Where automation stalls, and why

Almost every firm we speak to uses AI. Far fewer can say what it changed. The owner knows the team has ChatGPT or Copilot, suspects it helps, and cannot point to a number that moved.

That gap is why we use a ladder. It gives each routine in a business a place, names the next step, and shows why so many attempts to automate stall on the way there.

The ladder in brief

The autonomy ladder has five levels. Each one is defined by who does the work, who decides, and who presses the button that changes the system of record.

• Level 0, Manual. A person does it, every time.
• Level 1, Assisted. People use AI to work faster. The process is unchanged.
• Level 2, Connected. AI works inside your systems. A person approves each run.
• Level 3, Delegated. It runs. People handle what it flags.
• Level 4, Autonomous. It runs itself. People watch the outcomes.

A level belongs to a routine, never to a company. A firm with enquiries at level 2, reporting at level 1 and contracts at level 0 is normal.

The ladder is a business translation of published research on AI autonomy, by Feng, McDonald and Zhang in 2025 and by Morris and colleagues at Google DeepMind in 2023. Both find that more autonomy is not a better system, and that the right level depends on the task. The full ladder is on the Levels page.

Where most firms sit

Most of the work in most firms runs at level 0 or level 1. Level 1 is the crowded rung because it is the easiest to buy. A licence, a login, and the team starts drafting emails and summaries with an assistant.

It helps the person. It leaves the process where it was. Someone still reads the output, fixes it and types it into the CRM, the ledger or the job system. Nothing is logged, so nothing can be counted.

MIT's Project NANDA saw the same pattern at scale. Its 2025 report, The GenAI Divide, drew on a review of 300 public AI initiatives and interviews at 52 organisations. It found that over 80 per cent of organisations had explored or piloted tools like ChatGPT and Copilot, and that those tools lift individual productivity, not the profit and loss.

A licence moves a routine from level 0 to level 1. It does not move it to level 2.

The first real change happens at level 2. The software reads from and writes to the systems of record, a person approves each run, and every run leaves a record behind. That is the move most firms have not made yet, and it is where most attempts go wrong.

Why the move stalls

The failures we see are rarely about the model. They come from how the move was chosen, built and looked after. Five patterns account for most of them.

The tool is mistaken for the change. A firm buys seats, runs a lunch-and-learn, and counts itself done. A year later the spend is visible and the return is not, because a level 1 routine produces no record of what it saved. When the budget review comes, nobody can defend it.

The problem is never pinned down. RAND interviewed 65 experienced data scientists and engineers about AI projects that failed, and published five root causes in 2024. The first is that the people involved misunderstand, or miscommunicate, the problem the AI is meant to solve. Another is chasing the newest technology instead of the real problem of the people who will use it. In a smaller firm this sounds like "let's do something with agents" in place of "our quotes take three days".

Rungs get skipped. A routine that nobody has written down, and that has never had an approval step, is handed straight to an agent that runs unattended. There is no history of what normal looks like, no exception rate, and nobody watching. Gartner, an analyst firm, forecast in June 2025 that over 40 per cent of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls.

The approval becomes a rubber stamp. Level 2 is only as safe as its approval step. When a person is shown a whole document fifty times a day, they stop reading and start clicking. Research on automation bias by Parasuraman and Manzey found that it affects experts as well as novices, and that training alone does not remove it.

Nobody owns it after launch. A vendor changes its API and a connector breaks. The mix of work shifts and exceptions creep up. The person who looked after the queue moves on. Without an owner, the routine slides back down a level without anyone deciding it should, and the hours quietly return to where they were.

How to keep it moving

Each failure has a habit that prevents it. None of them is exotic, and all of them cost less than a second attempt.

Place one routine, not the company. Pick the routine where the hours are and ask one question: who presses the button that changes the system of record? If a person types or pastes the result in, it is at level 1, however clever the assistant. Which routine to automate first shows how to choose between candidates.

Count before you build. Write down the hours a week at the step where most of the time goes, and pick one measure at the function it changes. Time to first reply, days to invoice, quotes out the same day. Without a before, there is no after.

Move one level, usually from 1 to 2. Level 2 is where the hours stop being a story and become a count. It needs the same four things every time: a connection to the system of record, an approval step, a log of every read and write, and a way to undo each write. Never plan more than two levels above where a routine sits today.

Design the approval to be read. Show the approver what changed, not the whole record. Keep batches small. Pull a few runs a week for a full review. That is what turns "a person approves it" from a comfort into a control.

Earn level 3 with evidence. A routine is ready to run on its own after about ninety days at level 2 with a low, steady exception rate. Level 3 then adds an exception queue with a named person behind it, a written list of what the system may not decide, an alert when exceptions drift, and a stop switch the owner can use without calling anyone.

Give it an owner and a calendar. One named person, a monthly look at the run log, and a check every quarter on whether the routine still sits where it should. Test it again whenever the model or the connected software changes. If exceptions rise, move the routine down a level, not up. Why automations slide back covers the warning signs.

Decide where to stop. The further a mistake can travel before someone sees it, the earlier a person needs to be in the loop. Most routines belong at level 2 or 3. Level 4 suits a narrow class of work where errors are cheap, reversible and caught at once.

Four questions for this week

You can place your own routines without us. Take the one that costs the most time and ask:

1. Who presses the button that changes the record?
2. How many hours a week does the slowest step take, and how do you know?
3. Could you show last month's runs, with the approvals and the errors?
4. If it broke tomorrow, who would notice first, and who would fix it?

If the answers are "a person", "a guess", "no" and "not sure", the routine is at level 1, and the move to level 2 is where the return is. The free assessment places your routines on the ladder and names the next step for each.