Agents The declined fix that comes back three times gets flagged for you
It is the third visit this year to the same rooftop unit, and the service manager is staring at three separate job summaries that each mention the same worn component — noted, quoted once, declined, and never connected.
Source: ServiceTitan, "Pantheon 2026: Live coverage from ServiceTitan" — https://www.servicetitan.com/blog/pantheon-2026-live-coverage — the coverage describes technicians logging findings by voice, photo, or text, an application that scans notes, photos, and job summaries, and a flag when the same issue appears three times so sales can revisit a declined recommendation; a companion agreement-review agent consolidated scattered agreement data and, on its first live run, flagged a problem in the first two quarterly invoices. Adapted here for a small shop; the named shop’s growth and margin figures from the session are not used.
The problem
A declined repair rarely gets declined three times. It gets declined once, out loud, and then mentioned quietly — a line in a job summary, a photo of the same corrosion, a voice note saying “still running, still loud.” Each mention lives on a different work order, and nobody’s job is reading all three side by side.
So the pattern surfaces the expensive way: when the component finally fails, on the customer’s busiest week, and the account asks why nobody told them it was getting worse. The shop did tell them — three times, in pieces. It just never assembled the pieces while the fix could still be planned.
How the agent workflow works
- Techs log findings the way they already work. Each finding goes in by voice, photo, or short text, attached to that customer and unit — no separate form, no end-of-day write-up from memory.
- The agent scans what nobody rereads. It reads notes, photos, and job summaries across completed work orders, including problems mentioned in passing but never written up formally.
- Repeats get grouped by asset and issue. Mentions of the same problem on the same unit are matched visit by visit, so a pattern builds in the background instead of relying on one tech’s memory.
- The third appearance raises the flag. Sales or the service manager gets a revisit packet: the original recommendation, when it was declined, and what each later visit saw, photos attached.
- A person decides how to reopen it. The agent does not call the customer. The manager picks the moment — often the next scheduled agreement visit — and reframes the fix as planned work with evidence.
- The same consolidation checks the agreement itself. A companion review pulls agreement terms, visit history, and invoices onto one screen and flags mismatches — in the source account, its first live run caught a problem in the first two quarterly invoices.
What you need to set it up
- One field platform where techs already log the job, with voice, photo, and text findings turned on — a second app just for findings will be skipped by Friday.
- Consistent customer and equipment records. A repeat only counts if all three mentions attach to the same unit; three units logged as one will invent patterns.
- A written definition of a finding versus a casual note, and declined recommendations recorded with their dates — the flag is only as honest as that first record.
- A threshold and routing rule the shop sets itself: three appearances is the source’s number, and the flag goes to a named owner of the revisit. Cost, honestly: the source publishes no flat price, so expect it inside a platform quote; internally, the cost is standardizing how findings get logged.
Where it breaks
Techs describe the same fault three different ways. Worn, noisy, and end-of-life can be one problem or three — and a photo shows a symptom, not a cause. The matching will miss real patterns and merge unrelated ones, so the packet is a lead for a human to verify, not a diagnosis to defend in front of a customer.
Some findings should never wait for three. Anything safety-shaped — a cracked heat exchanger, a frayed disconnect — needs a same-day escalation path that bypasses the pattern count. Three strikes is a rule for slow problems only.
Declined sometimes means decided. The customer may be budgeting a full replacement or selling the building. A flag that ignores that context turns good record-keeping into resell pressure.
The bottom line
Run it on agreement accounts first, where the visit history is densest. The win is simple: a recommendation the customer said no to once stops quietly expiring in old work orders, and comes back — with three visits of evidence — while it is still a planned fix instead of an emergency.