My AI agent is confused. Apparently, that's billable.

I've been trying to build an AI-native sales tool.

Five attempts so far. None particularly successful.

Those attempts have made me think quite a bit about the distance between watching an AI system do something impressive and getting it to produce a useful, reliable outcome.

And somewhere along the way, I became interested in a question.

What happens when the software we buy to eliminate inefficiency starts charging us for its own confusion?

The meter is running

Traditional enterprise software had a reasonably predictable commercial arrangement.

You bought licenses. Paid subscriptions. Budgeted for maintenance.

If your CRM contained terrible data, someone had to clean it. If your HR system couldn't reconcile two job titles, someone opened a spreadsheet. If a workflow broke, someone raised a ticket.

All of this cost money, of course. People's time isn't free.

But the software wasn't necessarily charging you more every time somebody struggled to make sense of it.

Now consider an autonomous AI agent.

You give it a task. It reasons, reads data, calls tools, interprets results, makes decisions and occasionally discovers that its previous decision was spectacularly wrong.

So it tries again.

Perhaps it consults another system. Perhaps it retrieves more information. Perhaps it spends an impressive number of tokens reaching a conclusion that a reasonably competent human would have arrived at by asking one question.

All of that activity consumes resources. Depending on how the system is priced, those resources can translate directly into API charges, infrastructure costs or usage-based fees.

We've effectively invented a software worker that can charge us for being confused.

And unlike a human employee, that worker can be running across thousands of workflows at once.

Now, a few extra tokens aren't going to bankrupt anyone. Sometimes spending more on reasoning is precisely what makes automation worthwhile.

An agent that uses twice as much compute but consistently produces the correct outcome could be considerably cheaper than one that gives a quick, useless answer.

Which is why I think measuring the cost of an agent by its token consumption alone misses the point.

The more interesting measure is the cost of getting something done correctly.

That includes execution, retries, exceptions, human intervention and the consequences of whatever goes wrong.

An agent that costs very little to run but creates hours of corrective work isn't necessarily efficient.

Neither is an agent that completes a task autonomously after consuming more resources than the task was worth.

We've given organisational confusion a compute budget

A significant part of my professional world involves understanding how enterprises define work, organise jobs, identify skills and make workforce decisions.

Imagine a multinational company with five versions of the same job title across four regions.

A Business Analyst in one country performs a different function from a Business Analyst somewhere else.

One system calls the role Senior Analyst. Another calls it Business Consultant. A third has apparently decided that naming conventions are an unnecessary administrative burden.

Now introduce an AI agent.

Ask it to identify which employees have the skills to move into a particular role.

The agent examines the available information. Encounters inconsistencies. Attempts to reconcile them. Retrieves more records. Consults another source. Revises its interpretation.

If properly designed, it might recognise the ambiguity and ask for clarification.

If poorly designed, it might continue trying to solve the problem, or worse, confidently produce an answer.

Either way, the underlying ambiguity hasn't disappeared simply because something intelligent encountered it.

We've taken organisational confusion and given it a compute budget.

What's especially interesting is that this problem becomes more expensive as automation scales.

A human struggling to reconcile contradictory information might waste an afternoon.

An agent encountering the same contradiction across ten thousand transactions can create ten thousand opportunities for unnecessary work.

Of course, an agent might also resolve the problem far faster than a human could. That possibility is precisely what makes the technology valuable.

But the economics depend on whether the agent is solving the underlying problem or repeatedly compensating for it.

And there's a difference.

If every execution requires the system to rediscover how your organisation defines a role, interprets a policy or routes an approval, you're effectively paying to reconstruct institutional knowledge that should probably have been available in the first place.

The funny thing is that none of this is particularly new.

Enterprises have struggled with fragmented data, inconsistent definitions and unclear processes for decades.

We just had a different way of absorbing the consequences.

A person would investigate. Someone would send an email. Another person would clarify. Eventually, the work would move forward.

Now we're increasingly asking machines to navigate those same contradictions.

And we can measure some of the effort in tokens, tool calls, execution time and infrastructure consumption.

Problems that once disappeared into the general fog of organisational inefficiency can now appear as a rather specific operating expense.

Perhaps intelligence isn't the bottleneck

I'm beginning to suspect that some of the most valuable improvements in enterprise AI won't come from making the models more intelligent.

They'll come from making the organisations they operate within less confusing.

Consistent job definitions. Reliable skills data. Clear organisational structures. Processes with actual owners. Approval rules that don't change depending on whom you ask.

Rather unglamorous things.

Nobody gets particularly excited about cleaning up job architecture.

It's considerably easier to get a room interested in autonomous agents than in agreeing what the word Manager means across fourteen business units.

Yet the second problem might determine whether the first investment delivers anything useful.

A smarter agent can sometimes compensate for messy information. It can recognise inconsistencies, ask better questions and make educated inferences.

But there's a limit to how much intelligence can compensate for an organisation that hasn't agreed on its own version of reality.

Which brings me to something I've been wondering about.

What if a significant portion of the repetitive work we want AI to eliminate exists because the organisation itself is unnecessarily complicated?

What if the process is broken, the data contradictory, and the rules understood by exactly three people, one of whom retired in 2021?

We could build an increasingly sophisticated agent to navigate all of that.

And perhaps it would do an excellent job.

Perhaps it would even be cheaper than asking a human to do so.

But wouldn't it be better to fix the mess?

I find the possibility rather amusing that, after spending extraordinary amounts of money developing machines capable of reasoning, enterprises might discover that one of the highest-return AI investments is simply getting their own information in order.

The future of intelligent automation could depend, at least partly, on some profoundly unintelligent housekeeping.

The agent isn't always malfunctioning. Sometimes it's doing its best to understand the organisation we've built.

We just happen to be paying for every attempt.