Template · Agent · Official
GmailSlackDiscordNotionHubSpot

Customer support that answers from your own docs

Agents resolve the questions your docs already answer and hand the rest to your team with full context — history, account state, and a suggested reply.

Get this template

Opens in AGNT and takes about a minute.

The short version

Tools it needsGmail, Slack, Discord, Notion, HubSpot
Setup4 steps, about ten minutes
ApprovalYours, per category — nothing is sent on your behalf unless you say so
Where it runsYour own machine. Credentials stay in a local vault.

What the agent actually does

The documented questions stop reaching you

Responses are grounded in your actual documentation and previous resolutions, with the source attached. If the docs do not cover it, the agent says so and escalates rather than improvising an answer that sounds plausible and is wrong.

Human gates where you want them

You decide which categories may be answered unattended and which always reach a person. Billing, cancellations and anything touching an account state are typical for the second list, and the gate is enforced by the runtime rather than by asking the model nicely.

Every resolution logged

Each ticket leaves a receipt: what the customer asked, which documentation was used, what was sent and who approved it. That record is what makes it safe to widen the agent’s remit over time.

The tools this job needs

Gmail

Where the work arrives. The inbox most operations still start in. The agent watches it and reads what turns up in full, rather than matching a rule against a subject line.

Slack

Context the agent pulls in before deciding. Team chat where operations actually surface. It is read, not just referenced — which is what lets the decision account for it.

Discord

Context the agent pulls in before deciding. Community chat with an automation-friendly API. It is read, not just referenced — which is what lets the decision account for it.

Notion

Where the result lands. The team wiki agents can actually keep current. Nothing is written here until the agent has formed a view and, where you asked for it, you have approved it.

Set it up

  1. Point the agent at your documentation and your resolved-ticket history.
  2. Define which categories can be answered directly and which must always escalate.
  3. Connect the channel — email, Slack, Discord, or your helpdesk — and the CRM for account context.
  4. Run in draft-only mode until the receipts justify letting a category answer unattended.

You can see exactly what it did

Nothing happens behind your back.

Every run leaves a receipt: what the agent read, which tools it called, what it decided and why, and precisely what it changed. Anything irreversible — sending, paying, publishing, deleting — waits for you to approve it. It runs on your own machine, with your own credentials, and the whole trail is yours to read afterwards. The point is not that you trust it. The point is that you never have to.

The brief it works from

This is the actual instruction set the template installs — what the agent is told to do, and what it is told never to do. Every line of it is yours to edit in AGNT after install.

Read the full brief

You handle the first pass on customer support, and you are measured on time-to-correct-answer, not on how many humans you avoided.

Answer only from the documentation

Retrieve the relevant material and answer from it, with the source attached. If the documentation does not cover the question, say exactly that and escalate. Never answer a policy, pricing, billing or account question from general knowledge — you will eventually state a refund policy incorrectly with total confidence, and that is a worse outcome than every ticket you deflected.

"This is not documented, escalating to the team" is a correct and expected answer. Use it without hesitation.

What an escalation must contain

The human should not have to re-read the thread. Include what the customer asked, their account state, what has already been tried, which documentation you consulted, and a draft response to accept or rewrite. Handling an escalation should be a minute of judgment rather than ten minutes of reconstruction.

Categories that always escalate

Billing, refunds, cancellations, anything that changes an account, anything legal, and anything where the customer is clearly angry. That last one is not a technical judgment — it is a recognition that a frustrated person needs a human, and routing them to a better-worded automated reply makes it worse.

Tone

Match the team's existing voice. Be direct and brief. Do not apologise three times. Do not promise a timeline, a fix or a credit — those are commitments only a human can make.

Every question you could not answer is a documentation gap

Record it in the customer's own wording. Collected over a month that list is the best documentation backlog the team will ever have, because it is ranked by what people actually ask rather than by what someone assumed.

Shown exactly as it ships. The agent inherits whichever model you already use.

Deflection is the wrong target on its own

Optimising purely for tickets closed without a human produces a support experience everybody recognises and nobody likes: a system that works hard to avoid connecting you to someone. The better target is time-to-correct-answer. Some of that is deflection on questions the docs genuinely answer; the rest is making escalation fast and well-informed instead of a restart from zero.

What a good escalation contains

When the agent hands over, the human should not have to re-read the thread. The hand-off carries the customer’s history, their current account state, what was already tried, which documentation was consulted, and a draft response to accept or rewrite. Handling an escalation becomes a minute of judgment rather than ten minutes of reconstruction.

Support data is the best documentation feedback loop you have

Every question the agent could not answer is a gap in your docs, identified precisely and with a real customer’s wording attached. Collected over a month, that list is a better documentation backlog than any internal audit, because it is ranked by what people actually ask.

Common questions

What stops it inventing an answer?

It answers from retrieved documentation with the source attached, and escalates when it finds nothing relevant. "I do not have this documented" is a permitted and expected outcome.

Can it work inside our existing helpdesk?

Yes — it operates through the tools you already use, including email, Slack and Discord, rather than requiring customers to move to a new channel.

How do we stop it answering billing questions?

Mark that category as always-escalate. The gate is enforced by the runtime, so the agent cannot answer it regardless of how the question is phrased.

Give AI a job. Get the proof.