Real-Time Analytics and Agent Assist

Post-call analytics informs future behaviour. Real-time systems intervene during the call, which changes the engineering, the value and the ethics.

What real-time systems do

Live transcription, streamed with low latency.

Trigger detection. A phrase, a topic, a compliance condition.

Prompting. Displaying a suggested response, a required disclosure, a knowledge article.

Supervisor alerting. Flagging a call for intervention.

Automated notes and post-call summarisation, increasingly.

Where it genuinely helps

Compliance prompts. Reminding an agent of a required disclosure at the right moment, and confirming it was said. This is the clearest value and it reduces a real failure mode.

Knowledge surfacing. Bringing up the relevant article when a topic is detected, which addresses the long-silence problem directly.

New agent support. The period where an agent knows the systems but not the answers is where prompting helps most.

Escalation alerting to a supervisor, based on acoustic measures rather than emotion inference.

Post-call summarisation, which is the least glamorous and possibly the highest-value application, because it removes after-call work.

Where it fails

Latency. A prompt that arrives after the moment has passed is noise. Real-time transcription is less accurate than post-call, because less context is available.

Cognitive load. An agent reading prompts while listening to a customer is doing two things badly. Too many prompts make the agent worse.

False triggers. A prompt fired on a misrecognised phrase interrupts for no reason, and agents learn to ignore prompts.

Scripted-sounding agents. Heavy prompting produces reading rather than conversation, which customers detect.

Accuracy under real-time constraints, which is worse than the batch figures a vendor quotes.

Designing it so agents use it

Few prompts. Five well-chosen triggers beat fifty.

High precision over recall. A prompt that is usually right is used; one that is often wrong is dismissed reflexively, which also disables the ones that matter.

Unobtrusive placement. Not a modal, not a flashing element.

Dismissible, with the dismissal logged so you can see which prompts are consistently ignored.

Measured. Track prompt display, prompt use, and whether prompted behaviour actually occurred. Most deployments do not, and cannot say whether the feature does anything.

The surveillance question

Real-time monitoring is different in kind from post-call review, and agents experience it that way.

Continuous live monitoring of every call is a different working environment. It is not the same as knowing that a sample is reviewed later.

In several jurisdictions, employee monitoring carries specific obligations — notification, consultation with employee representatives, proportionality assessment. Real-time systems are more likely to engage them.

Agents should know what is monitored live, what triggers a supervisor alert, and what is recorded. A system that flags calls to a supervisor without the agent knowing produces a workplace where people are being watched by something they cannot see.

The transparency question has its own note, and it applies most sharply here.

The realistic assessment

Compliance prompting and knowledge surfacing are genuinely useful and have a clear mechanism.

Automated summarisation removes real work and is the application with the best return currently.

Live coaching prompts on soft skills are oversold. The evidence that they improve outcomes is thin, and the cognitive cost is real.

Emotion-triggered supervisor alerts inherit every problem in the emotion note and should use acoustic measures instead.

Buy it for the first two, be sceptical about the third, and refuse the fourth in its emotion-based form.

Measuring whether prompts do anything

Real-time features are deployed and rarely evaluated. The measurement is available.

Log every prompt displayed, with the trigger and the timestamp.

Log whether it was dismissed, ignored or acted on where the interface allows.

Check whether the prompted behaviour actually occurred in the transcript afterwards.

Compare outcomes on calls where a prompt fired and was acted on against calls where it fired and was not.

Track prompt precision. How often the trigger fired appropriately, sampled by hand.

Retire prompts that are consistently ignored. An ignored prompt trains agents to ignore all of them, including the compliance ones.

Most deployments cannot answer any of these, and consequently cannot say whether the real-time feature they are paying for changes anything.