twimco.Let’s talk ↗︎
Menu
Field notes / AI & judgment

The AI summary dropped the one thing the customer said twice.

"Customer happy, wants to renew, send quote." It reads well, it's mostly true, and it leaves out the condition the whole renewal depends on. Reading one call against its summary, line by line, shows what to ask these tools for and what a person still has to check.

Editorial engraving of a bright meeting nook where an account manager highlights one passage on a long ribbon of printed transcript beside a laptop showing a short summary, as a colleague with a headset leans in to point at it.

AI call summaries tend to be fluent and broadly accurate, and still lose the details that change what happens next: conditions, commitments your team made, numbers and dates, who said what, and hesitation. Ask for structure instead of a paragraph, with commitments, conditions, open questions and figures each linked to the moment in the recording. Have the person who was on the call confirm anything that changes a record, and test the tool on past calls where you already know what mattered.

A good summary of the wrong call

Here's a summary of the kind an AI note-taker produces after a customer call. The call itself is made up for this note, but the shape will be familiar.

Customer is happy with the service and wants to renew for another year. Asked about pricing for next year. Next step: send renewal quote.

It's tidy. Every sentence is true. A sales manager skimming the CRM would move this account into "likely to renew" and think no more about it.

Now here are four moments from the call it summarizes, a thirty-minute conversation with the operations director of a property management company that pays for a monthly maintenance service:

  1. About eight minutes in: "We'd renew, but only if the on-site visits stay monthly. Your new contract says quarterly."
  2. About fourteen minutes in, the account manager: "Let me check with our operations team whether we can keep monthly at the current rate. I'll come back to you."
  3. About twenty-two minutes in: "Honestly, another vendor quoted us lower, but we'd rather stay if the visits work."
  4. About twenty-eight minutes in, almost as an aside: "I need something in writing by Friday. Our board meets Monday. And it'll be my colleague Marisol signing this year, not me."

Near the end, the customer repeats the first point: "So, to be clear, monthly visits, or we're out."

The summary isn't wrong about the call, exactly. It's just wrong about what the call was about.

What summaries tend to drop

Reading the two side by side, the missing pieces fall into a few kinds, and you'll see the same kinds in most summaries of real calls.

Conditions. "We'd renew, but only if" became "wants to renew." Summaries compress, and conditions are the first thing compression flattens. The word "but" carries a lot of meaning that rarely survives.

Commitments your own team made. The account manager promised to check with operations and come back. That promise is now nobody's task.

Numbers, dates and deadlines. "By Friday" and "the board meets Monday" are the most time-sensitive facts in the call, and they're gone.

Who said what, and who decides. The person signing has changed. That matters for who gets the quote.

Tone and hesitation. The competitor's lower quote, offered reluctantly, is the clearest signal of risk. Summaries tend to report sentiment as an overall verdict, "customer is happy," which smooths over exactly this kind of mixed message.

None of this means the tool is bad. A summary can be accurate sentence by sentence and still be wrong about the call. The question is what you ask it for, and what you check.

A summary can be accurate sentence by sentence and still be wrong about the call.

Read it line by line

If you rewrite that summary to capture what the call actually contained, it looks less like a paragraph and more like a set of fields:

  • Outcome: Renewal likely, conditional.
  • Condition: Monthly on-site visits must continue. New contract specifies quarterly. Said twice.
  • Our commitment: Account manager to confirm with operations whether monthly can continue at current rate.
  • Deadline: Written proposal by Friday; customer's board meets Monday.
  • Decision-maker: Marisol will sign this year.
  • Risk: Customer has a lower competing quote; prefers to stay.

Each line points to a moment in the call. Someone reading this knows what to do today, and who the quote goes to. The paragraph version would have produced a standard renewal quote, sent to the wrong person, describing quarterly visits, arriving after the board met.

Decision guide

What the summary said vs. what the call contained

ConsiderationParagraph summaryWhat the call contained
OutcomeWants to renew.Will renew only if visits stay monthly. Said twice.
Our promiseNot mentioned.Account manager to check with operations.
DeadlineNot mentioned.In writing by Friday; board meets Monday.
Who signsNot mentioned.A different colleague this year.
RiskCustomer is happy.Has a lower competing quote.
Illustrative composite call written for this note; not real tool output or client data.

Ask for structure, not prose

Most summarization tools can be set up, or built, to produce fields rather than a paragraph. For calls that feed a CRM or a project tracker, structure is almost always the better choice.

Useful fields for a customer call:

  • Commitments, with who, what and by when, for both sides.
  • Conditions and objections, quoted word for word rather than paraphrased.
  • Open questions that nobody answered on the call.
  • Numbers, dates and names, each with the moment it was said.
  • Next step, with an owner.

Two features matter more than the rest. Every field should link to the timestamp in the recording or transcript, so anyone can check it in seconds. And conditions should be quoted, not summarized, because the paraphrase is where meaning gets lost.

Decide who confirms what

A summary is a draft written by a tool that wasn't in the room. Treat it that way.

The person who was on the call should confirm the commitments and conditions before they become tasks or update a record. That takes a minute while the call is fresh, and it's much faster than writing notes from scratch, which is the real time saving. Anything that changes a customer record in a way that matters, such as a renewal date, a price, a contract term or a new decision-maker, should be confirmed by a person, not written automatically.

There's also the question of the recording itself. Everyone on a call should know a note-taker is present, and consent rules for recording vary by place, so check the ones that apply to you and your customers. We covered this, along with the rest of a workable staff policy, in your team already uses AI. Write the policy for that team.

How the work moves

From recording to record

  1. Call recorded with consent

    Everyone knows a note-taker is present.

  2. Structured draft

    Commitments, conditions, questions, figures, each linked to a timestamp.

  3. Rep confirms

    Corrects and approves while the call is fresh.

  4. Tasks created

    Each commitment gets an owner and a date.

  5. Record updated

    Changes to terms or contacts confirmed by a person.

Illustrative workflow, not a client screenshot or measured result.

Test it on calls you already understand

Before you rely on a summarization tool for anything important, test it on calls where you already know what mattered.

Pick ten recent calls that your team remembers well: a renewal that nearly slipped, a complaint that turned into an upsell, a call where something was promised. Run them through the tool. For each, have the person who was on the call mark what the summary got right, what it missed, and what it got wrong. Use your own categories. You're not looking for a score, you're looking for patterns. Maybe it reliably catches dates but flattens conditions. Maybe it's fine on short calls and poor on long ones.

Then decide where it fits: drafting notes for a person to confirm, populating certain fields automatically, or not yet. It's the same discipline we described for documents in the AI reads the invoice fine. Plan for the one it doesn't.

We build AI applications that bring context into the moment work happens, with sources visible and the important calls left to people. If your CRM is filling up with summaries nobody quite trusts, tell us how your calls flow today.

Related work: PatientIQ.

twimco.

We build alongside the people doing the work.
Bring us the problem you’re thinking about.

Related work

PatientIQ ↗︎