Coaching

Coaching collects the changes Tandem recommends to the resources your agents use, in one list you work through. Each recommendation has two answers: Accept, which applies the change, and Reject, which clears it from the list.

The page lives under Automations → Agents → Coaching.

Suggested FAQs

The first kind of recommendation is suggested FAQs. These are questions people actually asked that one of your knowledge bases couldn't answer. Both channels feed it, phone calls and web chat alike, for every agent a knowledge base is attached to. A question has to come up a few times before it's listed, so a single unusual phrasing doesn't clutter the page.

Every row names the knowledge base the question was asked against, how many times it was asked, and when it was last seen. Recommendations are ordered by how often the question came up, so the ones affecting the most conversations are at the top.

Each one also carries a label:

• No matching content — nothing in the knowledge base covers this yet. Write an answer.
• Near match — existing content may need tuning — the answer is probably in there already, but retrieval isn't finding it. Check your existing content before adding a duplicate.

Also asked as expands a row to show the other wordings people used for the same question, along with links to the conversations it came from. It's worth a look before writing the answer: the phrasings tell you the words people actually use. Some rows have no conversation links, and that's normal. The wordings and the count stand on their own.

Accept

Click Accept, fill in a Title and an Answer, then click Add FAQ. Tandem can't prefill the answer, because a suggested FAQ is by definition a question your knowledge base has no answer for. The title is prefilled with the question itself, and you can change it.

This creates a typed FAQ in that knowledge base, exactly as Write an FAQ does on the knowledge base's own page, and the recommendation clears itself from the list.

Reject

A suggestion that isn't worth an FAQ can be cleared with Reject. There's no confirmation dialog: the toast that follows carries an Undo if you change your mind.

Agent changes

Tandem also proposes specific configuration changes to your agents, built on the calls a supervisor coached or took over and, if you turn on automatic reviews, on all of an agent's calls. Each is backed by evidence from the calls it read. This is a separate kind of recommendation from suggested FAQs above. The Coaching page has two tabs: Agent Changes and Suggested FAQs.

Only voice calls are reviewed. A website chat is never read by a review pass and never produces one of these recommendations. An agent that only handles chat won't see anything here, however many conversations it has.

The agent list

Agent Changes lists your agents down the left side, one card each, showing how many recommendations are waiting for that agent and how long ago the latest one was made. Agents with the most waiting are at the top. Once you have five or more agents, a search box above the list narrows it by name.

Click an agent to show it beside the list. The agent at the top is shown until you pick one, and each agent has its own address, so a link to one agent's coaching opens on that agent. In a narrow window, the list and the agent take turns: clicking an agent opens it on its own, and All Agents goes back to the list.

An agent's summary gives its counts (Pending, Applied, Dismissed), when its latest review ran (Last Review), and how many calls that review read (Calls Reviewed). Open Agent goes to the agent's editor, and Automatic reviews sets whether the agent's calls are reviewed automatically (see below). Below the summary, the Pending tab lists what is waiting on a decision and the Decided tab lists what was already applied or dismissed. An agent with nothing pending opens on Decided, so its history is still one click from the list.

An agent that doesn't store transcripts says Not reviewed — transcripts are off on its card, because a review has no calls to read. Its summary says how to turn transcripts on.

What triggers a review

A call a supervisor coached the agent on starts a review as soon as the call ends. The review reads the supervisor's coaching notes beside the call and turns them into changes. Coaching on a web chat doesn't start one, because reviews read calls only.

A call a supervisor took over from the agent also starts a review, once the call has ended and the supervisor has closed it out, so the review reads the whole call: why they stepped in, how they closed it out, and what they said after the agent left. A review that runs before then stops short of that call and reads it next time. A call nobody closes out within 12 hours is read by the next review without a close-out.

Unless you turn on automatic reviews, those are the only calls a review reads. Every recommendation then comes from something a supervisor said or did, and says who: Coached by and their name.

Where agent tests are enabled, a review also suggests up to 2 tests built from the calls it reviewed, and applying a recommendation suggests one test that checks the change holds. They appear under Suggested Tests below the agent's recommendations. See Suggested tests.

Automatic reviews

Automatic reviews, in the agent's summary, has the AI review all of the agent's calls too, every 10, 25, or 50 stored calls, as you choose. It's off for every agent until you turn it on. A review reads at most 50 calls and proposes at most 5 changes, which is why 50 is the longest interval. Calls a supervisor coached still count for more than the AI's own reading, and a recommendation that cites one still says who coached it.

While automatic reviews are on, Review Now beside the setting runs a review immediately rather than waiting for the count. Reload the page once it finishes to see what it recommends. A review needs stored transcripts either way: see Transcript retention for what controls that.

What a recommendation contains

Each card says where it came from and when. One built on calls a supervisor coached or took over says Coached by and their name. One from an automatic review says AI recommended and "found in X of Y calls reviewed": Y is how many calls that review read, and X is how many of them the recommendation cites as evidence. Underneath, in plain English, is the reason for the change.

A recommendation can only propose one of nine settings: the system prompt, the greeting, dispositions, tool-usage policies (steering rules), the knowledge relevance floor (Minimum score in the editor), knowledge passages per lookup (Max results), the idle re-prompt, how long to wait in silence before re-prompting (Silence before the agent speaks up), or the maximum call length (Maximum call duration). Nothing else: no workflow change, no routing, no knowledge base content.

What Will Change shows the setting's current value against the proposed one. On the Decided tab it's What Changed for an applied recommendation and What Was Proposed for a dismissed one, and shows the value as it was when recommended against the proposed one. Evidence links to the specific conversations the recommendation drew on, each with the quote that grounded it. In a very narrow window, a pending recommendation hides both, along with Apply and Dismiss; a decided one keeps them. Those links go to Call Logs, which defaults to the last 30 days and, on an organization without that page enabled, doesn't open at all — an old evidence link can look broken when it isn't.

Apply, Apply all, and Dismiss

Apply writes the proposed value. What happens next depends on whether the agent has been moved to versioning: on one that has, the write lands as a new published version, so a change you don't want can be undone by publishing the version it replaced; on one that hasn't, Apply overwrites the live configuration directly, and there is no earlier version to go back to. The Versioning page's own callout says which kind of agent you're looking at. Nothing here is ever applied without a click.

Apply asks you to confirm first. The confirmation shows the change itself (the same view as What Will Change) and says whether the agent can be rolled back: on a versioned agent, by publishing the version the change replaces; on one that isn't, there is no earlier version.

Whenever an agent has a pending recommendation that can be applied, Apply All, in the agent's summary, applies those recommendations in a single write. It leaves out any that are outdated, refreshing, or no longer needed (see below), which stay pending for you to deal with on their own cards. If two of the rest propose different values for the same setting, only the newer proposal is sent; the older one stays pending rather than being applied out of order. Its confirmation lists exactly the changes it will send and says how many others will stay pending. Once the newer one is applied, the older one is marked Outdated, and you can refresh it.

Dismiss clears a recommendation without changing anything, and that recommendation won't come back. Because it can't be undone, Dismiss also asks you to confirm. A later review can still raise the same change again under a different title.

When a recommendation is outdated

A recommendation is compared against the agent's configuration as it is right now, not the snapshot the review read. When that setting has changed since the recommendation was written, most often because another recommendation for the same setting was applied, the card is labelled Outdated as soon as the change lands. It shows no diff, rather than one measured against a value that no longer exists, and it can't be applied. Apply marks a card Outdated too if it finds the agent changed in a way the card couldn't see.

Refresh, beside the label, rewrites the recommendation against the current value. It keeps what the setting does now, including the change you just applied, and adds what this recommendation was for, from the same evidence. The card reads Refreshing… while that runs, usually well under a minute, then shows the revised change for you to apply or dismiss as usual. If the current value already does what it asked for, the card says No longer needed, and you can dismiss it. If the refresh fails, the card says so, and you can refresh it again.

Transferred calls

When a call is transferred from one agent to another, the stored transcript covers only the leg before the transfer. Nothing said after the handoff is recorded. A recommendation built from a transferred call rests on that first leg alone, and the card gives no indication that the call carried on elsewhere. Check the linked conversation in Call Logs before applying a recommendation you'd expect to depend on how the call actually ended.

The call record itself stays filed under the agent the caller originally reached, not the one they were transferred to. An agent that only ever receives transferred calls accrues no call records of its own, so it never crosses the review threshold and is never reviewed.

Next steps

• Knowledge base overview — build the knowledge base the accepted FAQs land in, and attach it to an agent.
• Transcript retention — turn on transcript storage for an agent so its calls can be reviewed.
• Versioning — how a published version works, and how to roll one back.