Quality

Callable Intelligence

Callable Intelligence reads your agents' finished conversations, finds the problems that keep coming back, and suggests specific fixes. You test each fix against your current agent before anything goes live, and nothing changes on an agent until someone on your team approves it.

It's included with the Starter, Growth, Scale and Enterprise packs. On the Free pack or top-ups only, the page shows what it does and how to unlock it.

What it does

  1. Analyses finished conversations. Phone calls, SMS conversations and email threads, a few minutes after they end. Each one is scored, with what the customer wanted, whether they got it, and any problems: a question the agent couldn't answer, a booking it didn't offer, a customer asked to repeat themselves.
  2. Finds patterns. A single bad conversation isn't a pattern. A problem becomes an insight once it shows up in at least 3 conversations on the same agent (2 if it's serious), from more than one customer when your agent has several.
  3. Recommends a fix. For each insight it writes a recommendation: the problem, why it matters, the expected benefit, the risks, and a concrete change to the agent's instructions, knowledge or FAQs.
  4. Tests the change. Turn a recommendation into a proposal and run it in the Simulation Lab: your current agent and the proposed one go through the same scenarios side by side, in a sandbox.
  5. Goes live only when you approve. An admin approves the proposal and deploys it. Every deployment can be rolled back in one click.
  6. Measures the result. After a deployment, conversations before and after it are compared, so you can see whether things improved.

Turn it on

Callable Intelligence is off until you switch it on.

  1. Open Callable Intelligence from the sidebar, under the Co-pilot.
  2. Go to Settings and switch Callable Intelligence on.
  3. Choose the channels to analyse, and save.

When you first turn it on, it looks back over the last 90 days of conversations so it has history to learn from. After that it keeps up with new conversations as they finish. Only workspace owners and admins can change these settings.

Web chat conversations, Test Suites runs and meetings aren't analysed. Conversations shorter than the minimum duration (15 seconds by default) are skipped.

The tabs

TabWhat you'll find
OverviewConversations analysed, the quality score and its trend, open issues, recommendations and changes waiting to deploy
ConversationsEvery analysed conversation, with its outcome, score, findings and transcript (personal details removed)
InsightsRecurring problems, how many conversations and customers they came from, and the conversations behind each one
RecommendationsSuggested fixes, with the exact change and the reasons behind it. Create a proposal or reject a recommendation
OptimizationProposals: the current and proposed text side by side, simulation results, approval and deployment
Simulation LabPick a proposal and scenarios, see the cost, and run the side-by-side comparison
Version HistoryEvery proposal, approval, deployment and rollback, with before-and-after results
SettingsTurn it on or off, channels, sample rate, minimum duration, frequency, monthly budget and retention

Each agent's builder also has an Improve link that opens Callable Intelligence filtered to that agent.

How it decides what to fix

Recommendations come from evidence, not from a single conversation:

  • Repetition. At least 3 conversations on the same agent and topic, or 2 when the problem is serious.
  • Different customers. When an agent has several customers, the problem has to come from more than one of them, so one caller can't steer your agent.
  • Share of traffic. Very rare problems (under 2% of an agent's analysed conversations) don't become recommendations.
  • Outages aren't agent problems. Failures from an outside service, or tool errors bunched into a short window, are reported but never turned into agent changes.

Every Insight shows why it does or doesn't qualify yet, for example "Seen in 2 conversations from 1 customer; needs at least 3 conversations."

Safety checks on every change

A suggested change is checked in code before you see it:

  • It can only touch the agent's instructions, knowledge, FAQs, greeting and tool descriptions. Never phone numbers, integrations, keys or billing.
  • It can't add a web address, email address, phone number or price that isn't already in your agent's setup. If a fix needs one, the recommendation says so and leaves it for you to add.
  • It can't remove banned phrases or stop the agent handing callers to a person.

Anything that fails a check is shown on the recommendation as Stopped by safety checks, with the reason.

Who can do what

RoleCan
Owner, AdminEverything: change settings, approve, deploy and roll back
DeveloperCreate proposals and run simulations, but not approve or deploy
MemberView insights, recommendations and results. Transcripts are hidden

If your workspace has more than one admin, a change has to be approved by an admin other than the person who proposed it. Approving without a simulation needs a written reason.

Costs and budget

  • Each workspace has a monthly budget (default $25). When this month's analysis reaches it, analysis pauses and the page says why. It resumes next month, or straight away if you raise the budget.
  • To spend less, analyse only some channels, set a sample rate (for example 20% of conversations), raise the minimum duration, or analyse daily or weekly instead of continuously.
  • Simulations run through Test Suites and use credits like any test run. The Simulation Lab shows the cost and asks you to confirm before anything runs.
  • To stop completely, switch Callable Intelligence off in Settings. Nothing more is analysed, and everything found so far stays.

Privacy

  • Phone numbers, email addresses and other personal details are removed from what's stored and shown.
  • Customers are counted with a one-way code, never by their number or address.
  • Analyses are kept for 180 days by default (set anywhere from 7 to 730 days in Settings).
  • Workspaces whose data must stay in Saudi Arabia aren't analysed.

Measuring results

After a change is deployed, conversations from the 14 days before it are compared with conversations since. The verdict is improved, no clear change, regressed or not enough data yet. A verdict needs at least 30 conversations on each side.

The comparison shows how results changed after a deployment, not proof that the change caused it. Other things can change at the same time, such as who's calling and why. If a change is measured as worse, you can choose to be notified in Settings. Callable never rolls back on its own; rolling back is always your decision, from Version History.