Use tools on calls

An approved tool does nothing until a playbook uses it. You decide where: when it runs, what it sends, what the assistant may do with the answer, and what happens if it fails. Because this lives in the playbook, it is drafted, tested, published and rolled back like everything else.

First approve the tool under Tools. See Add tool servers.

Add a tool to the playbook#

Open the playbook and go to the Tools tab. Under Tools on the call, press Use another tool, pick an approved tool from the Tool list and press Add. A tool marked changes something is one its approver said changes another system.

Each tool gets a card. Its On switch keeps it in the playbook without running it. What it does is the name your team sees on the call record, for example "Check unit availability".

When it runs#

When it runs has two kinds of answer.

When the agent decides to look it up. The assistant chooses to use it during the conversation, the way it searches your knowledge. Tick the parts of the call it may use it in: Greeting, Asking questions, Answering questions, Booking or Wrapping up. Use this for a question whose answer lives elsewhere: "is a two-bedroom open in March?"

Automatically, at a fixed moment, every time. The assistant does not choose, and never sees the tool.

Automatically...	Timing
when the call starts	In the background. It never delays the greeting.
once the caller confirms who they are	In the background.
when an answer is recorded	You pick Which answer. In the background.
once all required answers are in	In the background.
before offering times	Before your booking rules decide, so a rule can use its answer. A holding line covers the wait.
after a meeting is booked	After the call, retried if it fails.
when the call ends	After the call record is written, retried if it fails.

Only when… adds a condition, built like a booking rule, so the tool is offered or run only when it holds, for example only once the care type is known.

A tool that runs during the call has 3 seconds to answer. If it is slower, it counts as a failure on this call.

What it sends#

What it sends lists every field the tool takes. For each one, Comes from says where the value comes from:

Choice	What is sent
Not sent	Nothing.
The agent fills it in	The assistant supplies it from the conversation. Only when the agent decides to use the tool. You write What the agent is told about it.
A fixed value	The same value every time.
Their answer: one of your questions	What the caller answered.
This location's CRM ID, name or time zone	From the community's profile.
The call ID, The number they called, When the call started	From the call.
The caller's phone (once known)	Only after the caller has given it or confirmed who they are.
Their CRM contact (once they confirm who they are)	Only after the caller has confirmed who they are.
Every answer so far	All the answers collected.
Your account ID	Only when you pick it. Never sent otherwise.
The booking, who it is with, when and what kind	Only for a tool that runs after a meeting is booked.

Values from the call are filled in by AnswerStack, never by the assistant, so a caller cannot talk it into looking up another community or another person's record. Whatever the assistant fills in is checked against the tool's own rules before anything is sent.

What happens with the answer#

Use the answer as decides what the tool's answer is for:

Named values, for the agent and your rules. Pick up to 10 values out of the answer. For each, give a Name for rules, say Where in the answer it is (a path such as $.available or $.units[0].date), its Kind (Text, Number or Yes or no) and a Label.
Text the agent can speak from. The answer is given to the assistant as text.
Nothing: just record that it ran. Only success or failure is kept.

Most characters the agent ever sees caps the answer (600 by default, between 50 and 2,000).

What the assistant may say#

Each named value has a switch: The agent may say this, or For rules only: never in the conversation. A lead score or an internal ID belongs in rules only. The assistant never sees a rules-only value, so it cannot say it.

The defaults follow who chose to run the tool. Values from a tool the assistant decided to use may be said. Values from an automatic tool are for rules only until you switch them on. Text from an automatic tool is recorded on the call but not given to the assistant, and nothing from a tool that runs after the call can be said to anyone.

The assistant presents an answer as current information from your system, in its own words, and does not add to it. Anything the answer does not cover gets "the team will confirm that". An answer that contains instructions for the assistant is thrown away whole, and the tool counts as failed.

Using values in flows and rules#

Named values appear in the rules builders you already use, under their label: booking rules, and branch and not-a-fit rules in questions and flows. A rule could transfer a caller when the lead tier is A, or send them down a different line of questioning.

A tool that runs once all required answers are in can also feed a line's talking points: values the agent may say appear under Lookup results in a Talk about point, and the assistant waits for the tool before it says the point. If the tool fails or is too slow, the point is left out. While a talking point uses a tool, the tool cannot be turned off or moved to another moment without changing the point first.

If the tool failed, its values are missing, like an unanswered question. When a booking rule reads a tool's value, the editor warns you until another rule says what happens when the tool could not answer.

If it fails or takes too long#

Choice	What happens
Say the team will confirm, and note it for follow-up	The default. A follow-up note is added to the call.
Carry on as if it were not set up	Automatic tools only.
Offer a callback	The assistant moves toward taking a callback.
Offer to connect them with a person	Hands over to the community's human backup.

Holding line is what the assistant says while a lookup takes a moment, for example "Let me check what's open."

The same applies when a tool cannot run at all: its server needs reconnecting, is paused, or changed the tool since it was approved. A server that fails three times on one call is skipped for the rest of it. The call itself never stops.

Tools that change something#

A tool approved as changing something (it opens a ticket, reserves a unit) follows stricter rules:

If the assistant uses it, The agent reads back what it will do, and goes ahead only when the caller agrees must be on. Publishing refuses it otherwise.
Most times on one call is Once or Twice. Never more.
It never runs for a caller AnswerStack knows nothing about, and it always carries a key that stops a retry from doing it twice.
A tool with no field for that key cannot run after a booking. Run it when the call ends instead.

When a tool needs more from the caller#

Some tools answer "I need one more detail", such as which floor plan. With When the tool needs one more detail, the agent may ask the caller on (the default for tools that only look things up), the assistant asks in its own words, one question at a time, and tries again with the answer. The caller can decline.

At most three details per request and two rounds per lookup.
Never passwords, codes, card or bank details, or ID numbers. A tool that asks for one fails, and the question is never spoken.
A tool that runs automatically waits for the assistant's next turn to ask. One that runs after the call has nobody to ask, so it fails instead.

Limits#

At most 4 tools the assistant can choose to use in any one part of the call. The editor warns at 4 and blocks more, because fewer choices keep the assistant reliable.
At most 20 tools per playbook.
A tool runs at most three times on one call (twice for one that changes something).

Communities#

A community's Customizations tab lists the playbook's tools like any other item. A community can Keep a tool or Turn off it, unless the playbook locks it. It cannot change what a tool sends: values such as This location's CRM ID already follow the community. See Create your communities.

Where facts come from#

The assistant states facts only from your brief, the community's profile, your knowledge, and the answer of an approved tool on this call. A price or a date that none of those contains is caught before it is spoken. See What it will not do.

Before you publish#

Publishing checks every tool: that it is approved and unchanged since, that its server is connected, and that everything it sends is something the tool takes. Use Try it under Tools to see what a tool answers before callers hear it. See Test and publish.