Meridian: an agent nobody could steer
An illustrative case study — a composite scenario, not a real client — built to show what an engagement looks like when things are genuinely messy. Meridian is a legacy logistics operator that bolted an AI agent onto three systems that don’t agree with each other. The result came in stages, not overnight.
An AI agent bolted onto three systems that didn’t agree — with no record of its actions and no one owning sign-off.
A capable agent, pointed at chaos.
The agent wasn’t the problem — the ground under it was. It acted on whatever data answered first, with no record and no owner. Trust collapsed, and staff started routing around it.
- It acted on stale data from whichever of three systems replied first
- Approvals had no owner — some things sent, some stalled, nobody knew why
- No record of what the agent did, or why
- Staff had quietly gone back to doing it by hand
The agent stopped guessing.
A staged fix, not a silver bullet.
You can’t rebuild trust in a week. But you can stop it getting worse on day one, then rebuild in stages.
Stage 1 — Stabilise · weeks 1–3
A human approval gate on every outbound action, and log everything. The bleeding stops immediately.
Stage 2 — Untangle · weeks 4–8
One source of truth per decision. The agent reads from that and flags conflicts instead of guessing.
Stage 3 — Earn it back · months 2–3
With reasoning visible and actions reversible, staff stop routing around it. Measured value returns.
You can’t fix a trust problem in a week. You can stop it getting worse on day one. Then you rebuild, in stages.
A staged fix: stop the bleeding on day one, then rebuild trust over the next three months.
Messy is normal.
If your AI is bolted onto systems that don’t agree, that’s exactly the conversation to have. Start with an audit.