TL;DR: When AI tools don’t connect to your systems, people fill the gap by copying, pasting, re-checking, and correcting. That’s transport work, and it’s invisible because it feels productive. Sort your team’s tasks into transport and judgment, count the transport minutes for two weeks, and you’ll have a number worth acting on.
There’s a specific kind of busy that has shown up in businesses over the last two years, and it looks exactly like progress.
Someone pulls information out of one system because another can’t reach it. A customer’s details get pasted into an AI tool to draft a reply, and the result gets pasted somewhere else. A report gets checked by hand because nobody trusts what came out automatically. Prompts get rewritten to give a tool context it should already have.
There’s a name for it: human middleware. Your people become the connective tissue holding disconnected systems together.
The awkward part is that everyone involved will tell you AI made them faster. They may be right. They may also be completely wrong, and there’s now decent evidence people can’t reliably tell the difference.
What is human middleware?
People moving information between systems so the software can do its job. Copying between apps, re-entering data that already exists elsewhere, verifying outputs by hand, and fixing results that were almost right. None of it is anybody’s actual job, and all of it fills real hours.
You’ll spot it once you know the shape. Any time a person is the reason data gets from A to B, and the only skill required is care and patience, that’s middleware.
It grows quietly because each instance is small and reasonable. One tool here, one automation there, a new AI feature in something you already pay for. Nobody decides to build a workflow held together by copy and paste. It assembles itself.
Why it feels productive while it’s happening
Here’s the finding worth sitting with. In 2025, the research group METR ran a randomized controlled trial with experienced developers working on their own real projects, half the tasks with AI tools allowed and half without. The developers took 19% longer with AI. Before starting, they’d predicted a 24% speedup. Afterward, having done the work, they still believed AI had made them about 20% faster.
Be careful how you use that. METR has since treated the speed result as historical, ran a follow-up with much wider error bars, and is redesigning the experiment. Tools have changed since early 2025 and the honest read is that nobody knows the current speed effect.
The other half of the finding is the durable one. The gap between what people believed and what the clock showed ran nearly forty percentage points, among skilled professionals doing familiar work. That’s the mechanism underneath busy days that feel productive.
So asking your team whether AI is helping gets you a sincere answer that may be unrelated to reality. You need something you can count.
Transport work and judgment work
Every task your team does with these tools falls into one of two buckets, and the difference is the whole point.
Judgment work is the part software can’t do because it doesn’t have the information. Which client is fragile right now. Whether this draft sounds like your firm. Whether the number in the report is plausible given what happened last month. Whether this customer needs a phone call instead of an email. Nobody outside your business knows any of that, and no tool will.
Transport work is moving data from one place to another. Software is excellent at this. It’s what computers were built for. When a person does it, that’s not value being added, it’s a plumbing failure wearing a human costume.
Both feel like work, which is why they never get separated. Both fill the day and produce a sense of having been busy, and only one is worth paying a person to do. Sorting them is the most useful hour a business owner can spend on AI right now.
How do you measure it in your own business?
Run a two-week tally. Open a shared document and ask everyone to log any task where they moved information between systems by hand, with a rough guess at the minutes. No precision required and no judgment attached, or people underreport.
Two weeks is enough to see the pattern. You’re looking for the same three or four tasks appearing across different people. Those are your candidates and they usually account for most of the total.
Then convert it. If four people each lose 30 minutes a day to the same handoff, that’s roughly 40 hours a month of transport work. Now you have a number to hold against the cost of fixing it instead of a vague sense that things feel harder than they should.
Most businesses are surprised twice: by the total, and by how concentrated it is in a couple of gaps nobody had named.
What do you do with the number?
Pick one of three answers for each recurring task: connect the systems, drop the tool, or accept the manual step and say so out loud. All three are legitimate. What isn’t legitimate is leaving it undecided, which is what’s happening now.
Not everything is worth integrating. A five-minute weekly copy-paste costs less than the integration that would remove it, and admitting that is a real decision rather than a failure. The point is that someone chose it.
MIT’s 2025 research into enterprise AI found the deployments producing measurable value were narrow ones embedded in a single workflow, not general-purpose assistants dropped into a business. The lead author’s summary was that the winners pick one pain point and execute well. Your two-week tally is how you find which pain point.
That’s the right sequence for AI strategy work generally. Find the expensive handoff first, then choose a tool. The other order is how businesses end up with four assistants and more admin than they started with, and it’s the same reason day-to-day IT support should be in the room when tools get bought.
Why copy and paste is also a security problem
Because data pasted into an outside tool has left your business and nobody logged it. Verizon’s 2026 breach research found regular AI use on company devices tripled in a year, two-thirds of it through personal accounts, with source code the most common thing pasted in.
The connection to middleware is direct. People paste into whatever tool is at hand because the proper path doesn’t exist. Fix the plumbing and much of that behavior stops on its own, because the workaround was never the point.
Meanwhile, two things help: an approved tool on a business account so there’s a sanctioned option, and a written line about what data can go into it. That belongs in the same conversation as your broader security review, not a separate project.
The takeaway
Your team shouldn’t spend the day helping software talk to other software. But you won’t fix it by asking whether AI is helping, because people are poor judges of their own speed with these tools.
Count instead. Two weeks, one shared document, rough minutes. Sort what comes back into transport and judgment, fix the two or three biggest transport tasks, and leave the judgment work with the people who have the context.
If you’d like help running that exercise, book a short call. It usually starts with a list you already have in your head.
Frequently Asked Questions
What is human middleware?
When employees spend time moving information between systems by hand so the software can function: copying between apps, re-entering existing data, checking outputs manually, correcting results that were nearly right. It’s work created by disconnected tools, not work the business needs.
Can AI make a team less productive?
It can. A 2025 randomized trial by METR found experienced developers took 19% longer on real tasks with AI tools, while believing they’d been 20% faster. The researchers have since revised the speed finding, but the gap between perception and measurement is the lasting lesson.
How do I measure whether AI is actually helping?
Track it for two weeks. Have your team log tasks where they manually move information between systems, with rough time estimates. Look for the same handoffs repeating across people. That gives you a number instead of an impression.
Should we integrate every manual step?
No. Some handoffs cost less than the integration that would remove them. The goal is to decide deliberately rather than drift, so each recurring manual task gets one of three answers: connect the systems, drop the tool, or accept the step knowingly.
Why do employees paste company data into AI tools?
Usually because no approved path exists for the task. The workaround is a symptom of missing integration, not carelessness. Providing a sanctioned business-account tool with clear rules about what data is allowed removes most of the incentive.
Find the Handoff That’s Costing You Most
Most businesses adopting AI end up with more tools and no fewer manual steps, because nobody mapped how information actually moves before buying. Z-JAK helps Louisville companies find the expensive handoffs first, then decide which are worth connecting, alongside the systems and security work underneath. Start a conversation and we’ll work through yours.