Design the Escalation Path Before the Agent's Happy Path
What makes an agent safe in production is not its accuracy but what happens in the small share of cases where it is wrong. Silent failure is the default state of every scheduled automation ever built.
Bence FedorIT Solution ArchitectFour failure modes to design against
Confidently wrong: the output looks correct and there is no signal. Silently stopped: the scheduled job failed eleven days ago and nobody noticed. Partially completed: the record was created but the notification never sent — worse than doing nothing, because the system now says it is handled. And looping: the agent chases the same client every morning.
Controls, in order of value
A daily heartbeat message to one channel stating that it ran, processed N items and escalated M. The absence of the heartbeat is the alarm. Then volume caps on any outbound action, idempotency so a rerun cannot double-send, and a plain-language audit log of what the agent did and why.
Escalation design is four decisions: who receives it, in which tool they already use, with what response time, and what the agent does while waiting. Undefined waiting behaviour is where duplicates and abandoned cases come from. Never route escalations to a shared inbox — shared ownership is no ownership.
Agree the kill criteria while everyone is optimistic
If escalation rate exceeds a stated threshold for two consecutive weeks, or a client complains about agent output, it is switched off and reviewed. That is a much easier conversation before an incident than during one.
An 85%-accurate agent with clean escalation beats a 97%-accurate agent that fails silently, every time. Accuracy demos well; recoverability determines whether you can leave it running unattended.
For every automation currently running, confirm you would know within twenty-four hours if it stopped. Wherever the answer is no, add a daily heartbeat message into one channel this week.
Designs the API and integration layer that automation depends on, across systems never built to connect.
Meet the team →Questions we get asked
What single question should I ask an AI vendor?
How will I know within twenty-four hours if this stops working? Most have no answer. Silent failure — not inaccuracy — is the default state of every scheduled automation.
Is a more accurate AI agent always better?
No. An 85%-accurate agent with clean escalation beats a 97%-accurate agent that fails silently. Accuracy demonstrates well; recoverability determines whether you can leave it running unattended.
Where should an AI agent send cases it cannot handle?
To a named person, in a tool they already use, with a defined response time and defined agent behaviour while waiting. Never to a shared inbox — shared ownership is no ownership.
Related insight
The AI Assistant That Does Everything Does Nothing
An agent's reliability collapses as you add systems it must be simultaneously correct about. The only agents that reach production…
AutomationYour Software Is Fine. The Gaps Between It Cost You Forty Hours a Week.
The recoverable waste in most operating companies sits in the transfers between systems, not inside any one of them. No vendor is …
AutomationAutomate a Broken Process and You Buy a Faster Broken Process
Automation makes a process permanent. Any step you encode without questioning becomes structurally harder to remove than it was th…
Where is your business still running manually?
We review your processes, identify the highest-impact automation opportunities and estimate their financial impact — before anything is built.