TL;DR
- A general call score answers "how was the call"; a custom agent answers your specific questions: was X said, was Y verified, was Z promised.
- An agent is defined in plain language, no code: describe what to check, what counts as compliant and what to do when it is not, and it runs on 100% of calls.
- The common uses: industry compliance, enforcing promotion rules, spotting product-specific opportunities and internal quality standards.
- A good agent checks one clear thing. Ten focused agents beat one agent trying to check everything.
A good conversation intelligence system ships with the horizontal capabilities: transcription, scoring, summaries. But in every business, after a month, the same question comes up: "can it also check that...?" And the rest of that sentence is always something nobody else has: the chain's refund wording, the mandatory brokerage question, the promotion that only runs until month end. Custom AI agents are the systemic answer: instead of requesting a feature from the vendor, you build the check yourself.
What an agent actually is
An agent is a checking instruction that runs on every call after it ends (or during it, when wired to the AI copilot). You define it in three sentences: what to check, when it applies, and what to do with the result. For example: "On every new-policy sales call, verify the rep presented the exact management fees. If not, tag the call and send an alert to the supervisor". That is it. From that moment, no policy call passes unchecked.
The four uses that repeat on every floor
- Industry compliance: fair disclosure, mandatory declarations, binding wordings. Health declaration control is a live example of such an agent in insurance.
- Rotating business rules: promotion terms, current prices, stock. An agent updated with one sentence when the promotion changes, instead of a briefing everyone forgets.
- Opportunity spotting: a customer mentioning a need that fits a complementary product, a move, a life event. The agent tags it, and the next sale is half made.
- Internal quality: your standards beyond the script, value before price, a meeting offered on every fitting call, a respectful tone even in rejection.
What separates a good agent from noise
Rule one: one check per agent. An agent checking "that the call was good and proper and professional" checks nothing; an agent checking "was the VAT-inclusive price presented" gives a sharp, actionable answer. Rule two: every agent has a recipient and a response, otherwise it produces a report nobody reads. Rule three: calibrate on real calls. Run the agent over fifty past calls, see where it was right and where it missed, and sharpen the wording before going live.
The real power: questions you did not know to ask
Once the basic agents run, the interesting stage begins: research agents. "On closed deals, what did the rep do after the price objection?", "In which calls did the customer ask for time to think, and what were they promised?". Every business question becomes a check running across all calls, feeding cross-call analysis and coaching. The floor turns from a place that collects recordings into a place that asks them questions.
How it is built in Saleso
You build an agent in a plain-language interface: describe the check the way you would explain it to a new employee, choose which calls it runs on and what happens with the result, a tag, a score, an alert or a CRM field. Start from templates for your industry and sharpen. And when something truly complex is needed, the implementation team builds it with you. The definition stays alive: update, pause or duplicate an agent at any moment.
Frequently asked questions
How many agents should run in parallel?
Start with two or three that solve a clear pain, usually compliance or a business rule that breaks often. A mature floor runs ten to twenty active agents comfortably, because each is focused and quiet: it only speaks when there is a finding.
What happens when an agent is wrong?
Mark the finding as incorrect with a click, and it becomes a calibration example that sharpens the agent. Keep the direction in mind: an agent catching ninety percent of cases already beats any manual review, which samples five percent of calls at best.
Can an agent also act in real time, inside the call?
Yes: an agent wired to the AI copilot can surface a reminder to the rep at the relevant moment, for example when a stage requiring a mandatory sentence arrives. The same definition also powers the post-call check, so control and assistance live from one source.