- Published
- Updated
- 10 min read

TL;DR: Forrester found 55% of employers regret laying off staff because of AI, and most of those cuts came before any working AI system existed. The lesson for a small business is about what to automate: pick repetitive tasks with clear rules and a human path for exceptions, not whole roles. Here is how I would choose.
Updated October 2026: this note now cites each figure to its source. We removed numbers from the earlier version that we could not trace to a primary source, corrected which firm made which prediction, and removed details about our own team that no longer belong here.
Deciding what to automate is the whole game. The companies now regretting their AI layoffs mostly did not get the technology wrong. They pointed it at the wrong target: a headcount number instead of a piece of work. I study behavioral psychology, and this pattern is familiar. When a decision is framed as a cost line, people optimize the cost line and lose sight of what the cost was buying.
What does the research say about AI layoffs?
Forrester's analysis of the 2026 future of work put a number on the regret. "According to the analysis, 55% of employers now regret laying off staff because of AI." (Computerworld (opens in a new tab)) The same report expects many of those cuts to unwind: "half of AI-attributed layoffs are likely to be reversed," often with the work handed "to lower-wage human workers, offshore or at lower salary." (The Register (opens in a new tab))
Gartner made a narrower prediction about one function. It "predicted that half of companies that reduced customer service headcount due to AI will rehire staff by 2027 to perform similar functions." (CX Dive (opens in a new tab)) Two research firms, two different scopes, and the same direction.
Most cuts came before the AI did
The most telling line came from Forrester's follow-up forecast in January 2026. Asked whether companies planning AI layoffs had a working system ready, the answer was blunt: "When we ask if they have a mature, vetted AI app ready to fill in those jobs, nine out of ten times, the answer is no." (The Register (opens in a new tab))
Forrester's own release calls this "AI washing," the habit of "attributing financially motivated cuts to future AI implementation." The same forecast says AI could account for 6% of US job losses by 2030, while it expects AI to "augment 20% of jobs over the next five years." (Forrester (opens in a new tab)) In other words, the research expects far more jobs to change shape than to disappear.
The people buying AI expect more staff, not fewer
One figure from the earlier Forrester report gets less attention than it should. "More people in charge of AI investment expect it to increase headcount (57 percent) than to decrease it (15 percent) over the next year." (The Register (opens in a new tab)) The people closest to the tools are the least convinced that the tools replace people.
A large survey of executives points the same way. The National Bureau of Economic Research summarized responses "from nearly 6,000 CEOs, CFOs, and senior finance managers," and found that "More than 90 percent of executives report no effect of AI use on employment over the past three years." (NBER (opens in a new tab))
Why did replacing people with AI backfire?
The best known example is Klarna. In February 2024 the company said its AI assistant "has had 2.3 million conversations, two-thirds of Klarna's customer service chats," and was "doing the equivalent work of 700 full-time agents." (Klarna (opens in a new tab))
About a year later the company changed course. It wanted customers to "always have the option to speak with a human," and its CEO told Bloomberg: "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality." (CX Dive (opens in a new tab))
Notice what did not happen. The chatbot was not switched off. It kept handling routine chats. What changed was the assumption that it could carry the whole job, including the conversations that need judgment. That is a decision about what to automate, and the first version drew the line in the wrong place.
The pattern behind the regret
Across these reports I see the same three mistakes about what to automate:
- The target was a role, not a task. A job is a bundle of tasks. Some are routine, some carry judgment, some hold relationships. Cutting the role removes all of them at once.
- The system did not exist yet. Nine out of ten planned cuts had no vetted AI ready, by Forrester's count. The savings were booked before the capability was built.
- Cost was the only scorecard. Klarna's CEO said it directly. When cost is the main measure, quality loses, and quality is what customers notice.
What does the evidence say AI is good at?
The strongest evidence I know of points to augmentation. A study of 5,179 customer support agents found that access to a generative AI assistant "increases productivity, as measured by issues resolved per hour, by 14% on average, including a 34% improvement for novice and low-skilled workers." (NBER (opens in a new tab)) The paper was later published in the Quarterly Journal of Economics.
Read that slowly. The biggest gains went to the newest people. The tool spread what experienced agents already knew. That is a reason to keep a team and make it better, not a reason to shrink it.
Where the AI projects that pay off focus
MIT's research on enterprise AI, as reported by Fortune, found that "about 5% of AI pilot programs achieve rapid revenue acceleration; the vast majority stall." The problem it named was not model quality but a "learning gap" for tools and organizations, and it found "the biggest ROI in back-office automation." (Fortune (opens in a new tab)) One researcher put the winning approach simply: successful teams "pick one pain point, execute well, and partner smartly."
Pick one pain point. That is the most practical sentence in the whole debate about what to automate.
How should a small business decide what to automate?
Small businesses are already in the middle of this. The U.S. Chamber of Commerce reports that "58% of small businesses self-identified they use generative AI," and that "82% of small businesses using AI increased their workforce over the past year." (U.S. Chamber of Commerce (opens in a new tab)) The Census Bureau, which asks a narrower question about using AI in a business function, puts overall use between 17% and 20%, and lower for the smallest firms. (U.S. Census Bureau (opens in a new tab)) Either way, most owners are still early, which is the best time to choose well.
Here is the order I would use to decide what to automate.
Step 1: List tasks, not jobs
Take one role and write down everything that person does in a normal week. Mark each task as routine, judgment, or relationship. Routine tasks follow rules most of the time: data entry, scheduling, first drafts, sorting the inbox, pulling a weekly report. Judgment and relationship tasks are where your business actually earns trust. Only the routine column is a candidate for what to automate.
Step 2: Score the routine tasks
To narrow down what to automate, ask four questions about each routine task:
- How often does it happen? Daily work pays back faster than quarterly work.
- How clear are the rules? If two people would do it the same way, a system can too.
- What does a mistake cost? A wrong meeting time is cheap. A wrong invoice to a client is not.
- Can you tell a good result from a bad one quickly? If you cannot check it, you cannot automate it safely.
The best candidates are frequent, rule based, cheap to get wrong once, and easy to check. That short list is what to automate first.
Step 3: Build the human path before launch
Every automation needs a place to send what it is unsure about. Decide who gets the exceptions, how fast they need an answer, and what the customer sees while they wait. This is what AI quality gates are for: a check on every output, and a person for the cases that fail the check.
Step 4: Measure before and after
Time the task for two weeks before you change anything. Then measure the same thing after. If the hours saved do not show up, or the error rate climbs, you have learned something cheaply. Time saved and time spent checking the output are both real, which we wrote about in our note on why AI is not saving teams the time they expected.
Step 5: Reinvest the time on purpose
Part of choosing what to automate is deciding, in advance, what your people will do with the hours they get back. Follow up with old customers. Fix the onboarding process. Write the proposal you keep putting off. Time that is not assigned tends to disappear into more email.
What should you not automate?
Knowing what to automate is half the decision. The other half is what to leave with people:
- Conversations with an upset customer. Klarna's reversal started here.
- Anything where being technically correct and being right are different answers. Refund exceptions, pricing for a long time client, a scheduling conflict with a family emergency behind it.
- Final approval on anything that leaves the building. A draft can be automated. Sending it to a client should still have a person who said yes.
- Work nobody has written down yet. If the process lives only in someone's head, automate after you document it, not before. We cover that problem in our note on labor hoarding and undocumented knowledge.
I wrote a longer walkthrough on how to figure out where AI actually fits in your business (opens in a new tab), and Brian Gagne, Kief Studio's CTO, has written a practical guide to AI for small business owners (opens in a new tab) that covers what to buy and what staff may use.
Is a small team better placed than a large company?
In one way, yes. A small business can see the whole workflow on one page. There are fewer layers between the person doing the work and the person deciding what to automate, so the feedback loop is short. When an automation goes wrong, you hear about it the same day.
The risk runs the other way too. A small team has less slack. If one person's routine work moves to a system that fails quietly, there may be nobody else who knows how to do it by hand. That is why the human path in Step 3 matters more for a team of eight than for a company of eight thousand. Be honest about that when you decide what to automate.
The research above points to the same habit for both: start with one painful, repetitive task, prove the result, and keep your people in charge of the judgment. AI implementation done this way is slower to announce and much cheaper to undo.
Frequently asked questions
Do companies regret replacing workers with AI?
Many do. Forrester found 55% of employers regret laying off staff because of AI and expects half of AI-attributed layoffs to be reversed, often at lower pay or offshore.
Why did AI layoffs backfire?
Most cuts came before a working system existed. Forrester says nine out of ten companies planning AI layoffs had no mature, vetted AI ready. Cost was the main measure, and quality suffered. They chose what to automate by headcount, not by task.
How do I decide what to automate first?
List one role's tasks, mark the routine ones, and score them on frequency, clear rules, cost of a mistake and how easy the result is to check. Start with the task that scores best on all four.
What should a small business never automate?
Upset customers, exceptions that need judgment, final approval on anything sent to a client, and any process nobody has written down yet. Deciding what to automate starts with deciding what stays human.
Does AI make employees more productive?
In customer support, a study of 5,179 agents found a 14% productivity gain on average and 34% for newer and less experienced workers. The gains came from helping people, not replacing them.
Are small businesses cutting staff because of AI?
Most are not. The U.S. Chamber of Commerce reports that 82% of small businesses using AI increased their workforce over the past year.
Where we fit
Kief Studio builds AI automation that removes bottlenecks, not people. We pick the workflow with you, define what good output looks like and measure how it runs today, then build it with retries, error handling and quality gates on every output. Anything the model is unsure about goes to a person. We run it for you on LTFI, so nothing is installed on your side, and LTFI's Automate layer (opens in a new tab) is built so nothing goes out until a person approves it.
If you are not sure yet what to automate, our technology advisory work starts with what you run today, what it costs in time, and where it holds you back, then writes the order of steps in plain language.
First conversation is free. No commitment.




