Humans Shouldn't Be the Middleware

Companies have spent the last few years investing heavily in automation, AI tools, and connected workflows. 

In many cases, the gains are real. Teams are moving faster, tasks are taking less time, and more work is being handled through software.

But many employees are still spending their days doing work no system was supposed to require. 

They are checking outputs, moving information between platforms, correcting inconsistencies, and filling in the gaps when tools fail to communicate clearly with each other.

In a growing number of organizations, humans have become the layer holding disconnected systems together.

‍

What that coordination work actually looks like

The work is hard to see from above because it does not have its own line on the org chart. 

It shows up as small, recurring tasks that take a few minutes each and add up to most of someone's week. 

Inside a typical operation, it looks like this:

• A teammate keeps a private spreadsheet because the dashboard and the system of record have never quite agreed about reality
• A weekly standing meeting exists only to reconcile what two systems say happened
• A Slack thread has become the workflow for handling exceptions because the system of record cannot capture them
• The AI-drafted reply gets rewritten before it goes out because the context the model had was missing the two things the customer mentioned on last week's call
• Someone on the team is valuable mostly because they remember how three of your tools connect when nobody else does
• The same data gets entered twice because nobody trusts the sync between the CRM and the operations system

None of this looks like "work" in the way the team's job description describes it. It looks like keeping things from breaking, which is hard to put on a sprint board or a quarterly review. 

That is the coordination layer humans are absorbing, and it grows every time another tool gets added to the stack.

‍

The cost shows up in time, attention, and trust

Those scenes add up to more time than most leadership teams realize, and the research has started to put numbers on it.

Start with the time. 

Research from Asana found that the average knowledge worker switches between roughly ten different applications up to 25 times a day, and that task switching can make workers up to 40% less productive. 

The same research found that knowledge workers now spend about 60% of their time on what Asana calls "work about work": coordination, status chasing, tool switching, and the small reconciliations that keep workflows moving. The actual deliverables get the remaining 40%.

Then there is the attention cost. Gloria Mark's long-running research at the University of California, Irvine has found that after a single interruption, it takes an average of 23 minutes to get back to the level of focus that was lost. In a workday spent alt-tabbing between ten apps and answering Slack questions about what the dashboard actually means, the deep-work time disappears. 

The team is busy without producing as much as the calendar suggests they should.

The cost that is hardest to put on a slide is trust. 

When the systems do not agree with each other, people stop trusting any single one of them. 

Operations leaders start double-checking numbers before standups because being wrong in public has a cost, and finance ends up building parallel reports to explain discrepancies it did not create. 

At the top of the org chart, executives delay decisions, less because they lack conviction and more because the conviction is not matched by confidence in the data underneath it. Meetings get longer, and decisions get softer.

None of this appears on a productivity dashboard. It appears in how tired the senior people on the team look, and in how much of their week was spent making sure something did not go wrong instead of moving the operation forward.

‍

What happens when this scales to the org level

These dynamics scale, and they do not always stay invisible. A real version played out at Commonwealth Bank of Australia in 2025. 

The bank announced it was cutting 45 customer service roles, citing a new AI voice bot that the company said had reduced call volumes by 2,000 a week. 

On paper, the math was clean. On the ground, the union representing the affected workers took CBA to the workplace relations tribunal and produced a different picture: call volumes were actually rising, the bank had been paying staff overtime to handle the load, and team leaders had been pulled in to answer calls. 

The voice bot had not absorbed the work. It had moved the work somewhere management was not looking.

By August, CBA had reversed the decision and rehired the 45 workers, with the bank admitting in writing that the redundancies had been an "error."

What is worth holding onto about that story is how long the dashboard and the operation disagreed without anyone catching it. 

A 2,000-call reduction is the kind of number that goes into a board update and a press release. 

The reality was that the same calls were getting handled by humans on overtime, and the only reason it surfaced was a union grievance. 

The bank itself could not see it from inside its own reporting.

The scale at which AI is being deployed makes this kind of gap more common, not less. Across 2024, companies collectively spent more than $250 billion on AI, and 74% of them saw no measurable value from the investment. 

By mid-2025, 42% of companies had abandoned the majority of their AI initiatives, up from 17% the year before. A lot of that spend went into rollouts where the dashboard looked good while the operation underneath was absorbing the work the AI was supposed to do.

Most companies running this pattern do not end up in a tribunal hearing. 

They end up with the same dynamic at smaller scale. Overtime expenses no one can fully explain. 

The senior people on the team burning out faster than the org chart suggests they should, and the deliverables that should have shipped last sprint sliding into the next one.

‍

Why the burden grows instead of shrinks

The intuitive expectation is that as software gets smarter, the operational work goes down. The CBA example, and the macro pattern behind it, is showing the opposite.

Every layer of tooling an organization adds creates more outputs to verify, more data to reconcile, and more confusion about which version of the truth is current. 

The coordination layer expands with the tool count, and the people responsible for it keep absorbing more of the work.

What is supposed to be reducing labor is in fact moving labor: from the original task to the work of keeping the tools that handle the task in sync. 

That is the shape of operational work in 2026 in a lot of companies. The team has not stopped doing things by hand. 

The things being done by hand have shifted from the work itself to maintaining the appearance that the systems are handling it.

We believe that business is built on transparency and trust, and good software is built the same way. 

That principle gets tested between systems now, and it gets tested inside the people holding them together. 

When the tools cannot agree on the state of the operation, the cost lands on the team first. 

Eventually it lands on the customer. By the time it lands on the dashboard, it is already in the renewal numbers.

‍

What better software actually does

The way out of this is different software, built around a different premise. 

Better systems reduce the amount of coordination work the team has to do to make tasks actually happen, instead of just automating individual tasks on top of a stack that already does not agree with itself.

That means:

• Reducing the number of places someone has to check to answer one question
• Centralizing the definition of what a given record means, so the same number looks the same in every tool
• Carrying context across handoffs so the team in operations does not have to re-derive what the team in sales already knew
• Making it possible to trust the dashboard well enough to stop maintaining the shadow spreadsheet
• Lowering the verification tax on AI-generated outputs by giving the model better context to work from in the first place

This is the work we focus on at Big Pixel. 

The version of software that does this well does not feel like another platform added to the stack. It feels like the operation finally fits together, and the people inside it can spend their time on the work the team was actually hired to do.

That is also the version that holds up when AI capability changes. 

A workflow built around clean context can absorb a better model later. 

A workflow built around coordination work that humans are doing in the gaps cannot.

‍

Where the coordination tax is hiding in your operation

The most useful next step is small. Spend an hour with your team mapping where the coordination work actually lives.

A few questions that surface it quickly:

• Which standing meeting on the calendar exists only to reconcile what two systems are saying?
• Who on the team is valuable mostly because they remember how the tools connect?
• How much of your team's daily Slack volume is workflow that the system of record could not capture?
• For each AI output your team uses, how much time goes into verifying it before it goes anywhere?
• If your most senior operations person were out for a week, what part of the operation would stall?

If the answers point to more invisible labor than you would have guessed, the next big improvement is probably the work of reducing the coordination tax the team is already paying to keep the systems you own from contradicting each other. 

Adding more AI on top of that layer usually makes the tax higher, not lower.

The shift in question, from how much can we automate to how much coordination can we eliminate, is where some of the strongest operational gains in 2026 are coming from.

When you are ready to do that mapping seriously and want a second set of eyes, our free strategy session walks through exactly that exercise.

We help you find where the coordination tax is highest and what to do about it before the next AI investment gets approved.