Agent assist

A new employee on the phone makes mistakes not out of laziness but because they do not remember: delivery times, return conditions, what is included in the set, how late bookings are taken. Agent assist is a screen where answer cards from your knowledge base appear by themselves during the conversation: the guest asked — the hint is already there.

Open: Admin panel → Call assist (admin.cenaly.com/call-assist). The section is switched on by the Operator coach card in the app store — it is hidden by default.

What agent assist does#

The assist listens to the conversation, recognises the speech and looks for a matching card on every guest reply. An important detail: no language model is called during the call. The answers are pre-compiled from the knowledge base into a "pack" with computed embeddings, so the hint arrives while the conversation is still going — the target time to the first hint is just over a second.

The section has four tabs: Call · Past calls · Coaching · Trainer; the "Agent assist knowledge bases" block sits on top, the settings at the bottom of the "Call" tab. The open tab lives in the address (?tab=…), so F5 and a link to a colleague both work.

What you need to start#

Step 1. Build the hint pack#

At the top of the section sits the Agent assist knowledge bases block: chips for every knowledge base collection ticked "Agent assist", plus "Whole knowledge base". Each chip shows its state: "pack ready", "building" or "no pack yet". The selected chip is the active base for the whole screen: the trainer, the manual search, the coaching and the "unanswered questions" all work against it.

A collection ticked "Agent assist" builds its pack by itself whenever the knowledge base is rebuilt — the Rebuild button is needed for the main base and for a forced rebuild. An empty list means no collection has the tick: the door leads to the knowledge base.

What the pack card shows:

• what it is made of — the "N from the knowledge base · M from fragments" badge: articles come first, document chunks fill up the remainder;
• staleness with a reason: "documents changed" or "articles changed";
• the golden report — the share of reference questions that landed on the expected card;
• access marks: a card built from an internal document is labelled "🔒 internal — do not read out" — the operator sees it but must not retell it to the guest; documents with a restricted-internal label are never compiled into a pack.

Step 2. Where the assist hears the call from#

There are two sources, and they do not exclude each other:

The settings live at the bottom of the "Call" tab: switch the assist on, choose the recognition provider (its per-minute price is shown right there), the language hint, whether to play the recording notice to the caller and which knowledge base new calls start with. The text and the language of that notice are set not here but in telephony — in the "default answer" block under the routing rules.

Step 3. What the operator sees during a call#

The Call tab is the working screen during the conversation:

• the live transcript coloured by speaker, the caller's number and the duration;
• hint cards: each with a confidence value, a "ask next" block and the sources it was built from. A shown card is not redrawn on every word, and only a noticeably better candidate can displace it;
• 👍 / 👎 feedback — it feeds the statistics and the "unanswered questions", instead of just disappearing;
• "Search the knowledge base" — a manual query into the same base when no hint appeared on its own (it is marked as a manual search);
• a base selector right inside the call: the operator switches the specialised base to fit the task ("delivery" → "booking" → "complaints"), and the guest's last reply is instantly rematched against the new pack. If the chosen base has no pack yet, it honestly says "pack not built" — hints are never silently substituted from the main base;
• "My extension" — a widget in the header showing the operator their own internal number and whether the softphone is online. Previously only a technician could answer "is my phone registered".

The panels next to the hints#

Early hints#

A separate toggle in the settings — "Show the topic and headings while the guest is still speaking" (off by default). With it, while the guest is finishing their sentence, a topic line and up to three heading-only cards marked "preliminary" appear in the hint area: a heading expands on click, and such a click does not count as using the hint. Once the guest has finished, the final match either confirms the card (the mark disappears and the card updates in place) or withdraws it.

The display is deliberately frugal: no more than three objects on screen, no more than one substantial update per 0.7 seconds, a shown card lives for at least a couple of seconds, and topics about returns, commitments and legal restrictions are only shown at very high confidence — a wrongly shown return condition costs more than a second of waiting. The verdict is computed locally on the PBX machine, so this mode does not add any model requests.

The trainer#

The Trainer tab is the same thing without a guest: the operator types a guest's line or dictates it in the browser and sees exactly the cards they would get on a live call. Below the confidence threshold the trainer honestly says "no hints" — which is a result too: it means there is a hole in the knowledge base. The trainer runs the active base (the chip selected at the top), its name is shown as a badge above the input. The call-stage script from the pack is visible there as well.

Past calls and unanswered questions#

The Past calls tab holds the sessions of the selected month: the full transcript, which hints were shown, the time to the first hint, and whether personal data was masked. The assist does not store audio — only text. A "conversation summary" and a list of follow-up tasks are computed on request: that is a separate AI expense.

The Unanswered questions block collects what guests ask while the knowledge base stays silent, ordered by frequency. AI writes a draft article from such a record, but saving it into the knowledge base is a human decision. Each gap remembers which base was silent, so the selector at the top filters the list by the chosen collection.

Coaching: the week card and objections#

The Coaching tab is visible to the account owner only and is computed without AI, from data already collected:

• The week card — the team over the week: calls, answered and missed, the average quality score, verdicts, a comparison with the previous week; the checklist items that fail most often; the best calls of the week with a link into the Calls section; the assist's own metrics (how many sessions, in what share hints appeared, how fast); per-operator numbers — as a list, not a ranking. Where there was not enough data, it says so.
• The objections collection — the reasons calls were lost, the unanswered questions and the failed checklist items in one place. AI writes a draft objection card, and the "add to knowledge base" button puts the document into the active collection, not into a common pile.

Limitations#

• The assist only knows what got into the pack. The model does not reason or improvise during the call: what is not in the knowledge base will not be suggested, and the question goes into "unanswered questions". That is a deliberate trade: speed and predictability instead of improvisation.
• The pack has to be built. A collection ticked "Agent assist" builds it automatically when the base is rebuilt; the main base uses the button. Until the pack exists the chip says so honestly, and hints are not substituted from another base.
• Recognition providers depend on the cloud. On cenaly.ru only Yandex SpeechKit is available; in the AWS circuit OpenAI Whisper is always there, and SpeechKit appears if a key is configured. An unavailable provider is not dropped silently: it is removed from the list with a warning.
• Early hints are off by default and are computed on the dedicated PBX machine — an operator working through the bridge will not get that mode.
• Hint anticipation is currently only being measured. It has no toggle in the panel and shows the operator nothing: it is a comparison of how right it would have been.
• The bridge requires confirmation that the other party is notified about recording — without that checkbox assisting from a headset does not start. The assist cannot listen "from the browser", without the bridge and without your own PBX.
• Audio is not stored: the session keeps only text with personal data masked. While the assist is switched off, speech is neither recognised nor stored at all.
• Settings, pack building and the week card are owner-only; an operator gets the feed, the trainer and their own "My extension" line.
• Georgian is recognised poorly — on ka calls the hints will be noticeably worse than on Russian or English ones.

Troubleshooting#

Symptom-by-symptom triage lives in troubleshooting; if that does not help, write to us — how to get help.

FAQ#

Do I need my own PBX?#

No. If the telephony is someone else's or the calls go through a softphone, install the Cenaly bridge on the operator's computer — it listens to the headset and the speaker. Your own PBX simply saves you from installing anything on the PC.

Does the guest hear the hints?#

No. The assist is a screen for the employee: the guest only hears the human. The only thing a caller may hear is the recording notice, if you switched it on.

Is the conversation recording kept?#

Not by the assist — it stores text only, with personal data masked. The audio lives where it was recorded: on your own PBX and in the Calls section, under its access rules and retention.

Can I keep several different bases?#

Yes, that is the point: a knowledge base collection ticked "Agent assist" becomes its own pack, and the operator switches bases during the conversation. The trainer, the coaching and the "unanswered questions" are computed against the selected one.

Why did no hint appear even though the answer is in the documents?#

Either the pack was built before the document appeared (check the staleness), or the document made it into the pack as a low-scoring fragment. Test the wording in the trainer: it shows exactly what a live call would show.

How much does it cost?#

The hint search itself makes no extra model requests. What costs money is speech recognition (the provider's per-minute price is shown in the settings) and the optional "conversation summary" and draft articles. The remaining AI balance is visible in your subscription.

Does it work for a face-to-face conversation rather than a phone call?#

Yes, through the bridge: it listens to the computer's headset and speaker, so hints work at the counter too — provided the other person has been notified about recording.

Related articles#

• Telephony: your own PBX — lines, operator extensions and the recording notice
• Calls: recordings and speech analytics — where recordings live and how call quality is scored
• AI receptionist — the robot on inbound calls that hands over to a human
• Knowledge base — collections of documents and the "Agent assist" tick
• Cenaly bridge — assisting from a headset with any softphone
• SIP phone setup — why the operator's softphone is offline
• Staff and shifts — who can open the section and whose numbers appear in coaching