# The exception path is the operation

> Every agent demo shows the easy case. Production is the exceptions, and the company that designs where they go owns the real operation.

- Type: Opinion
- Published: 2026-10-30
- Author: Sentynel
- URL: https://sentynel.ai/sygnals/the-exception-path-is-the-operation/
- Language: en

Most AI agent demos show the happy path: a request arrives, the agent handles it, everyone claps. **In production the happy path is the part that takes care of itself. The operation is whatever happens to the cases the agent cannot finish.** If you have not designed that path, you have not designed the operation.

## The easy case is not where the work is

Every process has a main road and a long tail of odd cases: the customer who writes in two languages, the order with a missing field, the request that touches two departments. People have always absorbed that tail without noticing, because a colleague just knew who to ask.

An agent does not know who to ask. Give it the tail and it will do one of three things: guess, stall, or hand the case to nobody. **Each of those is a design decision you made by not making one.**

## Three questions every agent needs answered in advance

- **What counts as an exception?** Write it as a rule the agent can test: confidence below a line, a field missing, an amount above a limit, a customer flagged as sensitive. "When it seems unusual" is not a rule.
- **Who receives it?** A named role with a queue, not a channel everyone watches and nobody owns.
- **What does the receiver get?** The original request, what the agent already did, and why it stopped. Otherwise the person starts from zero and the agent saved nothing.

**An exception path, not a dead end**

1. Work arrives — one owned queue
2. Agent attempts it — inside written limits
3. Exception test — a rule, not a feeling
4. Named owner — gets the full context
5. Outcome recorded — and fed back into the rules

## The tail is also where you learn

A well-designed exception path is not a cost. It is the cheapest source of improvement you have. Each case a person resolves shows you a rule the agent was missing, or a rule that should never have been delegated. Review them weekly and the agent's share of the work grows on evidence, not on optimism.

The opposite is the silent tail: cases that fall out of the flow and are never counted. The agent looks excellent because the hard cases left the dataset.

## What a measured operation looks like

In Sentynel's Rodo case, WhatsApp, social and web enquiries were brought into one service operation with a visible path to resolution: New, In progress, Resolved. In the first measured window, 82% of 604 live-created enquiries reached Resolved, with a median of about one hour. The same record shows the other side: roughly one in five had not reached Resolved, and because every enquiry sits in a stage, those are visible and countable rather than lost.

**Rodo · first measured window**

- 82%: of live-created enquiries reached Resolved
- ~1 h: median to resolved

Source: Sentynel: Rodo operational report, 24 September 2026. Cohort 31 August–24 September, 604 live-created enquiries. Figures describe pipeline records, not all conversations. First measured window, no comparable earlier baseline.

That is the point of the example. The headline number matters less than the fact that the remainder is counted.

## What to do next

- **List the last ten cases your team had to rescue** from an automated step. Those are your real exception rules.
- **Write each as a testable condition** and decide the owner.
- **Count exceptions as a metric** next to the success rate. A rate with no denominator for the tail is a flattering story.

A question for your next review: when your agent gives up, does a specific person find out, with the context, in the same hour?
