How do I hand ticket triage to AI without losing control?

Why owners stall on this decision

Owners stall because a wrong call on a client's ticket lands on them, and vendor pages answer that fear with an accuracy figure. Around 65% of Australian SMEs not using AI cite distrust in its decisions or a wish to keep human control. Accuracy alone will never settle that for you.

That figure comes from the National AI Centre's monthly survey of at least 400 SME owners and decision-makers¹. Vendors know the fear. ConnectWise promises triage "while keeping you in control"⁴, Mizo states 95%+ classification accuracy³, and a partner on Kaseya's page reports around 98% within a week².

Those numbers come from vendor pages, and none of them says what the result was checked against. Control comes from two things you can set yourself: a fair test and a clear line on who decides what.

Test your techs before you test AI

Your category field makes a poor answer key. One partner on Kaseya's page says 20 to 30% of its tickets are not categorised correctly today. If you score AI against tickets like those, you punish it for disagreeing with a mistake and reward it for copying one. Measure your own desk first.

Run this test in an afternoon:

1. Pull your last 100 closed tickets, worst two line emails included.
2. Hide the category, priority and team on each one.
3. Ask two senior techs to set all three, apart from each other.
4. Count the tickets where they disagree. That rate is your baseline.
5. Have them settle each disagreement. The settled list becomes your answer key.

Now you hold a pass mark you can defend in front of your team. AI that matches your answer key as often as your two seniors match each other has reached the standard your desk runs on today. Your rule list comes out of this too, because each disagreement points at a category or priority nobody wrote down.

Worked example, not a client result: the three fields take 2.5 minutes a ticket by hand, so two techs and 100 tickets is 500 minutes. At A$55 an hour⁶ the whole test costs you A$458.

Who decides what once it goes live

Split the work by what a mistake costs you. A wrong category takes a tech two seconds to change. A wrong message reaches a client, and a wrong billing call costs you money or a relationship. Hand AI the first group on day one, and let the other groups earn their way out or stay with a person.

The original words stay on the ticket too. Ticket Clarity adds a clear title and short summary and keeps the user's words as they wrote them⁷, so your tech can always read what the client said.

Your off switch

Control means you can stop it in one move. Cobotic works through a dedicated user you create in your own PSA, with only the access the workflows need. Switch that user off at any time and every workflow stops, and your tickets, your set-up and your data stay where they always were.

The rest of the safety net looks like this:

• If AI goes down, your tickets stay in the normal queue, untouched.
• Your runbook shows what each workflow does, how to check it, how to change it and how to switch it off.
• Your PSA audit log shows every action the Cobotic user takes⁹.
• Cancel and we delete your PSA key within 24 hours⁹.

Worked example: what triage costs and what control costs

Triage costs far more than the control around it. Take 500 tickets a month, with the four triage fields taking 3.25 minutes a ticket by hand. That is 27 hours a month, or A$1,490 at A$55 an hour. Checking the review queue takes minutes a month, even if one ticket in 20 lands there.

This example describes no client. The 5% queue rate and the 30 seconds are our assumptions, so set your own.

A$11.46 of review buys you a person's eye on the tickets that need one. Sending work up the line costs far more. MetricNet's benchmark, 2018, prices a ticket at US$22 on the service desk and adds US$69 when it moves to desktop support, and it counts 17.8% of desktop tickets as work Level 1 could have closed⁵.

Four numbers to watch in the first month

Read four numbers each week and you will know whether to widen the handover or tighten it. Your PSA already holds most of them. Each one moves on its own, so a drop in one tells you which rule set to fix. Compare each number with the baseline from your own test.

Override rate does the most work. Falling week on week means your rules are settling. A spike on one client or one category points at a rule to fix, and you send that change to Cobotic as part of the plan. Other Category Cleanup and Closing Category Fix then keep the list honest on each resolved ticket.

The handover, day by day

You can run this whole handover inside a month, and you can see results before you give anyone access. The Test Drive needs a ticket export from you and nothing else. First Month then runs on your real tickets, with messages to users held for approval during the first two weeks.

1. Today: run your last 50 tickets through the free Test Drive. Your browser removes contact details before anything is sent, and you see exactly what we receive. Compare its calls with your techs' calls.
2. Same afternoon: run the 100 ticket test above and set your pass mark.
3. Day 1: start your first month for A$99. Create the Cobotic user and all AI workflows run in your PSA in under 15 minutes. No category list or priority matrix? We build them from your recent tickets.
4. Days 1 to 14: your techs approve each request for missing details. Pick the first Self-Fix Reply categories, starting with the simplest.
5. Day 15 onward: details requests go out on their own. Read the four numbers each week.
6. Day 25: we send a reminder. Cancel in one click, or the plan starts on day 31 at A$990 a month for your first 1,000 tickets, then A$0.99 a ticket.

Already using the triage in your PSA? Keep it. Our comparison shows the workflows it leaves on your techs.

What control looked like at a Perth MSP

A 20 person Perth MSP handed sorting and routing to six AI workflows inside its PSA. The team kept the judgement calls and the customer conversations. Every ticket now arrives sorted, assigned and with a likely fix attached, and the case study records $200,000 saved.

Read the Perth case study¹⁰. A five person MSP in Tasmania ran the same split on a much smaller desk, and its owner came off the tools.