Human Middleware: The cost of making AI usable, and how long it will be paid

The first draft is now cheap. The second draft is the cost.

Human Middleware. This term names the cognitive labour that sits around the model, the act of transforming plausible output into usable output.

AI strategies currently depend on an informal layer of quality assurance that hasn't been factored into capacity planning or talent development frameworks. This shows up as employees that are willing to take on the relentless, frustrating Human Middleware work without any changes to their contract, job description or salary.

AI strategies currently depend on an informal layer of quality assurance that hasn't been factored into capacity planning or talent development frameworks.

More than writing and interpreting, Human Middleware includes reading the system before the words even arrive. Response time becomes a clue, not a measurement. A fast answer on a hard task suggests a shortcut was taken. A slow answer on an easy task suggests the model misread the request. A slow answer on a hard task could mean either deep processing or confusion, and you cannot tell which unless you follow the reasoning feed in real time to see what was happening.

Increasingly, using LLMs properly already means using more than one model, holding multiple mental models, each with different strengths and tendencies. The systems themselves change without warning, so those mental models need constant recalibration. This interpretation layer is expensive to build and expensive to maintain. It does not compress or stabilise because the ground keeps shifting.

If it stays unnamed, it will stay unaccounted for.

It cannot stay that way.

Could Human Middleware be a transition phase?

Some of the burden is learning how to drive the tool. The other burden is making its output fit a specific reality with consequences, where omissions and phrasing can become commitments.

It resembles supervision, but it behaves differently. Unlike correcting a person, correcting a model typically doesn't create lasting change in the base system, often the same correction is required each time.

Models may reduce some of the burden as they become easier to steer and better at holding context. But that is not the same as the burden of the conversion work diminishing.

The interpretation work that surrounds a single document can exceed the time the model took to produce it. A thirty-second draft becomes a thirty-minute negotiation with the output.

Two forces push against a reduction.

Many cloud model providers update systems without user control and the user's working assumptions need re-testing because the boundary of what the system will do has moved without an internal decision.

The other is the pull toward generic plausibility. A model can become more capable and still default to language that sounds reasonable in the abstract while missing what is locally true, locally risky, or locally unacceptable.

Models are improving and context windows have been expanding, but neither of these updates changes the structural pull toward generic plausibility.

Organisations do not need what reads well, they need what reads right.

Opting out

There is a group who are not engaging because they have decided the tax is not worth paying. They can already produce quality work without it and they perceive that their level of expertise means the tool does not offer enough to justify the friction.

This is a bet, not a permanent position. They are betting they can continue delivering without engaging deeply, at least for now. Some will discover the environment shifted faster than their timeline. But for now, they are not carrying Human Middleware because they have opted out entirely.

This creates a governance blind spot where strategic assumptions about capability distribution may not match operational reality. The cost of this decision falls on the board. It shows up as uneven adoption, bottlenecks around the trusted reviewers, and a two-speed organisation where the AI story looks strong in slides but patchy in work.

What people are calculating

Employees facing this reality likely face a calculation.

If this is temporary, if the interpretation layer eventually thins to nothing, then the work that currently justifies their role disappears with it. They are valuable now because AI needs them. Remove that need, and the question of what remains becomes uncomfortable.

If this is not temporary, if the conversion work persists because drift and generic gravity never go away, then this is not a transition they are enduring. This is the shape of work for the rest of their careers, where the frustration does not ease and the calibration does not end. The grind continues until work stops or they do.

People adjust behaviour to what the job rewards and drains, whether anyone names it or not.

The biggest threat to AI adoption may not be hallucinations or data security. It may be annoyance.

The cost beneath the cost

Organisations are inadvertently absorbing this work into existing workflows without recognising it as a distinct capability requiring investment, much like how 'data hygiene' was once invisible until it became a competitive differentiator.

The durability of AI adoption may ultimately depend less on technical reliability than on the sustainability of the human oversight layer, particularly the cognitive load it places on high performers whose discretionary effort determines execution quality.

In general, the promise to employees was that AI would enable them to operate at a higher level of abstraction, with judgement and strategy rising as machines handled execution. That shift requires people to move beyond interpretation into leverage. If the interpretation layer keeps absorbing their capacity, the shift stalls, perhaps indefinitely, and that has significant consequences for your AI strategy and for the people propping up the illusion of machines that can think.