What Happens When Your AI Workflow Hits an Exception

At 9:30 in the morning on October 28, one of our own outreach jobs opened a list of eight restaurant contacts and sent nothing. The run the day before had sent twelve. Nothing failed. The job compared each contact against the list of people it had already messaged, found all eight were duplicates, and closed with duplicate 8. That quiet moment is the one worth designing for: a workflow meets a case it cannot finish on its own, and someone has decided in advance what happens next.

Where a workflow actually breaks

A clean automation has a happy path. An invoice arrives, matches a purchase order, the amount lands in the spreadsheet, a confirmation goes out. The cases that stop it are ordinary. The invoice carries no purchase order number. The supplier billed $840 where the order said $800 because two items were added at the door. The supplier’s name appears as “Bright Farms” this time and “Bright Farm Supply” on the last order. Two copies of the same invoice arrive a minute apart.

Every one of those is a decision. It either stops the line or goes to a person who opens two windows and types the answer.

Two words worth defining

MCP (Model Context Protocol) is a standard plug. It lets the AI reach the software you already keep, such as the shared inbox and the order spreadsheet, without someone building a custom integration between each pair of tools. A sub-agent is a worker that owns one workflow end to end: it reads what came in, checks it against your records, makes the update, sends the reply, and writes down what it did.

Both matter at a stop. The sub-agent is the piece that sees the missing field, and it needs a rule for whether to guess or to hold. Guessing is how a wrong number reaches a customer.

Decide the exception path first

Before you automate the happy path, write down what happens at each stop. When a field is missing, stop and ask for it. A missing purchase order number costs a minute to look up. A guessed one sails past review and shows up on someone’s bill.

When an item looks like a duplicate, hold it. A held item costs a glance. A processed duplicate costs a refund conversation and an apology.

When the numbers disagree, hand the item to a person with both figures side by side, so the question they answer is which one is right.

Each stop should leave one line a person can read later: what arrived, what did not line up, what the system would have done without the check. That line is what separates an exception queue from a silent failure.

Read the queue

Count the stops per hundred items and then read what is in them. If the same supplier shows up every week with no purchase order number, your automation just made a supplier problem visible. Our October 28 run is the small version of that: eight items held, zero sent, one glance to confirm the holds were right. A workflow that never stops is usually guessing anyway, and a guess travels further than a typo.

What this costs

For a company under 50 people, this kind of control used to belong to enterprise software with six-figure licensing and a consultant to configure it. EchoAI runs the workflow and its exception queue at a fraction of that. Starter is $299 a month for one role and one workflow, and the rest of the plans are on the pricing page. The products page shows the digital employees that own this work.

Book a demo and bring the workflow that currently stops when the data gets messy. We will map its exception path with you.