Most operations teams have automated something. A bot that copies invoice data into the ERP. A workflow that routes forms by a dropdown value. A script that moves files from one folder to another. On the dashboard, the process looks automated.
Then you walk the floor and find people doing the same work they did five years ago. They open documents, read them, decide what they are, re-type the important parts and pass them on. The automation hasn't failed. There are just too many exceptions.
Rules need a tidy world
Classic automation, whether RPA or rule-based workflows, does exactly what it is told. That is its strength and its limit. It expects a known input in a known format and follows a logic path.
Business inputs are rarely like that though.
Invoices come in hundreds of layouts.
Insurance submissions mix forms, scans, spreadsheets and free-text emails.
Tax and finance teams work with documents in several languages and tables that change from one source to the next.
Every variation is either a new rule to write or a case that drops out to a person.
So teams end up maintaining a growing rulebook, while the exception queue keeps growing anyway. This is where the hidden cost of automation is.
The hidden cost sits in the queue
Exceptions are expensive in ways that rarely appear in a business case:
- Re-typing. People spend hours typing stuff from documents into systems. Every manual step is a chance for an error that surfaces later, when it costs more.
- Waiting. A case that drops out of the automated path waits for someone to pick it up.
- Scaling. When volumes rise, the only lever left is hiring or outsourcing. The automation doesn't absorb growth; the queue does.
- Lost knowledge. How to handle the odd case? This knowledge lives in experienced people's heads, not in the system.
If you want a quick test, ask your team one question: of everything that is "automated", how much still passes through a human inbox? The answer is usually higher than the dashboard suggests.
What changes with AI agents
An AI agent doesn't need every case to look the same. It reads a document the way a person would, works out what it is and what matters in it, and uses your existing tools (mail, documents, databases, ERP or CRM) to act on it. When something is genuinely unclear, it doesn't fail silently. It hands the case to a person with the relevant information already pulled together.
That changes the role of your team. People stop executing routine steps and start supervising: checking the cases that need judgement, and correcting the agent when it gets something wrong, so it improves over time.
Old automation followed rules. Agents get the work done.
Where to look first
Before thinking about tools, look at your own exception queue. Three signals point to a process worth rethinking:
- Information arrives in documents or emails, not in clean structured fields.
- The same "automated" process still needs people to read, classify or re-type.
- Volumes grow faster than the team, and the plan for next year is more headcount.
If all three are true, the bottleneck isn't effort. It is the type of automation.
At Two Impulse we build AI agents that work inside existing systems for organisations across Europe, including in insurance and tax advisory.