Newsroom

Every operations leader I talk to has the same queue somewhere. It might be called "manual review", "exceptions" or "needs attention". It is where work lands when a bot can't finish it: the invoice in an unusual layout, the claim from a portal that changed last week, the onboarding file with a scanned document instead of a typed one.

The bots upstream of that queue are usually doing their job. They take the clean, structured, predictable items and process them fast. The trouble is that the queue never empties, and the people working it are often the most experienced people on the team.

This post is about that queue, and why the answer to it is not more scripts.

Why scripts hand work back

Scripted automation is built on rules: go to this screen, find this field, copy this value, click this button. When every input looks the same, rules are fast and reliable.

Exceptions are, by definition, the inputs that don't fit the rules. A scanned PDF instead of a structured file. A portal where the button moved. A Citrix session where the bot can't see the underlying fields at all. A case where someone has to decide whether a discrepancy matters.

You can keep writing rules for each new variation. Most teams do, for a while. Then the cost of maintaining the rules overtakes the value of the work they automate.

What a computer-use agent does differently

A computer-use agent works the way a person at the keyboard does. It looks at the screen, works out what it is looking at, and decides what to do next. Three things follow from that:

1. It isn't tied to a layout. If a field moves, the agent finds it where it is now.
2. It can read what a rule can't. An odd invoice format or a scanned form is something to read, not a reason to stop.
3. It works through the screens you already have. Web portals, desktop applications and Citrix sessions, with no new APIs and no changes to your systems.

And for the steps that need judgment, the agent routes the item to a person, with the work prepared. You set the thresholds: what it can close on its own, and what always goes to review.

Keep the bots

This is not a rip-and-replace argument. If a bot runs reliably on a stable screen, leave it alone. It is cheap, it is fast, and it is doing its job.

The opportunity is in the work bots hand back. In one deployment, H agents took on invoice validation for a BPO that had been doing three-way matching by hand in legacy systems. Processing got 40% faster and errors fell by 30%, with operators reviewing the agent's output rather than keying every invoice. In another, agents took over a KYC onboarding workflow that crossed eight to twelve steps and several systems, and now handle more than 80% of it, with a person signing off the final step.

In both cases the existing automation stayed. The agents took the part that had always landed on people.

How to start without a project

The safest way to find out whether this works on your queue is to try it on your queue, in a sandbox, for two weeks:

• Pick one exception queue with clear success criteria, ideally 1,000 or more items a month
• Give the agent a user-level account in a test environment, read-only to start
• Agree what "done" means with the process owner
• Measure the completion rate, and look at every item the agent routed to a person and why

That is what our Exception Queue Sprint is: one queue, two weeks, and a completion rate your sponsor can act on. Your bots keep running the whole time.