Support, service desks and intake teams

Answer faster without answering blind.

Support queues fail in predictable ways. The same questions arrive over and over, urgent items sit behind routine ones because nothing triaged them, and first response time slips because the person who can answer is busy answering something else. None of that needs a bigger team. It needs the routine traffic handled and the rest routed properly.

Book a scoping call Try the 2 minute fit check

Four workflows, in your vocabulary

The work, not the buzzword.

Each one says what arrives, what the agent does with it, where a person still stands in the way, and what gets counted afterwards. If a workflow cannot be written that way, it is not ready to be built.

Triage on arrival

What arrives: Everything the queue receives: email, web form, portal message and the notes typed up from phone calls, all landing in the order they happened to be sent.
What the agent does: Reads each item as it lands, works out the category, the urgency and the account it belongs to, applies your priority rules and routes it to the right queue or person with a short summary at the top.
Where the human gate sits: Priority overrides stay with the team lead, and anything the agent categorises with low confidence goes to a person rather than to a queue where it may sit.
What gets measured: Time from arrival to correct queue, the re-routing rate after triage, and the share of urgent items identified within your target window.

First response drafted

What arrives: A customer message that has a real answer somewhere in your knowledge base, past tickets or product documentation, which the agent handling it will spend most of their time locating.
What the agent does: Finds the supporting material, drafts a reply in your tone with the source attached to each claim, and presents it in the ticket ready to send, edit or reject.
Where the human gate sits: Nothing sends without a person releasing it, until you decide otherwise on a specific category and only after the accuracy on that category has been measured.
What gets measured: First response time, the share of drafts sent unedited, and the edit rate by category, which is what tells you where autonomy is safe to widen.

Repeat question deflection

What arrives: The questions that arrive every week: hours, status, how to reset it, where to find the form, what the fee is, each one a small interruption with a documented answer.
What the agent does: Answers from your own documented material only, quotes the source, and hands anything unsupported straight to a person rather than composing a plausible answer.
Where the human gate sits: The customer is told they are dealing with an automated reply and how to reach a person, and any request for a human ends the automated exchange immediately.
What gets measured: Deflection rate on the categories in scope, the reopen rate on deflected items, and the escalation rate, which matters more than the deflection number on its own.

Escalation and complaint routing

What arrives: An item that is not a routine question: a complaint, a vulnerable customer, a legal or safety concern, or a case the agent could not support.
What the agent does: Detects the signals you define, stops any automated handling on the item, assembles the history, the account context and what has already been said, and routes it to the named owner for that category.
Where the human gate sits: Complaints and sensitive categories are never resolved by an agent. They go to a person, and the routing is recorded with a timestamp so the response clock is evidenced.
What gets measured: Time to a named human owner, and the share of complaints identified at the first message rather than the third.

The results we can actually evidence

No invented support study. Here is how yours would be measured.

faster intake to instruction

45% Client's own tracked operating KPI, before and after the same definition

per land sale contract reviewed

1-2 hrs to ~10 min Client's existing time tracking, same review step before and after

Both measured results come from legal and conveyancing clients, against metrics those firms were already tracking. We will not publish a support figure until a support client produces one. The method is the part you can hold us to now: first response time from your helpdesk timestamps on the same queue definition before and after, an eight week window each side. Your helpdesk already holds both numbers, which is what makes this workflow straightforward to baseline honestly.

Read how both were measured

Privacy, complaints and telling people what they are talking to

What the sector will not let an agent do.

Support is the one function where the agent may speak to your customer, so the obligations are about disclosure and about knowing when to stop, not only about where the data sits.

Privacy Act obligations over customer information: The agent reads only the systems you connect, with the permissions you grant. Australian hosting is available across the deployment options, customer conversations are not used to train any model, and prompts and outputs are not retained by the model provider. Log retention is a setting you choose.
Complaint handling: Complaints are a detection and routing job, never a resolution job. When the signals you define appear, automated handling stops on that item, a named person becomes the owner and the timestamp is recorded so the response window can be evidenced.
Telling the customer what they are dealing with: Automated replies say so, and every one carries a route to a person. Asking for a human ends the automated exchange straight away. An agent that pretends to be a person is a reputational risk with no upside.
Vulnerable customers and hardship: Treated as a stop condition, not a category to be answered. The signals are defined with you during solution design, and the item is handed to the person who is trained to handle it.
What the customer was told, and when: Every draft, release, deflection and escalation is logged with the source material behind it, so the record of what was communicated exists without anyone reconstructing it from a mailbox.

This is a description of how the deployment is built, not legal, financial or compliance advice. Your own obligations should be confirmed with your own advisers. The security page sets out hosting, model choice, retention and access in full.

How it starts

A pilot, not a transformation programme.

One workflow, a baseline agreed before anything is built, an agent running supervised inside the tools you already use, and a keep-or-kill number at the end. Pilots start from $5,000.

Book a scoping call Try the 2 minute fit check

Common questions

The ones people actually ask.

Will it reply to customers on its own?

Only if you decide it should, and only once accuracy is proven on a supervised queue. It starts by drafting for a person to release, and any category can stay approval-gated permanently.

What happens when it does not know?

It escalates. An agent that guesses on a support queue creates more work than it saves, so it is configured to hand over rather than improvise, and it answers only from your own documented material.

How do you handle complaints?

As a stop condition. When a complaint, a hardship signal or a safety concern is detected, automated handling stops on that item and it goes to a named person with the history attached and the time recorded.

How do we prove it helped?

First response time, deflection rate and reopen rate, all from your own helpdesk, measured the same way before and after on the same queue definition.

Where else this shows up

Document-heavy operations Professional services Melbourne and Victoria

All industries Use cases by workflow Security and trust Case studies How we work