Helpdesk · AI agent
An AI agent that does the work
Most support AI writes a reply and stops. This one opens Stripe, checks the invoice, issues the refund, and sends the answer citing what it read — and asks you first whenever you have said it should.
Free 14-day trial. No credit card required.
Counted in our own production workspaces — our email-deliverability business DMARCLY, and Airclou support — from the agent’s first action in May 2026 to September 2026. Exact counts, never rounded up.
How it works
Read the real systems, then act on them
The answer to a support ticket is rarely in the inbox. It is in the billing system, the account record, the live DNS. The agent goes and gets it — and the same connection it read from is the one it can act through.
It gathers the evidence before it writes
A ticket arrives and the agent goes looking: the customer in Stripe, their subscription and invoices, your knowledgebase, your own admin API. Then it drafts — and every claim in the draft points back at the article or record it came from, so a reviewer can check the reasoning instead of trusting it.
- Chains a lookup on its own — invoice number to customer to subscription
- Cites the knowledgebase article and the record behind every claim
- Withholds what it cannot ground rather than guessing
Every write stops for a yes
When the agent wants to do something — refund a charge, cancel a subscription, update a record — the run parks and the proposal comes to a person. The card shows the exact action, the amount, the customer it affects and what the agent verified to get there. Approve, edit, or reject.
- Gated by default — nothing writes until it is allowed to
- The proposal names the amount and the customer, not just “an action”
- Approve from the ticket, the queue, or your phone
Autonomy is a dial you turn
Start with everything gated. As the agent earns it, open up one action at a time: which reply kinds may send themselves, which writes may run alone, and up to what amount. Everything inside the bound just happens; everything outside still waits for you.
- Per action and per amount — a $50 refund cap is one number
- Reply kinds graduate separately, from clarifying questions to commitments
- Real-money writes need an explicit live-key acknowledgment
Go further
What it runs on
Airclou Helpdesk
The inbox the agent works inside — email, live chat and portal in one routed queue, with the full audit trail.
Explore HelpdeskAI grounding
The published playbooks the agent answers from — cited by name and version, never from drafts.
Explore AI groundingApprove from your phone
The iOS companion puts every waiting proposal a tap from sent — same envelopes, same audit trail.
Explore Helpdesk for iOSQuestions
Frequently asked
What can the AI agent actually do?
Read and write in the systems your answers live in. Out of the box it looks up customers, invoices, subscriptions and payment methods in Stripe, searches your knowledgebase, triages and assigns the ticket, and sends the reply. It also issues refunds, cancels subscriptions, creates quotes and updates customer records — those are writes, so they are gated by default.
What stops it doing something you did not want?
Every write is gated by default — the run stops, a person approves or rejects, and only then does it execute. You open that up deliberately: auto-approval envelopes say which actions may run alone and up to what amount. Anything outside the envelope parks for a human, and anything touching real money requires an explicit live-key acknowledgment before it can ever auto-approve.
Does it make things up?
It answers from what it retrieved, and every draft carries its sources — the knowledgebase articles and the records it actually read. It is held to that: claims it cannot ground are withheld rather than guessed, and it will not volunteer a link it did not verify. When it cannot answer, it hands the ticket to a person with what it found, instead of filling the gap.
Can it reach my own systems, not just Stripe?
Yes. An admin can define a connector against your own HTTP API — base URL, credentials, and the reads and writes you want to expose — entirely as configuration, with no code from us. Those actions flow through the same approval gate and the same audit trail as the built-in ones.
Is the AI agent a paid seat?
No. Pricing is per human teammate on your support team — the AI agent is included in every plan, not billed as an extra seat.
Get started
Give it the boring half of the queue
Start with every action gated, watch what it proposes, and open up only what earns it.
Free 14-day trial · No credit card · Cancel anytime