Check AI call summaries against the call before saving
Compares each AI-written call summary with the transcript: anything invented, a complaint, support need or promise left out, a wrong outcome.
Try it on this example
Call summary written by the AI model: Call summary Reason for call: Customer called about her home insurance renewal premium, which has increased from £431 to £612. Discussion: Agent explained the increase is due to higher rebuild costs and storm claims in the area. Customer expressed dissatisfaction with the increase as a long-standing customer with no claims. Agent suggested removing accidental damage cover to reduce the premium. Customer advised her husband has passed away and is still named on the policy; agent advised a copy of the death certificate is needed to update the policyholder. Outcome: Customer agreed to renew at £612. Agent has asked underwriting to review the premium. Next steps: Agent to call customer back by Thursday afternoon with a requote. No further action.
Call transcript with speaker labels
- Can the summary be checked from the transcript alone?Yes97%
- How well does the transcript support the statements in the summary?Contradicts the call100%
- Does the summary leave out a complaint the customer made on the call?Yes68%
- Does the summary leave out circumstances the customer mentioned that may mean they need extra support?Yes94%
- Does the summary leave out something the agent promised to do, or the time given for it?Yes96%
- Does the summary record how the call ended as the transcript shows it?Different outcome100%
- Does the summary attribute something said or decided on the call to the wrong person?Yes84%
- How far can the next person rely on this summary as the record of the call?Wrong in a way that matters100%
These are real answers stored from one run on this example.
The prism behind it
Check AI call summaries against the call before saving
Fields
- Call transcript with speaker labels
- Call summary written by the AI model
Context
These are summaries of phone calls with customers of Pellbrook Insurance, a UK home and car insurer. After each call, the contact centre's software writes a summary with a generative AI model, and the summary is saved to the customer's record. The next agent, the complaints team and auditors rely on it as the record of the call. This check runs before a summary is saved. It compares the summary with the call transcript; it does not write or correct a summary. A summary with a problem goes back to the agent who took the call, who corrects it. What every summary must record, when it happened on the call: - Why the customer called. - Any complaint. Our complaints policy counts any expression of dissatisfaction with our service, prices, decisions or staff as a complaint, whether or not the word complaint is used. - Any circumstance the customer mentions that may mean they need extra support, such as a bereavement, an illness or disability, money worries, or difficulty reading, hearing or using the phone, and any support they ask for. - Every commitment the agent makes, such as a call back, a letter, email or form to send, a referral to another team or a change to the policy, with any day or time the agent gives. - How the call ended: what was agreed, changed or left open, and what happens next. Rules for the summary: it states only what was said on the call; it attributes each statement, request and decision to the person who made it; it may shorten what was said but must not change it. The transcript comes from speech recognition, with the speaker labels Agent and Customer, and may contain small recognition errors. Code also checks each amount, date and policy number in the summary against the transcript and the policy system.
Questions
Can the summary be checked from the transcript alone? Yes / No
Read the transcript and the summary. Yes: The transcript covers the call from greeting to goodbye and is readable enough to see what was said, even with small recognition errors. No: The transcript is cut off, mostly unreadable, or plainly about a different call from the one the summary describes, so the summary cannot be checked.
How well does the transcript support the statements in the summary? Scale
Check each statement in the summary about why the customer called, what was said, what was agreed and what happens next against the transcript. Something left out of the summary is not judged here; the questions on omissions record it. Rate the worst statement, not the number of problems.
Does the summary leave out a complaint the customer made on the call? Yes / No
Read the customer's lines and the complaints rule in the context, then the summary. Yes: The customer expresses dissatisfaction with our service, prices, decisions or staff, and the summary does not record it as a complaint or as dissatisfaction. No: The summary records each complaint, even in a few words, or the customer expresses no dissatisfaction with us. Frustration with an event itself, such as a storm, with nothing aimed at us, is not a complaint.
Does the summary leave out circumstances the customer mentioned that may mean they need extra support? Yes / No
Read the customer's lines and the support list in the context, then the summary. Go only on what the customer says, never on accent, fluency or tone of voice. Yes: The customer mentions a circumstance on the support list, or asks for support such as large print or contact in writing, and the summary leaves out that circumstance or that request. No: The summary records each such circumstance and each request for support, or the customer mentions none.
Does the summary leave out something the agent promised to do, or the time given for it? Yes / No
Read the agent's lines, then the summary. Yes: The agent commits to an action, such as a call back, sending a letter, document or form, a referral, or a change to the policy or the record, and the summary leaves out that action or the day or time the agent gave for it. No: The summary records every commitment with any time given, or the agent makes none. Explaining how something usually works is not a commitment.
Does the summary record how the call ended as the transcript shows it? Choice
Read the end of the call: what the customer agreed, decided or left open, and what happens next. Compare it with the outcome the summary gives.
Does the summary attribute something said or decided on the call to the wrong person? Yes / No
Compare who says, asks for, proposes and agrees to each thing in the transcript with who the summary says did. Yes: The summary says the customer said, asked for or agreed to something the agent said or proposed, or the other way round. No: Every statement, request and decision in the summary is attributed to the person who made it, or the summary attributes nothing to anyone. A statement that was never made at all is judged under grounding, not here.
How far can the next person rely on this summary as the record of the call? Choice
Judge the summary as a whole against the transcript and the rules in the context. The answer is a suggestion; the agent who took the call corrects the summary. If more than one option fits, choose the one lowest in the list.
Lens columns
enough_to_judge, enough_to_judge_probability, grounding, grounding_average, complaint_omitted, complaint_omitted_probability, vulnerability_omitted, vulnerability_omitted_probability, commitment_omitted, commitment_omitted_probability, outcome_matches, outcome_matches_probability, wrong_speaker, wrong_speaker_probability, summary_fit, summary_fit_probability
Run it on your own text
Add this prism in the app, change any question, and test it on a file of your own.