Our position is simple. AI should do the repetitive work, and the judgment calls should stay with people. So when an automation disappoints, look first at the task it was given, not at the tool. Knowing what not to automate means spotting the task that only looked repetitive, where someone was making small decisions all along.
Dropped AI projects are common enough to take the question seriously. In S&P Global Market Intelligence’s 2025 survey, 42% of companies said they had abandoned most of their AI initiatives before production. A year earlier, that figure was 17%. Respondents also said that, on average, 46% of their projects are scrapped between proof of concept and broad adoption.
About these numbers: S&P Global Market Intelligence surveyed 1,006 midlevel and senior IT and line-of-business professionals in North America and Europe. It published the results in May 2025. The figures are self-reported and cover AI projects of every kind, not business-operations automation alone. They also don’t say why projects were dropped.
For the why, RAND’s 2024 study is more useful. RAND interviewed 65 experienced data scientists and engineers, so this is the technical view rather than the business owner’s. Its first root cause: stakeholders often misunderstand or miscommunicate what problem needs to be solved using AI. In our view, picking a task that only looked rule-shaped is one version of that mistake.
Four signs a task is ready to automate
This half of the test is the familiar part, so it gets the least space. A task is a good candidate when all four of these hold.
- Repetitive. It happens the same way every time, like copying a form into your CRM.
- Rule-shaped. A new hire could do it from a one-page instruction. Routing an inquiry by service type fits.
- High-volume. It recurs often enough that the setup pays back. Status updates that arrive without chasing are a good example.
- Reversible. If it goes wrong, you can fix it before anyone outside notices.
Reversibility borrows from an older problem: the cost of undoing a wrong outcome. In his 2015 letter to Amazon shareholders, Jeff Bezos wrote that some decisions “are consequential and irreversible or nearly irreversible.” For those, “you can’t get back to where you were before.”
He was writing about management, not software, and he still wanted even reversible calls made “by high judgment individuals or small groups.” So we’re borrowing only his question about undoing. If an automated step’s mistakes can be undone, automate freely. If they can’t, keep a person on it.
Three signs a task should stay human
Each sign below can hide inside a task that passes all four checks. That’s why deciding what not to automate deserves most of your attention.
A wrong call lands on a person
Some mistakes cost you time. Others land on someone outside your business, and they find out before you do: a declined prospect, a client whose invoice gets flagged as disputed, a contractor told their work wasn’t accepted.
Handing a call to a system doesn’t hand off the consequences. In February 2024, British Columbia’s Civil Resolution Tribunal ruled against Air Canada after the chatbot on its website gave a customer wrong information about bereavement fares. In the tribunal’s words, “It makes no difference whether the information comes from a static page or a chatbot.” The airline had to pay CAD $650.88 in damages, plus interest and fees.
That is one small-claims decision in Canada, not a rule that governs your business. Automation also adds scale. The US Equal Employment Opportunity Commission alleged that iTutorGroup programmed its application software to automatically reject women aged 55 or older and men aged 60 or older. iTutorGroup agreed to pay $365,000 to settle the suit in 2023.
Once a rule runs automatically, a flaw in it reaches everyone it touches. And reversible for you isn’t the same as harmless for them: you can un-decline a lead in your CRM, but not in the memory of the person who got the email. So ask who finds out first when a step goes wrong. If it’s someone outside your business, keep a person on the call.
The rules are judgment in disguise
Plenty of written rules aren’t really rules. “Decline inquiries under our minimum budget” sounds rule-shaped until a small inquiry arrives from a returning client. Similarly, “follow up after a few days” works until a prospect on leave gets the nudge.
In those cases, the written rule summarizes someone’s judgment, and the exceptions are where the business lives. As a related caution, RAND relays one practitioner’s view that AI is poorly suited to automating an organization’s internal processes. That holds “especially when subjective human judgment is required to determine how those processes should function.”
To spot this, ask the person who does the task now how often they say “except.” If the answer is often, automate the steps around the decision and leave the decision alone.
The output is the relationship
Some outputs are artifacts: the weekly summary, the meeting notes, the invoice. Assembling them is exactly the repetitive work AI should take. Other outputs are the relationship itself, like the apology after a missed deadline or the note telling a long-standing client that prices are rising.
Klarna is a useful case because it went big. In February 2024, a month after launching it globally, Klarna said its AI assistant was “doing the equivalent work of 700 full-time agents.” Even so, by its September 2025 IPO prospectus, Klarna described a “dual-track approach” and said it continues to offer all of its customers the option of a human representative. So even a company leaning this hard on AI keeps a human option, the line we also draw in when a chatbot helps and when it hurts.
The test here is one question. Would the message lose its value if the recipient learned a system sent it? If yes, the sending is the point, so keep it human.
Where the honest answer is “partly”
Most real workflows don’t sort cleanly. The useful move is to cut each one at the step where judgment enters: automate everything before it, and keep that step human. Here is where the cut falls in five common workflows.
| Workflow | Automate | Keep human |
|---|---|---|
| Lead handling | Capture, deduplicate, route by type, acknowledge receipt, set a follow-up reminder | Deciding which leads to turn away, and how to say no |
| Client replies | Drafting a reply from the thread and project notes | Reading it, changing it, sending it |
| Weekly numbers | Pulling the figures, building the summary, flagging anything out of range | Saying what the numbers mean and what to do next |
| Scheduling | Offering slots, booking, reminders, rescheduling links | Deciding who gets priority when two clients need the same week |
| Invoicing | Generating invoices from agreed scope, sending payment reminders | Deciding whether to chase, waive or pause for a client in trouble |
Make sure the person actually decides
A person who clicks approve on every draft isn’t making the call. IBM’s Phaedra Boinodiris and Jamie Mackenzie put the goal plainly: people keep “meaningful oversight and decision-making authority, not just a rubber stamp at the end of an automated process.” If nobody really reads the draft, the decision has been automated after all. So agree in advance on what approval actually means for each step.
There’s also a slower cost to weigh. In our view, some repetitive work is how a newer team member learns to make the calls, which we cover in what disappears with junior creative jobs. Count that before you automate the learning away.
None of this argues for automating less by default. NIST’s voluntary AI Risk Management Framework, published in 2023, even notes that “some AI systems may not require human oversight.” The aim is to put people where the calls are and let automation have the rest.
How to decide what not to automate
Pick one workflow you’d like off your plate. For each step in it, ask these seven questions.
- Does it happen the same way every time?
- Could a new hire do it from a one-page instruction?
- Does it recur often enough to be worth setting up?
- If it goes wrong, can you fix it before anyone outside notices?
- Does the result land on a person outside your business, such as a lead, a client or a contractor?
- Does the person doing it now often say “except”?
- Would the output lose its value if the recipient knew a system sent it?
Automate the steps that get a yes on the first four and a no on the last three. Keep the rest with a person, and make sure that person actually decides.
Anyone selling automation will happily draft your list of things to automate. The list of decisions that stay with you is the one worth writing yourself.
Not sure where the line falls in your own workflows? The Blue Mango’s AI specialists can talk through how your week runs and which parts are worth automating. There’s no fixed package, just a conversation about your work. Book a time to talk it through.