The human harness: why your AI spend went up and your value did not

Almost everyone has AI now. McKinsey's State of AI 2025 puts regular use at 88% of organisations, and 39% attribute any profit impact to it at all, most of them under 5%. S&P Global found the share of companies abandoning most of their AI initiatives rose from 17% to 42% in a year. In the Netherlands, the Exact MKB Barometer 2026 puts AI use at 70% of SMEs, up from 6% two years earlier, and 47% of the 1,700 owners and managers surveyed agree with the sentence "we do not really know how to use AI".

The usual explanations are all partly right: the wrong use case, no training, bad data, pilots that never left the pilot. Underneath several of them sits a mechanism that nobody in the building can see, because they are standing inside it.

The muted call

A monthly update. The director asks how the renewal pipeline looks. The account manager mutes, types the question into the assistant, unmutes, and reads the answer out. The assistant has a follow-up about one customer. The account manager asks Bob. Bob mutes.

It is a joke that does the rounds, and it lands because the room recognises it. Everyone on the call is working. Nobody on the call is thinking.

Who is the harness in your office

A language model produces text. That is all it does. To turn the text into something that happens, a file opened, an email sent, a record updated, it needs a layer around it that executes. People who build agents call that layer the harness.

Enterprise IT has known the human version of this for decades: the "swivel chair", the person who reads from one screen and types into another. In Workday's 2026 survey of 6,100 professionals, analyst Ray Wang calls them "human middleware", and the survey finds that 82% spend significant time moving data between tools while only 27% have AI connected to the workflows where the work happens.

When the assistant has no harness, the person is the harness. The assistant asks for the document; the person fetches it from the drive. It drafts the reply; the person pastes it into the mail client and presses send. It needs last quarter's figures; the person logs into the finance system and copies them out. The hands moved to the person. The thinking moved to the model. That is the opposite of what the licence was bought for.

The SOP that came back as a to-do list

The pattern is at its clearest in companies that did the work properly. Someone wrote a good standard operating procedure: how an order is checked, how a quote goes out, how a support case is closed. It was shipped to the assistant as a skill, so everyone runs the same procedure.

Then an employee asks the assistant to run it. The assistant reads the procedure back as a list. Fetch the order from the system. Check it against the price list. Put the confirmation on the company letterhead. Open the customer record. Copy the link. Send the email. Fifty steps, each one for a person to perform by hand, in the order the model chose.

Nothing was automated. The spend went up because the licence is per seat and everyone has one, and the value stayed flat because the work still takes the same hands. The author of the procedure did the deciding, the model did the sequencing, and the employee did every step.

The vendors now sell connections that close part of this gap. Some are off by default, some are billed on top of the seat, and which ones you get depends on licence tier and region. In the mid-sized offices we walk into, the accounting package, the sector ERP and the ticketing tool are not on the list, and most seats are still unconnected to the systems where the work lives.

What leaks

Three things go missing while people are the hands.

Judgement. The person who relays an answer did not form it. A 2025 study by Microsoft Research and Carnegie Mellon of 319 knowledge workers found that the more confident people were in the AI, the less critical-thinking effort they reported putting in. The sales call is the same finding with a mute button.

Accountability. Three weeks later something went wrong with that order. Who decided? The employee executed the steps. The procedure's author wrote them a year ago. The model picked the sequence. The audit trail says Marieke, and Marieke cannot tell you why.

Senior time. The people you pay for judgement are spending it on transfer. Workday finds the copy and paste burden shows up most clearly in managers and directors, and an eight-month field study at a 200-person firm by Ranganathan and Ye, published in Harvard Business Review this year, found AI intensified work rather than reducing it. More tasks, more switching, same day.

Two tests to run this week

Shadow one person for one day. Pick someone good at their job. Count every time something crosses between them and the assistant by hand: a file in, a paragraph out, a login, a yes. Each one is a missing connection. Multiply by headcount and you have the size of your human harness.

Ask why. The next time someone gives an answer in a meeting, ask them why. If the reply is "that is what the AI said", you have found where the judgement went.

A procedure delivers value when the agent can finish it

Can the agent complete step one without a person? If not, the procedure is a description, not an automation, and nothing after step one matters.

Ad hoc work belongs in the chat window. Claude, ChatGPT and Copilot are good at the one-off analysis and the question nobody will ask twice. The moment a procedure repeats, daily, weekly, monthly, it belongs in your own harness, with the connections to your systems attached, so the agent fetches, formats, files and sends on its own. The same procedure, moved, not rewritten. Where the agent lacks access, it asks for it and a person grants it; the team harness post describes that loop. We build this at DevOptix, and the harness is the part most companies skip, because the chat window is what they bought and the connections are what nobody priced.

Three levels of approval, and the bottleneck moves

Not every action needs a human yes. Sort each action in the procedure into one of three levels.

Level 1: the agent acts. The action is bounded and a deterministic rule can check it: the order total is under the limit, the customer is in good standing, the ticket is in the right state. A model can advise on these checks. It never overrides them. A mechanical no beats a model yes, every time.

Level 2: the person who asked approves. The agent brings the decision with its context: "I am running the quote procedure for customer Y. The request was Z. I need your yes to send the quote at this price." One click, with everything in front of you.

Level 3: a manager signs. Four eyes, the way those actions were always signed.

The rule is to push every action to the lowest level the risk allows. Every action at level 1 is capacity you did not have last month. Every action left at level 3 is the old queue outside one manager's door, and fifty employees generating level 3 requests will fill it before the week is out. Level 2 has its own failure: at volume, people stop reading and start clicking. Watch for it.

What the leader owns

Not logins. Not prompts per user. Two numbers.

The number of your systems the agent can reach without a person in between. The number of its actions with a named person who decided.

And one early warning: the override rate at each approval gate, and how long an approval waits. A gate nobody ever overrides is a gate nobody reads.

Give the agent its hands. Your people get their heads back.