You have more demand than your people can deliver, and the next hire is either more than you can pay or more than you want to add. The short answer: before any new role, split it into jobs, prove what AI can and can't do on each one, hand the jobs over in stages, and hire only for the hours that are left. By the end you will have a six-step order from your capacity sums to your next hire, the question to answer in writing before any new role, and a plan for the people you already have.
Six steps decide what AI takes on and what your next hire does
One owner of a three-person business wrote: "We think we can get to 12m without adding headcount." Our method starts from that kind of goal: keep the people you have, grow what they deliver, and treat a new role as the last step rather than the first. The six steps run in order, and each one is usable on its own.
1. Do the capacity sums before you write a job ad
Do this. Write down the hours each customer or order takes today, the cap on customers those hours allow, the number you want to reach, and how many hours per customer you would have to remove to get there.
It worked when you have one page with today's hours per customer, your target and the gap in hours, before anyone talks about a hire.
Why. Our rule of thumb is that a hire is the answer to a sum, so the sum comes first. In our view there are four ways out when demand outruns what you can deliver: charge more, serve more customers per hour of delivery, turn the service into a product, or hire. Each one turns you into a different kind of business, so pick on purpose. In our view AI belongs in the second way: it can change the hours each customer takes, which is the number a hire was going to change. Look first at the part of the service customers enjoy after buying but did not buy for; we treat it as the first candidate to cut, shrink or hand to software.
2. Split the role into jobs, not a job title
Do this. Take the role you were about to hire for, list every repeated job in it, and mark each job one of three ways: AI could do or draft it, a person must do it, or nobody needs to do it.
It worked when every job on the list has one mark, and the hours behind the marks add up to the hours the role would have filled.
Why. In our view a role is a bundle of tasks: some go to software, some stay with a person, and the hire, if any, is for what is left. Our rule of thumb is that a job earns AI only if it repeats, follows the same rules each time, and gives back more time than it takes to set up. Count the jobs nobody does today too: in our view an undone job can cost more than a wage.
Example. Lütke's memo asks: "What would this area look like if autonomous AI agents were already part of the team?" 1. Microsoft proposes a "human-agent ratio" and two questions for leaders: "How many agents are needed for which roles and tasks? And how many humans are needed to guide them?" 2. We read both as the same move: start from the work, then decide the people.
Keep with a person
Rare and high stakes: not worth setting up AI.AI drafts, a person approves
Worth automating, but a person signs off.Drop it or do it by hand
Not worth the setup time.Hand to AI
Repeats, same rules, low risk.3. Fix the job by hand before AI touches it
Do this. For each job marked for AI, write down the steps a person follows today and run them by hand on your next few real orders until the result is the one you want.
It worked when someone other than you can follow the written steps and get the same result.
Why. In our view AI speeds up a process that works; it does not replace having one. One person put it this way: "They bolt AI onto a broken hiring decision and wonder why it does not work. If you hired the wrong rep, AI just makes them fail faster." Compare fairly too: give the AI step the same written steps and training a person would get before you judge it against a process people have tuned for years.
Example. In an experiment with 758 Boston Consulting Group consultants, those using AI on 18 tasks inside AI's abilities completed 12.2% more tasks and finished them 25.1% more quickly on average, with significantly better quality 3. On a task chosen to sit outside AI's abilities, consultants using AI were less likely to produce correct solutions than those without it 3. We read that as the reason to test each job, not the role as a whole: the same tool can help on one job and hurt on the next.
4. Answer in writing why AI can't do the job before you add a role
Do this. Before any new role, write one page that answers two questions: what would this area look like if AI already did part of the work, and which jobs did AI try and fall short on?
It worked when the page names the jobs AI was tried on, what it got wrong, and the hours a person still needs.
Why. Shopify chief executive Tobi Lütke told employees that before asking for more headcount or resources, teams must show why they "cannot get what they want done using AI" 1. Von Ahn's email said "headcount will only be given if a team cannot automate more of their work" 4. In our view the written answer is the point: it turns a feeling that you need help into a list of jobs AI was tried on.
Example. One person described the moment: "When I really thought about it, something I was about to try to hire people to do I realized I should probably just learn how to automate." Our rule of thumb is to treat that moment, just before the job ad, as the time to run this step.
5. Hand jobs over in stages and keep a person on the risky cases
Do this. For the first 30 orders let AI draft and a person approve, widen what it does alone only as its record holds, and send unusual or risky cases straight to a named person.
It worked when every job handed over has a named person who checks it, and you can say how many of the last 30 outputs went out without edits.
Why. Our rule of thumb is to give AI more freedom only as its record grows, then let it run on a schedule. Send routine work straight through and route edge cases to a person, so speed doesn't cost you a customer.
Example. Klarna paused hiring for over 12 months while it focused on AI; its headcount fell 22% to 3,500, mostly through attrition, and its chief executive asked staff to use AI to fill the gaps left by departing colleagues 5. In May 2025 Klarna's chief executive Sebastian Siemiatkowski told Bloomberg it was hiring human customer service staff again, saying AI chatbots were cheaper but resulted in "lower quality" output 5. In our view, keeping a person on the risky cases costs less than rehiring once quality slips.
6. Hire for the hours that are left, and recheck every role that falls vacant
Do this. After 30 orders, count the hours per customer again; when someone leaves, run steps 2 to 5 on their role before you replace them, then hire only for the hours still left.
It worked when your next hire, or your decision not to hire, rests on the hours left after the recount, not on the old total.
Why. MIT NANDA's 2025 report found workforce disruption already under way in customer support and administrative roles, and that "Rather than mass layoffs, companies are increasingly not backfilling positions as they become vacant" 6. The same report found the changes concentrated in jobs previously outsourced because of their perceived low value 6. We read that as the quiet form of this method: the decision comes at the vacancy, not in a round of cuts.
Example. One person with a small team put it this way: "I have a small team with AI on the side. Boosted production for everyone. Don’t need to hire a lot of humans anymore."
We think AI lifts capacity job by job, and only on jobs inside what it does well
Across 5,172 customer support agents, a generative AI assistant raised issues resolved per hour by 15% on average and by 30% for less skilled, less experienced workers; the most skilled and experienced saw small speed gains and small quality declines 7. The same study found that AI assistance "facilitates worker learning", customers became "more polite", and attrition fell as newer workers stayed 7. The authors write that the AI tool "helps newer agents to move more quickly down the experience curve" 7. In an MIT study of 453 college-educated professionals given occupation-specific writing tasks, access to ChatGPT cut the time taken by 40% and raised output quality, as rated by independent evaluators, by 18% 8. In the same study, workers who scored lower on the first task benefited more from ChatGPT on the second 8. We read those gains as hours taken out of each customer or task, which is the capacity a new hire was meant to add.
MIT News noted that the tasks did not require precise factual accuracy or context about a company's goals or customers, and co-author Shakked Noy said fact-checking and writing prompts take time in real work 8. In our view that is why steps 3 and 5 exist: written steps and a person on the risky cases keep the checking time from eating the gain.
Demand outgrows the team, so the owner writes a job ad for a whole role.
The owner splits the role into jobs, proves what AI can do on each, and hires only for the hours that are left.
In the Census Bureau's 2023 Annual Business Survey, businesses that adopted AI between 2020 and 2022 most often reported that their "number of workers did not change overall" 9. When these technologies did change worker numbers, all except robotics were more likely to increase them than decrease them 9. We read that as a sign that adopting AI and keeping your headcount steady can go together.
US Census Bureau survey data from December 2025 to May 2026 show overall business AI use between 17% and 20%, and less than 20% among firms with four or fewer employees 10. Between December 2025 and May 2026, AI use rose among firms with at least 20 employees but did not change significantly among firms with fewer than 20 10. In our view that gap matters for a business of a few people, where one hire is a large share of the payroll.
Shopify and Duolingo wrote the rule down; Klarna went back to hiring people for quality
Lütke's memo calls "reflexive AI usage" a "baseline expectation" at Shopify and adds AI usage questions to its performance and peer review questionnaire 1. In our view the review questions matter as much as the hiring rule: they make the habit part of the job.
In April 2025 Duolingo chief executive Luis von Ahn told staff the company would "gradually stop using contractors to do work that AI can handle" 4. Von Ahn wrote that "this isn't about replacing Duos with AI" and that the changes are "about removing bottlenecks" 4. Von Ahn wrote that replacing a slow, manual content process with one powered by AI was one of Duolingo's best recent decisions, and that "Without AI, it would take us decades to scale our content to more learners" 4.
After the backlash, von Ahn told the New York Times: "This was on me. I did not give enough context" and "We've never laid off any full-time employees. We don't plan to" 11. Von Ahn said Duolingo's "contractor force has gone up and down depending on needs" 11. We read that as a lesson for any owner: say who the rule is about before you announce it.
Siemiatkowski said: "Really, investing in the quality of human support is the way of the future for us" 5. In our view Klarna's path is a reason to keep step 5's human route open while the jobs are handed over.
Amazon's chief executive Andy Jassy wrote in June 2025: "We will need fewer people doing some of the jobs that are being done today, and more people doing other types of jobs" 12. Jassy added that in the next few years Amazon expects AI to "reduce our total corporate workforce as we get efficiency gains from using AI extensively across the company" 12. Salesforce chief executive Marc Benioff, whose company offers Agentforce and other AI-powered tools, said in 2025: "I've reduced it from 9,000 heads to about 5,000, because I need less heads", speaking of customer support 13. We read these large firms' statements as being about cutting totals; this method is about growing what the people you have can deliver.
This can fail when AI becomes the work, or when a job sits outside what AI does well
I fell in this trap of outsourcing my thinking and it set me back months in my business. It got to the point where using ai to do the work BECAME the work.one owner
In our view the warning sign is when setting up and checking AI takes more of your week than the job it replaced. When that happens, stop and go back to step 3: fix the job by hand, then hand it over again.
Gartner predicted in June 2025 that over 40% of agentic AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls 14. In our view that is a reason to hand over one job at a time and keep the person who did it until the record holds.
A fair question one person asked: "Curious though, do you think small teams can realistically map all their workflows, or is that only doable once you hit a certain scale?" Our answer: you don't need to map every workflow. Map the role you were about to hire for, job by job, as in step 2; that is one role, not the whole business.
One company president put it this way: "Reduced labor costs equal additional profits. The trick though is to maintain the culture and experience of our retail operation while still adapting to change." In our view that trade-off is real: keep the jobs customers feel with a person, and hand the rest to AI first.
Put the hours AI frees into the people you already have
Von Ahn expects that "one person will be able to accomplish more, rather than having fewer people" 11. In our view the support and writing studies above point the same way: give AI to your newest and least practised people first, not only to your best.
One business consultant put it this way: "Hey, as a business consultant for the last 15 years, I can tell you there is probably a way to re-purpose (some or all of) those workers to take on more work or opportunities that AI generates. It may not always be so, OR some employees may be incapable of learning/pivoting, but making your business better often leads to more." One owner asked: "Where can I find employees who CAN use AI for my small businesses?" In our view the hire doesn't disappear; it changes. When you do hire, hire for the hours left and for someone who will run the AI steps, not only do the work by hand.
Within seven days, count the hours behind three jobs that grew with sales
Open the last 90 days of orders and pick the three jobs that grew when sales grew. Then run the checklist below, starting with the first job, for the next 30 orders.
before your next hire
You started with a hire you couldn't afford; now you have a way to decide it
You now have a six-step order from your capacity sums to your next hire. You have the question to answer in writing before any new role: what would this area look like if AI already did part of the work, and which jobs did AI try and fall short on? And you have a plan for the people you already have: give them the AI steps first, and hire only for the hours that are left.
Read before: How to Replace Yourself in Your Business (Before You Can Afford a Team). Read next: Are the AI Agents in Your Business Useful, or Just Busy?, to judge the jobs you handed over.
Where does your business sit on AI and capacity? Take the Scorecard.
Sources
- Shopify CEO says no new hires without proof AI can't do the job · The Verge, 2025-04-07
- The 2025 Annual Work Trend Index: The Frontier Firm is born · Microsoft, 2025-04-23
- Navigating the Jagged Technological Frontier · Harvard Business School Working Paper 24-013, 2023-09
- Duolingo will replace contract workers with AI · The Verge, 2025-04-28
- Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It on Calls, According to the Company's CEO · Entrepreneur, 2025-05-09
- MIT report: 95% of generative AI pilots at companies are failing · Fortune, 2025-08-18
- Generative AI at Work · The Quarterly Journal of Economics, 2025
- Study finds ChatGPT boosts worker productivity for some writing tasks · MIT News, 2023-07-14
- How AI and Other Technology Impacted Businesses and Workers · U.S. Census Bureau, 2025-09-17
- Large Firms With at Least 20 Employees Biggest AI Users · U.S. Census Bureau, 2026-05-26
- Duolingo CEO admits his controversial AI memo 'did not give enough context' and insists the company never laid off full-time employees · Fortune, 2025-08-18
- Message from CEO Andy Jassy: Some thoughts on Generative AI · Amazon, 2025-06-17
- Salesforce cuts 4,000 jobs due to AI, CEO says · Fox Business, 2025-09-03
- Over 40 percent of Agentic AI projects could face the axe by end of 2027: Gartner · W.Media, 2025-06-27
