Which tickets to hand over first, how to keep it safe, and how to know when the agent has earned more trust.
Maintenance work is necessary, repetitive and rarely anyone’s favourite job. And most have been wanting to hand it over to AI for ages, but we always hear these horror stories about how AI agents mess things up for real businesses.
Except at Jem, whose WhatsApp-based HR platform serves more than 250,000 frontline workers and sees 92% of support tickets now handled by AI.
We asked Jem co-founder and CEO Simon Ellis how his team handed its maintenance work to an agent without all the risks…
The move: start with a reviewer on everything, then step back
Jem has been working towards this for about 18 months. Everything started with an expert in the loop, someone who knew the system reviewing every piece of the agent’s work. The team only moved forward once they had done enough to feel comfortable. Today the agent is short of work.
“The team says the agent is actually struggling for work.”
How to hand routine engineering tickets to an AI agent
1. Separate maintenance from bugs and new features
Simon is clear that these are different kinds of work. Maintenance differs from fixing bugs, and both differ from building new features. Jem started with maintenance because it is the most repetitive and predictable. Fully autonomous feature work is what Simon calls the next frontier, not the starting point.
2. Begin with low-risk, well-defined tickets
Within maintenance, pick tickets that are frequent but easy to check: dependency updates, small fixes, test repairs, documentation. Keep sensitive areas out of the first wave, such as payments, logins and changes that delete or rewrite data. Those can follow once you understand how the agent behaves.
3. Have an expert review every change at first
In the early stage, nothing the agent produces goes live without a person who knows the system checking it. The agent proposes the change; the expert approves, corrects or rejects it. This is slower than full automation, but it’s how you learn where the agent is reliable and where it isn’t.
4. Measure how often its work needs no changes
Track what share of the agent’s changes get approved without any edits, broken down by type of ticket. PagerDuty’s engineering team uses this measure to decide which kinds of work an agent could handle alone. A high, steady number for one ticket type is the evidence that it’s ready for more freedom.
5. Hand over one ticket type at a time
When a type of ticket keeps passing review untouched, let the agent handle it with lighter checks, then move to the next type. Keep the right to take autonomy back if mistakes start appearing. That steady, reversible approach is how Jem moved from reviewing everything to an agent that resolves 92% of maintenance tickets alone.
The big payoff
For a 10-person engineering team, two days a month of maintenance is now largely handled by the agent. That time goes back into building the product, and the team has a tested path for handing over more work as the tools improve.
You might also be interested in
Want the full story?
Simon’s full panel from The Open Letter Cape Town 2026 is available to members inside the Founder Collab, where he goes further than we could cover here:
Why Jem nearly killed its AI experiments too early, and what it cost
How to get Claude to spot your repetitive work every week
How to share context across offices in three cities
How Jem helps customers like McDonald’s adopt AI without fear
Why he wants a thousand AI-native people, not ten
You’ll also get access to 40+ masterclasses from SA founders and operators on sales, fundraising, UX, paid media and more inside The Founder Collab.
Get more SA tech and business news, tips and business-building workflows and subscribe to The Open Letter.
