An agent on a broken system is a faster broken system
Why adding AI agents to the way a company already works rarely pays off, and why redesigning the operation first is the move that does.
Every company already runs on an operating system. Nobody installed it and nobody chose it. It is the sum of habits: the report someone builds by hand every Monday, the invoice that waits in the wrong inbox, the customer message that gets answered when someone remembers. Nobody calls it a problem. It is just how things work here.
Now the market is selling agents. An agent for support, an agent for sales, an agent for reporting. Each one is sold as a shortcut. Install it, and the work gets done.
Here is the uncomfortable part: an agent placed on top of a broken process does the broken process faster. If a report is late because the data lives in five places, an agent that writes the report still has to find the data in five places. If customer messages are lost because nobody owns them, an agent that answers them still has nowhere to send the ones it cannot solve.
The tool is not the bottleneck
Most operations do not fail for lack of software. They fail in the gaps between systems and between people: the hand-off nobody owns, the decision that waits for a number, the number that waits for a person.
Adding a tool adds another place for work to wait. That is why so many AI pilots look impressive in a demo and disappear in month three. The demo runs on a clean path. The operation runs on habit.
Redesign first, then automate
The companies that get real value from AI do something less exciting first. They redesign how the work gets done, and only then put agents inside the new design.
In practice, that means four questions, in this order:
- Where does work wait? Not where it is hard: where it sits. Waiting is the most expensive thing in an operation and the least visible.
- Who owns each hand-off? If the answer is "whoever sees it first", that is the gap an agent will fall into.
- What does done look like? A process without a clear end state cannot be automated, only accelerated.
- What should a person still decide? Autonomy is not the absence of people. It is people deciding only where judgement matters.
- 01Find the waitWhere work sits, not where it is hard
- 02Own the hand-offsNo gap left to “whoever sees it”
- 03Define doneA clear end state for every process
- 04Keep judgement humanPeople decide only where it matters
- 05Put agents insideNow automation compounds
Answer those, and agents stop being add-ons. They become the way the system runs.
What it looks like when it works
When we rebuilt event reporting for OCESA, the problem was not that nobody could write a report. It was that the evidence for one event lived in ticketing, sponsorship records, ratings, presentations and media files, and every question meant a hunt. The reporting operation was redesigned around the question first; the system that answers it came second. The reporting cycle went from weeks to minutes.
Source: Sentynel project case record (OCESA + Smart Sponsorship). Reported project outcome.
The agent was not the breakthrough. The redesign was. The agent made it run.
The question to ask before buying anything
Before you add an agent to anything, ask: if this agent worked perfectly, would the work still wait somewhere?
If the answer is yes, you do not need an agent yet. You need a new operating system, and then agents that live inside it.
Don't add. Replace.