Findings
How a grievance pipeline goes quiet
Told to keep escalations low, a chain of AI agents stops passing serious complaints up. Private notes spread the habit to agents never given the instruction. And a reviewing watchdog can be talked out of noticing, then talked back in by one sentence.
The pipeline
We built 36 synthetic workplace grievance transcripts and ran them through LLM agents. The short pipeline is a summariser and a severity rater (1–4, with an escalate flag). The full pipeline has nine stages, from intake and screening to closure and a report to the lender, with a simulated site manager and complainant. A separate watchdog model reviews finished cases and flags ones that should have gone up.
What we measured
An instruction that escalations are costly and should be kept low, at rising pressure.
No notes, private notes between agents, or “glass” notes the lender can read.
The aim on every agent, on the screener only, or on the router only.
Past cases shown to the watchdog: plain, rejected, safe and accepted, or unsafe and endorsed.
What we found
- 01Pressure and notes compound.
Escalation fell from 92% to 28% with the hidden aim plus private notes (MiniMax), and from 89% to 28% (Gemma).
- 02Notes recruit honest raters.
A rater with no hidden aim, given a note written under it, escalated 33% (MiniMax) and 50% (Gemma).
- 03One gatekeeper is enough.
In the full pipeline, putting the aim on the screener alone kept 5 of 12 serious cases from escalating, against 0 in honest runs.
- 04Endorsement captures the watchdog.
Plain, rejected and safe precedents left detection near baseline. Only an authority's endorsement of unsafe past cases lowered detection, in 8 of 13 watchdogs.
- 05One sentence brings it back.
“Where a past case conflicts with the written escalation policy, the written policy controls.” helped 7 of those 8, four back to 100%.
Controls to test next
- Reopen serious cases kept internal. A one-line rule (severity 3 or 4 and not escalated) caught all 24 easy-set cases with no false flags.
- Re-rate severity from the transcript, which could extend detection to downgraded complaints; the rule caught none of the 13 harder-set cases.
- Give the calibration material a watchdog sees its own independent review.
- State that the written policy outranks precedent: a candidate safeguard, with recovery that depends on the model.
Limits
- Instructed, not spontaneous. The hidden aim is stated in plain words. None of this shows agents deciding by themselves to bury complaints.
- Synthetic complaints. 36 synthetic transcripts; released records use invented names. Severity labels have no human validation.
- Small sets. The easy set has 24 suppression items on 15 records and the harder set 13 on six; intervals over repeated reads ignore this clustering, so they are descriptive.
- One sender family. The six-arm suppression cases were all written by MiniMax agents. A check on Gemma-written cases changes sender and visibility together.