The Human Is the Bottleneck: Why AI Works Slowly in Companies Even Though the Model Answers in Seconds

The model answers in seconds. The task it answers for still takes days. In practice, managing directors regularly describe this mismatch: the demo was impressive, but everyday work feels sluggish. The obvious explanation is that the tool is not good enough yet. That falls short. The time is not lost in the model. It is lost between the windows.

Three scenes from everyday work

A clerk receives a customer complaint. She copies the text into the chat window, fetches the customer number from the CRM, types in the contract status by hand, reads the suggestion, copies it back into the email client and adjusts the salutation. The model was done after seconds. The whole process took a quarter of an hour, and the clerk was busy for every minute of it.

Second scene. A team wants support tickets to be sorted automatically in advance. But the AI is not allowed into the ticket system. So someone exports a list every day, uploads it and copies the results back. The request for read access has been sitting with IT for weeks, because nobody knows which role an AI should be assigned to. During this time, the AI is not waiting for data. The task is waiting for a decision.

Third scene. For a quotation, line items are to be extracted from an inquiry, matched against the price list, written into the template and submitted for review. Every step is its own approval: ten times reading, ten times clicking, for a task the model could complete in one go. All three scenes come from consulting practice. They are examples from individual projects, not measurements.

The waiting point is the human

What the scenes have in common is the place where the work stands still. That is the first thesis: every copy, every tab switch, every approval is a point where the system waits for a human. The human is not idle in this, quite the opposite. He is constantly busy, but with transport instead of judgment. The picture is a powerful engine in a vehicle with the handbrake on. You can keep tuning the engine, but you will not get faster.

The latency a managing director feels is therefore not model latency. It is handoff latency. It arises in the seams between email client, CRM, ticket system and chat window, and it grows with every place where someone has to carry something from A to B.

What a harness is

A term helps here: the harness. It means the environment in which an AI can carry a task through to completion. It consists of context sources (knowledge, tickets, customer data), permissions (what the AI may read and change), tools (what it can trigger, such as creating a record) and feedback channels through which results and questions flow back to people.

Open standards for the connection now exist. The Model Context Protocol describes how an AI application accesses tools and data sources, but expressly does not define how it handles them. The standard solves the plug, not the organization behind it. And more connections are not automatically better. Anthropic describes a model's context window, meaning the amount of text it processes at once, as a finite resource with diminishing marginal returns and looks for the smallest set of high-signal information. The source is about the model, not about the number of connected systems. It still works as an analogy, and this is an assessment from practice: a harness that connects everything available creates noise. A good harness brings a few suitable sources into the workflow.

The capability is there, the environment is missing

The second thesis is less comfortable. Models can now complete longer tasks on their own than they could a short while ago. METR measures how long tasks can be that a model solves successfully with 50 percent probability, and estimates the doubling time of this task length since 2023 at around 131 days, with an uncertainty range of 107 to 161 days. These are benchmark tasks, not workflows in mid-sized businesses. The figure cannot be transferred to companies, but it shows the direction: the distance a model can cover on its own is growing fast. In the three scenes above, a different pace prevails. Permissions, connections and approval rules there still date from the time when AI was a chat window. Anyone who operates a capability designed for longer stretches in small bites, with a human before every mouthful, is not using it to the full.

Why the harness gets left behind

Honesty is in order at this point, without blame. The harness is invisible work, the wiring behind the wall: nobody admires it, but without it no lamp lights up. A connection to the ticket system, a permission model, a well-maintained knowledge directory impress nobody in a demo. A new tool or a new prompt impresses immediately. Where a budget is tied to what can be shown, infrastructure has poor odds, and the next tool gets bought while the connection is left lying around. This is not personal failure but an incentive pattern that operates in projects and budgets. It can only be broken by naming it.

Control that isn't

The third thesis concerns the feeling that keeps the intermediate steps alive: control. Anyone who reads every draft, approves every intermediate output and copies every result onward himself has a hand on the process. In fact he controls little. He checks things he cannot judge without context, and he checks things that need no checking because they have no consequences.

An analysis by Anthropic of Claude Code and the public API shows that new users of Claude Code approve all actions automatically in around 20 percent of sessions, and in over 40 percent after 750 sessions. This setting is called auto-approve: the tool no longer asks at every step. The same analysis finds that in around 73 percent of tool calls via the API a human is involved in some form, and only 0.8 percent of actions appear to be irreversible. The data basis is one provider's developer tool and interface, not mid-sized businesses, and the figures come from the provider itself. Still, the figures suggest two things. First: with growing experience, users relax approval, even if the majority keeps it. Permanent approval at the intermediate step does not seem to be a stable end state but a beginner phase. Second: if only a small fraction of actions cannot be undone, one approval per manual step is out of proportion to the risk. This is an interpretation; the study does not say how often a human actually clicks at every step. Anthropic itself concludes that effective oversight of agents requires new forms of post-deployment monitoring and new forms of human-AI interaction that help both sides manage autonomy and risk together.

Where an approval belongs

Real control lies in the harness: in defined rules, limited permissions, complete logs and feedback channels through which it is possible to trace what happened. Human approvals belong at the points where something can no longer be undone: money leaves the company, a statement goes out, a legal commitment is made. In its guide to building agents, Anthropic describes that these can pause at checkpoints or when blocked for human feedback. A checkpoint is a planned place, not a reflex before every step. That is where an approval belongs. One, not ten.

Three steps for the coming weeks

Find the breaks. Have someone count for one week how often a person copies, looks up or approves something for the AI. Check questions: How many systems does a typical task touch? Where does the AI wait for a human, and for how long?

Build the harness. Connect the context sources that appeared most often in the count: knowledge, tickets, CRM. Assign permissions to the AI as you would to a new employee, with a role, limits and a log. Provide tools and feedback channels. Check questions: Does the AI today have the same access as the person who operates it? Who decides on permissions, and how long does that take?

Move the approvals. Away from the intermediate step, toward the result and the few irreversible points. Check questions: Which approval truly prevents harm, and which merely reassures? What happens if the AI runs a task all the way to the proposal and a human decides only at the end?

The question shifts

In the end, the question a managing director asks about his AI adoption changes. No longer: How fast does the model answer? But: How many waiting points has the organization built around the model, and who is actually waiting there for whom? The human remains responsible, that is not up for debate. But responsibility shows in decisions with consequences, not in every click. An organization that understands this takes the human out of the role of messenger and gives back the role he was hired for: to judge where it matters.