where AM time goes day to day and which hours AI can recover, see a day in the life of an account manager. For the specific case of QBRs (where AI helps in prep and hurts in generation), read QBRs that drive expansion. For where these tools fit alongside the rest of your stack, see the AM tools and tech stack. And for the failure modes that compound when AI hides them, common pitfalls that sink AM careers is worth a read.

How Rework Fits In

Most AMs end up running their AI workflow in five different surfaces: call notes in one tool, account history in another, AI prompts in a chat window, follow-up drafts in email, action items in a task app. The friction that AI was supposed to remove gets reintroduced by the tool sprawl.

Rework CRM keeps the source data (usage trends, contract details, contact history, call notes) in one place, so the prompts above actually have something to read. And Rework Work Ops tracks the action items AI extracts, with owners and due dates, so commitments don't fall off after the call. CRM starts at $12/user/month, Work Ops at $6/user/month.

What This Comes Down To

Trust is your product. AI is a force multiplier on the parts of the job that don't require trust to produce: synthesis, drafting, pattern-spotting. It's a liability on the parts that do: tone, judgment, commitment, the conversation.

The AMs who win the next two years aren't the ones who refuse AI or the ones who outsource everything to it. They're the ones who learned the line, drew it in pen, and never crossed it under deadline pressure. The deck on slide 4 was wrong because the AM crossed the line on a Tuesday morning. Don't be that AM.

Frequently Asked Questions About AI in the AM Workflow

How much time should AI realistically save an account manager per week?

Realistic target is 5-8 hours per week, mostly on meeting prep, post-call notes, and follow-up drafting. AMs claiming 15+ hours per week are usually either skipping the review step (and creating downstream problems they're not counting) or counting time savings on tasks they should never have been doing manually in the first place. The savings need to redirect into customer-facing time, not just more admin — that's the test.

What's the single biggest AI mistake AMs make?

Sending raw or lightly-edited AI output directly to customers. Customers learn to detect it within two or three emails — the rhythm is too even, the phrasing is too generic, the warmth feels manufactured. They stop replying with substance because they're not sure they're talking to a person anymore. The relationship goes flat without a single dramatic moment, which is what makes it hard to spot until renewal time.

Should I use AI to generate QBR decks?

Use AI for prep — synthesizing usage data, surfacing patterns, drafting talking points. Do not use AI to generate the final deck end-to-end. Always rebuild the three most important slides by hand: the executive summary, the ROI / business outcome slide, and the renewal narrative. AI hallucinates numbers and misattributes quotes confidently. Decks look fine in preview and fall apart under questioning in the room. The hand-built backup of the critical slides is non-negotiable.

Is it okay to use AI for customer-facing emails?

Use it for v0 — the rough draft that gets you past the blank page. Always rewrite in your voice before sending. Maintain a list of phrases you do not use ("circling back," "touching base," "leveraging," "just wanted to," "synergies") and add them to your prompts as blocked phrases. The rewrite step is what makes the difference between AI as a tool and AI as a tell.

How do I evaluate AI churn risk scores?

Treat them as hypotheses, not verdicts. AI is good at spotting patterns in usage data — usage drops, executive disengagement, ticket volume spikes — but bad at judging which patterns matter for a specific customer. The right response to a "high churn risk" output is a conversation with the customer, not an over-discount or a panic email. The customer will train you to overreact if you let scores drive your behavior.

What should AMs not let AI do, ever?

Three things, no exceptions. (1) Compose escalation or bad-news communications without a human writing the actual words — tone is judgment and AI cannot read the room. (2) Generate a final QBR deck without slide-level human review of every number and quote. (3) Replace a hard conversation. If you're reaching for AI because the conversation feels uncomfortable, that's exactly when you need to write it yourself or pick up the phone.

How do I know if AI is actually helping or quietly hurting?

Track four numbers per quarter. Admin time saved per week (target 5-8 hours, redirected to customer-facing time). QBR prep time (target down from 6 hours to 2 hours, with quality flat or better). CSAT on AM-touch interactions (should stay flat or rise — falling CSAT is the signal customers can tell). Customer-facing communications walked back or corrected per quarter (target trending toward zero — every walk-back is a trust withdrawal). If those four numbers are trending right, AI is helping. If any one is going the wrong way, it's hurting and you haven't seen it yet.

Learn More