AI call scoring for CSRs: measuring booking rate from recordings

How does AI call scoring work?#

AI call scoring works as a chain: recordings are transcribed, split by speaker, classified by reason and outcome, and graded against a rubric your team defines. Most tools connect to the phone system for audio and to the field service platform, such as ServiceTitan or Housecall Pro, for the booking.

Every link in the chain can be wrong, and errors compound. A transcript that mishears a street name is harmless; a classifier that calls a billing question an opportunity quietly drags down a CSR's booking rate.

• 1Capture: the phone system records the call and passes audio plus metadata such as line, campaign and duration.
• 2Transcribe: speech becomes text, split into CSR and caller.
• 3Classify: the call is labeled by reason, such as new service, maintenance, existing job, billing or solicitation.
• 4Decide opportunity: the tool judges whether the caller could have been booked.
• 5Match outcome: the call is matched to an appointment or job created in the field service platform.
• 6Score: rubric items are graded, such as greeting, problem discovery, offering a time and asking for the booking.

Measuring booking rate correctly#

Booking rate is booked opportunity calls divided by all opportunity calls, and the definition of an opportunity decides whether the number means anything. Calls where a customer could have been scheduled count, even if the CSR never offered a time; calls about an existing appointment, a bill or a vendor pitch do not.

Verify the booked flag against the field service platform rather than trusting the transcript. A caller who says yes may hang up before a time is set, and a job created later by a CSR callback may belong to the original call. Write the rules down once and apply them to every CSR and every branch.

What does a scored call record contain?#

A scored call record contains the call's metadata, the transcript, the classification, the outcome and the grades, each traceable to the evidence behind it. The table lists the fields a complete record carries.

The last two rows are the ones most tools skip and the ones that give the record lasting value. A manager's correction shows where the AI was wrong, and the job outcome shows whether a well-scored call led to good work or to a canceled appointment.

Using scores for coaching without losing trust#

Call scores help coaching when CSRs can see the call, the rubric item and the passage behind each grade. A score with no evidence reads as surveillance and gets argued about rather than used.

Start with a short rubric and calibrate it by having managers score the same calls as the tool, then review the disagreements together. Each disagreement either fixes a rubric item or teaches the team something about a hard call.

Hold off on tying pay to raw AI scores until you know how often the tool misclassifies calls on your lines. Pay plans built on an unverified metric create disputes that outlast the tool.

Consent steps before recordings feed any AI#

Recordings should feed an AI tool only after you have checked that callers were told about recording, that your disclosure covers the intended use, and that the vendor's terms limit what it does with your audio. Call recording laws differ by state, and some require every party's consent, so counsel should confirm which apply to each of your lines.

Automated redaction helps but is not complete on its own. The open-source Presidio project, for example, warns that because it uses automated detection there is no guarantee it will find all sensitive information, and that additional protections should be employed. Plan for sampled human review of transcripts.

• 1Play a recording notice on every inbound line, including after-hours and overflow lines.
• 2Check outbound calls, where the notice is often missing.
• 3Tell employees in writing that calls are recorded and scored, and how scores are used.
• 4Read the scoring vendor's data use clause, and opt out of training on your recordings if you do not want it.
• 5Set retention periods for audio and transcripts, and delete on schedule.
• 6Keep payment card details out of recordings and transcripts, or pause recording while payment is taken.

Illustrative: a multi-trade contractor checks its booking rate#

Illustrative: a fictional HVAC, plumbing and electrical company runs a central call center for several branches. Its call scoring tool reports a booking rate the COO believes is too low, and CSRs have stopped trusting their scores.

A sample review shows the tool counting reschedules and billing calls as opportunities, and missing bookings made when a CSR called the customer back. The team tightens the opportunity rules, matches outcomes to jobs in the field service platform and adds manager review for disputed scores.

The decision: coach only on the corrected rate, and add the recording notice the after-hours line had been missing. The outcome is a booking rate CSRs accept as fair, and a scored call archive with linked job outcomes that the company owns and controls.

How SourceX looks at scored call archives#

Scored call archives with linked job outcomes are one of the record types AI developers ask about, because they connect real conversations to business results. SourceX assesses them through the SourceX five-step transaction: Supply, Rights, Preparation, Approval and Delivery.

The Rights step checks recording disclosures, employee notices and vendor terms. Preparation removes caller and employee identities, payment details and addresses under a plan the supplier approves, and the SourceX Evidence Packet records it. Nothing is shared at the fit check stage.