AI Strategy & Transformation7 min read·By , Independent AI Advisor & Fractional AI Lead·

Companies keep automating work that should no longer exist

When automation gets cheap, the question of whether a step should exist stops being forced. A deletion test for the work you are about to automate.

In the July–August 1990 issue of Harvard Business Review, Michael Hammer published an article titled "Reengineering Work: Don't Automate, Obliterate." His complaint was that companies had spent heavily on information technology to mechanize processes built for an earlier world of paper and control, and that speeding those processes up could not repair what was wrong with them.

His best-known case was Ford's accounts payable department. Ford had planned to cut about 20% of its more than 500 clerks. Then it compared itself with Mazda, which ran the same function with five people. That comparison pushed Ford to question the process rather than polish it. Under what Hammer calls invoiceless processing, the data items that had to match fell from 14 to 3, and where the new process was installed, Hammer reports a 75% reduction in headcount. The lesson was not that Ford needed better software. The matching work existed because of a rule about when to pay, and the rule could change.

Thirty-six years later the same mistake is cheaper to make. Prototyping an agent against an existing inbox or ticket queue takes far less effort than building integration software once did. That is my judgment, not a measured figure.

Cheap automation removes a question that cost used to force

Here is the argument I want to make, and it is my interpretation, not a finding from the studies below. Automation used to need a business case. To write a business case, someone had to describe what the process was for. Weak steps often failed at that stage, because the description was embarrassing: nobody could say who used the output.

When a prototype is quick to build, the description is easy to skip. The demo becomes the argument. And once a step is automated, it acquires a dashboard, an owner and a new reason to persist: it works. Automation can turn an unexamined habit into infrastructure.

The wider evidence points the same way without proving this mechanism. Stop adopting AI already covers McKinsey's 2026 survey (fielded May 4 to June 8, 2026, 1,719 respondents), so I will only add one cut of it: nearly three-quarters of the roughly 6% of respondents McKinsey classes as AI high performers said they had fundamentally redesigned workflows because of AI, against about one-quarter of everyone else. Two cautions. The answers are self-reported, and high performers are defined by outcome, so this is an association, not proof that redesign caused the results. It also does not say what was removed. Neither this survey nor Hammer's 1990 cases gives a number you should apply to your backlog.

That earlier essay argues that the should be the workflow. This one is narrower: what to ask about a single backlog item before anyone approves money to automate it.

Four reasons a step exists

Before asking whether a model can do a step, I would ask why the step exists. The answers I would expect fall into four groups. This is a working classification of my own, not a validated taxonomy.

Why the step existsQuestion to askUsual verdict
It produces an outcome someone needs: a payment, a decision, a delivered productIs the outcome still needed, and is this the cheapest route to it?Candidate for automation
It compensates for a defect elsewhere: re-keying, reconciling two systems, chasing missing dataCan the source be fixed?Repair upstream, then reassess
It produces evidence for a control that nobody reads: reports, approvals, logsWho uses the evidence, and what happens if it stops?Test the control; retire it or replace it with sampling
It carries information between people: status updates, handoffs, relaysDoes the information still need to move, or should the decision sit with someone who already has it?Redesign the decision route

The categories are illustrative, not case studies. The point is that only the first row is a clean case for automation. In the other three, automation makes the underlying defect cheaper to keep. A reconciliation agent does not fix the two systems that disagree. A report-preparation agent cannot tell you whether anyone acts on the report.

The safest answer your AI team can give you may be "we should not automate this" lists removing the work as the first option. This essay is only about how to decide whether that option applies, and it does not repeat that ladder.

The deletion test

Take the item at the top of your automation backlog and put four questions to the team that runs it, in this order.

  1. Who consumes the output, and what do they do differently because of it? A name, not a department. If nobody can be named, label the step an orphan.
  2. If the step stopped for a short, fixed period, what would break, and who would notice first? Write the prediction down before anything changes.
  3. What upstream defect or rule created the step, and does it still apply? Hammer's invoice-matching step traced back to a payment rule. Yours will trace back to something too.
  4. Which control or obligation requires it? Name the policy, the auditor or the regulation. If nobody can, ask the owner of the control rather than assuming.

Then, where the step is reversible and outside regulated, safety-critical or irreversible work, suspend or shrink it for a bounded period with a named person responsible for rollback. Compare what happened with the prediction.

Three outcomes are useful. If nobody noticed, retire the step and do not build the agent. If someone noticed, find the decision that changed and automate toward that outcome, which may be a much smaller job than automating the original step. If the step turned out to be a patch for an upstream defect, send the work to whoever owns the source.

The decision I would ask an executive to take is narrow: do not approve automation spend for a step that cannot answer questions one and two. A step that fails those questions has not been shown to be needed, so the automation case cannot be better than the case for the step.

What would make me wrong

Some steps look orphaned and are not. A report nobody reads may be the only record that surfaces a rare failure. The prediction written in advance and the bounded suspension are meant to catch this, but rare events can stay hidden during a short test.

Some steps exist because a regulator or auditor requires them. In that case the step is a fixed obligation, and automating it may be the right move. The test still helps, because it makes someone name the obligation and check whether it requires the step as currently performed.

Building the automation can itself reveal waste. That is fair, but the description is cheaper to do first, with a whiteboard and the people who run the work.

Speed has value. I would accept a deliberate, time-boxed decision to automate now and revisit later. I would resist the unrecorded version, where the revisit never happens.

Before the next approval

Hammer's title was a provocation in 1990. The tools are far more capable now, and the question has not changed: is the thing we are speeding up still worth doing? Ask it before the build, not in the post-launch review. If the team cannot say who uses the output and what changes without it, the right first deliverable is not an agent. It is a decision about whether the step should exist.

Where this connects

For the portfolio-level argument, see Stop adopting AI. For whether a use case has a business case at all, see The 10-minute test for whether an AI use case has a business case.

Sources and evidence notes

The state of AI in 2026: On the road to ROI

McKinsey & Company · 2026-08-25

Reengineering Work: Don't Automate, Obliterate (reprint hosted by NTNU)

Harvard Business Review, Michael Hammer, July–August 1990 · 1990-07-01

  • McKinsey: self-reported global survey (fielded 2026-05-04 to 06-08), association between and outcome-defined high performance, not causal evidence.
  • Hammer: first-hand case examples from a 1990 article; they illustrate the argument and are not a current estimate.

The four-reason classification, the mechanism about cost gates and the deletion test are Berk Bayri's argument, not findings attributed to either source.

Get new essays as they publish.

One email per essay. No noise between.

Berk Bayri

Independent AI Advisor & Fractional AI Lead

I help leadership teams decide where AI belongs, test it before it scales and build the teams that run it, drawing on 25+ years of building digital products, experiences and capabilities.