When Not to Automate: How to Tell Which Work to Leave Alone

The automation roadmap has forty items on it. Each one came from a workshop where someone asked, “What could AI do here?” Each one has an owner and a quarter.

Nobody asked the other question. Which of these should we not automate at all?

Six months later, three items have shipped. Two are being quietly rolled back. One is in front of the complaints team every week. And the people who used to do that work have moved on, so going back is harder than going forward.

The paradox: The list of what AI could do was easy to write. The list of what it should not do was never written, and that is the list that would have saved the most.

What We Found

Before we recommend automating anything, we study what the public record says about how such decisions go. For this article we went through an official audit report, decisions by a court and a tribunal on AI-generated answers, and analyst predictions about AI projects, from 2023 to 2025.

Four findings stand out.

“Could be automated” is often counted before “should be”. In March 2024 the UK’s National Audit Office reported that an indicative analysis had found almost a third of tasks in the civil service could be automated. The same report records that the analysis did not examine the feasibility of delivering those gains or make an assessment of cost, and that further scrutiny and evidence collection were recognised as needed. A count of what could be automated is a starting point, not a decision.

Handing a task to AI does not hand over the responsibility. In February 2024 a Canadian tribunal, in Moffatt v. Air Canada, rejected an airline’s argument that its website chatbot was responsible for its own answers. “It makes no difference whether the information comes from a static page or a chatbot,” the tribunal held. The airline was ordered to pay CA$812.02 in total. When a step that states policy is automated, the business still owns every answer it gives.

Some checks belong to a person by duty, not by habit. In June 2023 a US federal court in New York, in Mata v. Avianca, sanctioned lawyers who had filed “non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT”. The court reminded lawyers of their “gatekeeping role” to ensure the accuracy of their filings. A tool can help with the work. It cannot take over the duty to check it.

Many AI projects are predicted not to survive their own costs and risks. One analyst firm predicted in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, and in 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. Some of those projects, our reading suggests, were work that never needed automating.

What our review does not show: None of these sources sets out which work should be left alone. The audit report is about UK central government. The two legal decisions concern a chatbot and a drafting tool, not a decision system. The analyst figures are predictions, not measured outcomes; the first one’s date has passed and we did not examine whether it came true. The signs of leave-alone work below are our reasoning from this record and from how workflows behave, not findings from a study.

The List Nobody Writes

Our reading of the record is that most automation programmes write one list: what AI could do. The second list, what it should not do, is the one that protects the business, and it is rarely written down.

Leaving work alone is not a failure of ambition. It is a decision, and it deserves the same care as the decision to automate. These are the signs we look for that a step is better left with people, or with simple rules, than given to a model:

• Experts disagree on the right answer. If your own people cannot agree what is correct, no system can be measured against it. That is a policy question to settle first.
• The decision carries a duty a person must hold. Professional sign-offs, regulated judgements and hardship cases belong to someone accountable, as the court in the lawyers’ case made plain.
• An error cannot be undone. Where a wrong answer ends a relationship, moves money that cannot come back or harms someone, the bar for automation is very high.
• Volume is low and the work is varied. A step that happens a few times a week, differently each time, rarely repays the cost of automating and testing it.
• The rules change faster than they can be tested. If the policy moves weekly, any automation will lag it.
• The cost of one case does not fall. If review, rework and errors cost as much as the work saved, the step is cheaper left as it is.

The gap we see: many firms have a roadmap of what to automate and no record of what they chose not to. Firms that write both lists, step by step and with reasons, spend their automation budget where it pays and keep people where people matter. They can also explain both choices when a customer, an auditor or their own board asks.

This is our reading of the evidence. It is not a measured market figure.

Four Ways to Handle a Candidate Step, Weighed

For each step on an automation roadmap, a firm has four real options. Each one is right for somebody.

Option	When it fits	What it risks
Automate it fully	High volume, stable rules, errors are cheap and reversible	A wrong answer repeats at scale before anyone notices
Automate it with a person checking every output	The model is useful but answers are acted on	Checking becomes a rubber stamp; cost per case may not fall
Automate part of it (a model reads or drafts, people or approved rules decide)	The step mixes messy input with a decision that carries money or policy	Needs the split designed and tested, step by step
Leave it alone, and improve the process around it	Experts disagree, a duty sits with a person, errors cannot be undone, or volume is low	The opportunity cost of not automating, which should be stated, not assumed

Our recommendation: decide step by step, and write the reason down either way. Leave a step alone when any of the signs above applies, and say so in the plan. Where part of a step can use a model, keep the commitment (the price, the approval, the policy answer) with people or rules your team has approved.

Where You Are Now, and Where You Should Be

A recommendation is only useful if you can tell later whether it worked. So before any work begins we set a benchmark: a few measures taken from your own workflow, each with a target.

For this topic, the benchmark would look like this.

Measure	Typical starting point	Where it should be	How it is checked
A written reason for each step	Reasons only for what is automated	Every step has a decision (automate, part, leave alone) and a reason	The step-by-step triage
Steps where a duty sits with a person	Not listed	Each one listed, with the person or role who holds it	A map of the workflow
Steps where experts disagree	Found after go-live	Found and settled, or left alone, before any build	Expert review of real cases
Cost per case for each automated step	Not measured	Measured today and after, so steps that do not pay are stopped	Finance’s cost data
Rolled-back automation	Discovered by complaints	Rare, because leave-alone steps were never started	The change log

The starting points above are what we expect to find. In an engagement they are replaced by your own figures, taken in the first review. The targets are what we would work towards with you. They are not results we have already achieved for a client.

How We Work With You

1. We listen: a free Reliability Review. A first conversation about the workflow, what is on the automation list, and what is already causing trouble. You leave with a short written note of what we found.
2. We research. We research your sector, the regulator’s position and what the public record says about firms in your situation. This article is an example of that work.
3. We diagnose: the Workflow Diagnostic. It sorts each step of the workflow into compile, keep on an AI model, or don’t automate, and sets out the cost per case today against projected, a 100-rerun consistency test of your current pilot, a policy-to-rules map, the evidence gaps and a change plan. It may recommend another firm or tool, or not automating at all. We take no referral fees on model or platform recommendations.
4. If you approve, we prove it: a Proof Sprint. One workflow, on your own cases, measured for consistency and for accuracy against your own experts. You keep the code.
5. We review against the benchmark. Before work starts we set the benchmark. At each stage we show you where each measure stands, and the filled-in benchmark is yours at the end.

The Review is free. Every paid step is fixed price, agreed at scoping. Stop after any step. The findings are yours. The Diagnostic fee is credited against a Proof Sprint within 90 days.

What changes for the business:

Role	What they gain
COO	Automation goes where it reduces the workload, and the work that needs people stays with people, by decision rather than by accident.
CEO	An automation plan the board can trust because it also says what the firm chose not to automate, and why.
CTO or CIO	Engineering time goes to steps that pay, and fewer projects are built only to be rolled back.
Head of AI or Data	A programme judged on the steps it chose well, including the ones it left alone.

The cases in this article, and what IOanyT does about that kind of problem:

The case	What IOanyT does
The National Audit Office noted that tasks were counted as automatable before feasibility and cost were examined.	The Workflow Diagnostic sorts each step into compile, keep on an AI model, or don’t automate, with the reason and the cost per case.
In Moffatt v. Air Canada, the business was held responsible for its chatbot’s answer.	Commitments such as prices, approvals and policy answers stay with people or rules the client has approved.
In Mata v. Avianca, the court stressed the lawyers’ own gatekeeping role.	We list the steps where a duty sits with a person, and keep them there.
An analyst firm predicted many AI projects would be abandoned or cancelled over cost, value and risk.	We will recommend not automating, or another firm or tool, when that is better, and we take no referral fees.

IOanyT is a consultancy with its own engineering team: we study the problem, recommend, and prove it by building only what you approve. Some of our most useful advice is a list of things not to build.

Book a free Reliability Review →

Get a fixed quote →

See the Workflow Diagnostic →