Scoring and reviewing recorded calls with AI

How the bot works

Nine steps. The transcription, the masking, the criterion checks and the audit trail are ordinary code and cost almost nothing to run; the model is only used to read a conversation and to say whether a line is in it.

1. It works from the calls you already record. Under your own notice and consent rules, from your phone system or your platform. It starts no recording, enables nothing, and a call that was not recorded under your policy is not scored — there is no configuration in which it is.

2. It masks before it reads. Card numbers, bank detail, health information, anything on your list — stripped out of the transcript before a model is called, not after. The unmasked audio stays where it already lives, under the access rules it already has.

3. It transcribes and works out who said what. Both sides, labeled, with timestamps, tied to the call record and the account where your system already knows it. The transcript is the evidence for everything that follows.

4. It scores only what is checkable. Was the required disclosure said. Was the callback number confirmed. Was the address read back. Was the next step agreed out loud. Was the hold explained. Each one answered yes or no, with the line it came from and the second it was said.

5. It refuses to grade the judgment calls. Tone, empathy, rapport, whether somebody "owned it" — flagged with the passage and handed to a person, never scored. A number on empathy is a number somebody will defend in a review, and it should not exist.

6. It reports the team, not a league table. What customers keep asking, the objection nobody has a good answer for, the step that gets skipped when the queue is long, the disclosure that goes missing on Fridays. Across the team by default; a per-person view exists only if you turn it on for named people.

7. It points at the minutes worth hearing. Not a report to read instead of listening — the specific passages, with their timestamps, that a manager should play. Coaching is somebody listening to a call and talking to a person, and the bot's job is to find the minutes worth playing.

8. A manager owns the score. Every criterion can be overridden in one pass, and an override is recorded with its reason. Those overrides are the calibration set: when the bot and your managers disagree often on a criterion, the criterion is badly written, and that is a finding rather than a failure.

9. It keeps the evidence for every point. The rubric and its version, each criterion with the quoted line and timestamp behind it, what was flagged rather than scored, who overrode what and why, and what was masked. A score nobody can trace is a score nobody should act on.

What you configure

It can be configured to run in any business. This is what that actually means — the list of things that are yours, not ours.

Your scorecard

The criteria, in your words, written to be checkable — and the weights. Shorter is better: a rubric nobody can hold in their head gets argued with instead of used.

Your required lines

The disclosures, confirmations and read-backs that have to be said out loud on a call, and which call types each one applies to.

Your checkable and judgment split

Which criteria are scored from the transcript and which are only ever flagged for a person. Ours starts conservative and the line is yours to move.

Your recording rules

Which calls are recorded, in which places, what your greeting says, who may listen, and what happens when somebody objects. These come from you and your counsel; the bot only reads what your policy already covers.

Your masking list

What is stripped before any model sees a transcript — card and bank detail, health information, identifiers, anything else you name — and what is kept.

Your retention

How long transcripts and scores live, what is deleted when a recording is deleted, and who can ask for either.

Your coverage

Every call, or a share by queue, type or hour. Every call is possible; it is not always the right answer, and it is not a slogan.

Your visibility

Who sees scores, whether per-person views exist at all, who may hear a recording, and what the team is told about how this is used.

Your calibration loop

How often managers re-score a sample by hand, how disagreements change the rubric, and who signs off on a rubric version.

Transparent, and configurable by you

The dark band below is what the system will not do. This is what you can see and change — and it is how we build everything, not something invented for this page.

You see what it did, and why

Every point names the line that earned it and the second it was said, every flagged criterion names the passage instead of grading it, and every override is recorded with its reason. A score can be checked against the call in under a minute.

Versioned prompts you can tune yourself

The rubric, the required lines and the prompts behind reading a transcript are yours to tune. A change is a reviewable diff, replayed against calls you have already scored by hand so you can see exactly what it would have judged differently, and rolled back the same way.

A human in the loop wherever judgment actually matters

Transcribing, masking and checking the checkable criteria are automatic. Judging tone, deciding what a score means, choosing who gets coached and anything touching somebody's job are a person's — that boundary is a property of the system, not a promise about how carefully we'll use it.

Every run logged, along with what it cost

One row per call and one per rubric change: the rubric version, each criterion with its evidence, what was masked, what was flagged, who overrode what and why — plus token usage and cost for that run.

It can be configured to run in any business. The rubric and the rules are yours: your criteria, your required lines, your recording policy, your masking list, your retention, your visibility. The call wiring on this page is one configuration of it, not the limit of it.

Works with your tools

This runs on the phone system and the recordings you already have. We're not replacing your phones and we're not asking you to record anything new.

• A VoIP phone system
• A contact center platform
• A field service or dispatch tool
• Salesforce
• HubSpot
• Zendesk, Freshdesk or Help Scout
• A folder of recordings you control
• A spreadsheet you control
• Email, Slack or Microsoft Teams
• Your own transcripts, if you already have them

Works well for

• Contact centers — Quality review covers a sample, and the calls outside it are the ones nobody can say anything about.
• Insurance agencies — Required lines have to be said out loud on every call, and checking that is somebody's afternoon.
• Home service companies — Every booked job starts on the phone, and the script lives in whoever answered it.
• Healthcare providers — Front-desk calls carry patient detail, so masking and who may listen get settled first.
• Law firms — Intake calls are the one chance to get the facts and the next step right.

How a rollout works

Nothing is scored for real on day one. The first milestone is a set of calls your managers have already scored by hand, scored again from scratch and compared line by line.

1. 01Discovery

   The quote call, then a working session. We go through your scorecard, the lines that have to be said, which criteria are checkable and which are judgment, your recording and notice rules, your masking list, retention, and who may see what. You get the rubric written down, the exclusions and a written scope.

2. 02Dry run

   The bot scores calls your managers have already scored, with nothing written back. We compare criterion by criterion, and every disagreement is either a criterion that is badly written or a judgment call that should never have been scored. This is where the checkable and judgment split gets settled, and where your own numbers come from.

3. 03Assisted mode

   Live on one queue or one call type, with scores visible to managers only and every criterion overridable. Quality review stops being a sampling exercise and becomes a reading exercise. Who coaches whom does not change.

4. 04Optional limited automation

   Optional, and narrower than it sounds. The checkable criteria scored across every call in the queues you name, with the judgment criteria still flagged for a person, every hard stop in force, a sampled share re-scored by hand anyway, and the kill switch. Plenty of teams stop at step 03, and that's a good place to stop.

What we don't claim

• No numbers we haven't earned. No accuracy figure, no agreement rate with your managers, no coaching outcome, no conversion claim. We haven't listened to your calls. Whatever agreement looks like comes out of the dry run against calls you already scored by hand, and it would be yours.
• No customer claims on this page. No names, no case studies, no logos.
• This is not a monitoring product. It is built to make the checkable part of quality review cheap so that managers spend their time listening and talking. If what you want is surveillance of individuals, the guard band is not a phase one — the ranking, the per-person defaults and the employment decisions are off by design, and we are not the right people for that build.
• It will disagree with your managers, and sometimes it will be right. Those disagreements are the output you should care about in month one. The useful ones are criteria that were never written clearly enough to check, which is a rubric problem rather than a model problem.
• A transcript is not a call. A hard accent, a bad line, two people talking at once and a word the model has never heard all cost you evidence. Where the transcript is too poor to quote from, the call comes through unscored rather than scored badly.

Questions people ask

Can AI score our sales and service calls?

Yes, for the part that can be checked. An agent can read the transcript of a call you already record, answer the yes-or-no criteria on your own scorecard - was the disclosure said, was the number confirmed, was the next step agreed - and quote the line and the second behind every answer. Tone, empathy and rapport it flags for a person instead of scoring. It ranks nobody, and a manager owns the result.

What will it refuse to score?

Anything that is a judgment rather than a fact. Whether somebody sounded warm, whether they built rapport, whether they "owned" the problem: the bot finds the passage and hands it to a person. The reason is simple - a number on empathy looks settled, gets quoted in a review, and nobody can check it. The checkable criteria are checkable precisely because the words either were or were not said.

Do we have to tell our team their calls are being scored?

What you must do is a question for you and your counsel, and the bot claims no knowledge of recording or employment law - it reads your rules and follows them. What we will say is that we would rather you tell the team plainly: what is scored, what is only flagged, who sees it, and what it is never used for. Configuration includes who may see a score, because that question should be answered before anything runs.

Will it rank our people?

No. There is no leaderboard and no stack ranking, and per-person views exist only if you deliberately turn one on for named people. The default report is about the team and the process: the question customers keep asking, the step that gets skipped when the queue is long, the line that goes missing on Fridays.

Can it decide who gets coached, or fired?

No, and that is a hard stop rather than a setting. It produces evidence - a score with the quoted line behind it, and the passages worth listening to. Coaching is a manager listening to a call and talking to a person. Pay, discipline, promotion and termination are outside what this system is for, and using a machine score as the reason for one is outside what we will help you build.

What happens to card numbers and health information on calls?

They are masked out of the transcript before any model is called, from the list you define, and the unmasked audio stays where it already lives under the access rules it already has. If something on your list cannot be masked reliably in a given call, that call is not scored.

How accurate is it?

We do not have a number for your calls, and anybody who quotes you one has measured somebody else's. What we do instead is the calibration step: your managers score a set by hand, the bot scores the same set, and you look at the disagreements criterion by criterion before anything goes live. That is also how badly written criteria get found, and finding them is the point of the step.

Will it work with our phone system?

If you can get the recording or the transcript out of it, almost certainly - by API, by a scheduled export, or from the folder your system already drops files into. A VoIP system, a contact center platform, a dispatch tool that records, or a folder of audio with a spreadsheet beside it are all normal starting points.

Is there a record of why a call scored the way it did?

Yes, per call and per criterion: the rubric and its version, the quoted line and timestamp for every point, what was flagged rather than scored, what was masked, and every override with who made it and why. If a score is ever questioned months later, the answer is the call itself rather than a reconstruction.