A few months ago I wrote about the “messy middle” of transformation — the point where clean process maps meet the reality of approvals, exceptions, competing priorities, and systems that do not quite work the way everyone assumed they would.
I still think that is where transformation is won or lost. But there is another way to look at the same problem: coordination tax.
Coordination tax is the time people spend checking status, waiting for approvals, reconciling data, following up with another team, figuring out who owns the next step, and updating one system because something changed in another. None of that is really the work the business is trying to accomplish. It is the effort required to keep the work moving.
That may also be one of the more practical places where AI agents can help.
The work around the work
Take a typical quote-to-order process. The core work is straightforward enough: configure the product, determine the price, approve the commercial terms, create the order, and send it downstream.
What takes time is everything between those steps.
Sales needs engineering to confirm the configuration. Engineering needs more information from Sales. Pricing may need approval. Finance may need to review terms. Someone has to confirm inventory or lead time, and another person has to make sure everything is complete before the order moves into ERP.
The delay is often not caused by the work itself. It is caused by the coordination required to move from one step to the next.
Most businesses have accepted this as normal because it is spread across everyone’s day through Slack messages, emails, status meetings, spreadsheet updates, and quick checks in other systems. None of those activities looks significant on its own, but together they represent a meaningful cost.
That is the coordination tax.
This is where agents get more interesting
A lot of the current AI conversation is still focused on making individuals faster: drafting emails, summarizing meetings, analyzing spreadsheets, or creating proposals.
Those are useful improvements, but they do not necessarily change the process.
If a salesperson can build a quote in 20 minutes instead of two hours, but it still sits for two days waiting on engineering and another day waiting for approval, the overall business outcome has not changed very much.
The more interesting opportunity is using agents to reduce the coordination tax between those steps. An agent can monitor whether required information is complete, determine whether an approval is needed, identify the right person to involve, compare information across systems, check whether an order is ready, and trigger the next step without someone manually chasing it down.
That is not replacing the work. It is reducing the amount of effort required to keep the work moving.
Coordination tax is usually invisible
Most companies do not measure it directly.
It is buried in status meetings, follow-up emails, manual checks, duplicated updates, and small interruptions throughout the day. Because no single activity looks especially expensive, the total cost is easy to underestimate.
But those delays compound.
A one-day approval delay here, a handoff issue there, a missing field that creates rework, a system mismatch that requires manual reconciliation — eventually the process becomes slower and more expensive than anyone intended.
This is why AI agents could matter more at the workflow level than at the individual task level. The value is not just helping someone complete their piece faster. It is reducing the friction between all the pieces.
The business still has to be understandable
AI does not eliminate ambiguity.
An agent can only coordinate a process if the business has defined the rules underneath it. It still needs to know which system is authoritative, when an exception requires human judgment, who owns a decision, and what conditions must be met before something can move forward.
That was the underlying point of the original Messy Middle article.
The process breaks in the exceptions, handoffs, ownership gaps, and decisions that were never fully defined. AI does not solve those issues on its own, but once the rules are understood, agents may be very good at carrying the coordination burden that people have been absorbing for years.
A better question to ask
Instead of starting with:
What work can we automate?
A more useful question may be:
Where are we paying the most coordination tax?
Where are people waiting on each other, checking whether something is complete, moving information between systems, following up on approvals, or reconciling conflicting data?
Those are the places where cycle time disappears and where AI may have a more meaningful business impact.
We are already starting to see this play out with personal agents. The value is not always that they complete some complex task for you. Often, it is that they remove the mental tax around the task.
You do not have to remember the due date, keep checking the calendar, watch for the bill, or remind yourself to follow up. The system surfaces what needs your attention when it matters.
That is a subtle but important shift. The technology is not just helping you do the work faster. It is reducing the amount of energy required to manage everything around the work.
And expectations tend to move from our personal lives into business.
Once people get used to technology proactively telling them what needs attention, why should they accept spending half their workday checking status, chasing approvals, remembering dependencies, and figuring out what should happen next?
That is where I think enterprise agents are heading.
The biggest AI opportunity may not be automating the work itself.
It may be removing both the coordination tax and the mental tax required to keep the work moving.