Skip to content
Chat with AI Agent
Business & governance

What an AI-First Company Changes in Its Team: Roles That Shift and Headcount Rules That Backfire

Orange ITS — AI engineering team 10 min read

When the owner of a 60-person company reads that Shopify makes teams justify every new hire against AI, the first reaction is usually arithmetic. If a competitor can run on fewer people, shouldn’t we? A board member or an investor often asks the same thing within a week.

The memos that started this conversation are real, and so are the consequences some companies have seen after copying the cost logic and skipping the work design. In 2025 two well-known firms announced AI-first rules and took public criticism, and one of them apologised for how the message landed. A third reversed a customer-service bet that it had advertised loudly.

An AI-first company gains from changed roles and a different team shape, and the headcount rules that treat AI as a reason to cut tend to backfire on quality and trust. This article shows what the memos said, what the employment data supports, which jobs are changing, and a simple headcount rule that fits a Swiss SME. If you want a person who builds and owns agent workflows inside your team, our forward-deployed AI engineering service is built for that.

The role question has two siblings. For the software side, see from developer to AI product manager. For the operating model behind this article, see AI-native companies.


What the AI-First Company Memos Actually Said

Most of the debate runs on paraphrase, so it helps to read the claims closely.

Shopify, memo of March 2025, made public in April 2025. CEO Tobi Lütke told staff that reflexive AI usage is a baseline expectation. According to BetaKit’s coverage, teams must demonstrate why they cannot get what they want done with AI before asking for more headcount and resources, and AI usage questions would be added to performance and peer reviews. The press headline compresses this into “prove AI can’t do the job”. The memo asks for a demonstration, and the demonstration is the interesting part.

Duolingo, April 2025. CEO Luis von Ahn announced an AI-first approach in which the company would gradually stop using contractors for work that AI can handle, as quoted by Entrepreneur. Public backlash followed. In August 2025 he told the New York Times that “This was on me. I did not give enough context”, and Fortune reported that he said Duolingo had not cut full-time staff and had no plans to. That statement comes from the company.

Neither company has published a study linking its rule to results. A memo shows what leadership intends. It does not show that the rule works.

The “Prove AI Cannot Do It” Gate: What It Gets Right and Where It Fails

The gate has a real strength. A team that has to write down why an agent cannot handle a task must first describe the task, its inputs, its exceptions and its quality bar. Many requests for new hires turn out to describe a process nobody has mapped. That mapping is where process redesign starts, and it often finds waste before any model is involved.

The risks are equally concrete:

  • The test becomes theatre. A team runs a lazy trial with a chat assistant, it fails, and the hire is approved. Or the opposite: nobody dares to ask for people at all.
  • Quality loss stays invisible for months. Reduced capacity shows up in complaint rates and churn long after the headcount decision.
  • The best people read the memo as a threat. Skilled employees who can find work elsewhere may leave first.
  • Contractors absorb the pressure. Cutting external capacity is easy and quiet, but the know-how leaves with them.

No study isolates the effect of such a gate on productivity, quality or retention, so any company claiming proof either way is guessing. Treat the gate as a prompt for a structured review with a named owner and a written result, and do not treat it as a rule that decides alone.

What Klarna and Duolingo Show About Cost-First AI

Klarna is the most cited example. Its press release of 27 February 2024 said the AI assistant handled 2.3 million conversations in its first month and did the equivalent work of 700 full-time agents. That is a company-reported figure, and no independent audit exists. In May 2025, Entrepreneur reported, citing a Bloomberg interview, that CEO Sebastian Siemiatkowski described investing in the quality of human support as the way of the future for Klarna.

Two details matter for an SME. Klarna’s headcount fell by about 22% to roughly 3,500, mostly through attrition rather than announced layoffs. And it did not claim to have replaced all human service. It began recruiting human agents again and kept an option to speak with a person.

Duolingo shows a different failure. The product direction may be sound, but the memo’s tone hit customers and employees at the same time, and the CEO publicly conceded he had supplied too little context. The lesson is about communication and sequencing. Announcing a headcount rule before you can show colleagues what their work becomes invites exactly the reaction you were trying to avoid.

Both cases suggest the same reading. Cost is a legitimate outcome of automation. Cost used as the starting question, ahead of quality and customer experience, tends to produce results that have to be corrected in public.

What the Employment Data Says So Far

The evidence is thinner than the confident headlines suggest, and it points in more than one direction.

  • Firm-level survey. NBER Working Paper 34836 surveyed nearly 9,000 executives in seven countries. Respondents reported essentially no net employment effect over the past three years and expect little over the next three. They report AI raising labour-productivity growth by about 0.2% a year so far. These are self-reported views rather than measured outcomes.
  • Young workers in exposed jobs. Stanford’s “Canaries in the Coal Mine”, using payroll data, found a 13% relative employment decline for workers aged 22 to 25 in the most AI-exposed occupations, concentrated where AI is more likely to automate than assist.
  • A softer reading. Anthropic’s labour-market study found no systematic rise in unemployment among exposed workers, and only suggestive evidence of slower hiring of young workers. Its authors work for an AI vendor, which is worth remembering.
  • Layoffs attributed to AI. Fortune reported that Oxford Economics read US layoff data as attributing about 4.5% of reported job losses to AI in the first eleven months of 2025, and that it suspects some firms dress up layoffs as an AI story when other causes such as over-hiring are behind them. That figure covers one country and one period.
  • Block. Fortune reported that Block cut about 4,000 employees and that Jack Dorsey linked the move to smaller, flatter teams while also citing over-hiring during Covid. How much was AI is contested.

The honest summary is that the net effect on total employment is small so far. The early-career signal is the one to watch, and it should influence how you build entry-level roles in your own firm.

Roles That Shift Inside an AI-First Company

The shifts below are directional. No neutral dataset tracks job titles inside firms, so treat them as patterns to test rather than forecasts.

Role todayWhere the work movesWhat the person owns
DeveloperFrom writing first drafts to reviewing and orchestrating agent outputSpecifications, review, tests, and the decision to ship
Support agentFrom routine answers to exceptions, escalations and judgement callsHard cases, tone, and the feedback that improves the agent
Operations or admin staffFrom executing steps to owning a workflow end to endThe workflow’s accuracy, cost and exceptions
New: harness or forward-deployed ownerBuilding and maintaining the agent system around the modelContext, tools, permissions, monitoring and fixes
New: evaluation leadChecking whether agent output meets the standardTest sets, review rules and quality reports

Anthropic’s March 2026 Economic Index reports that coding tasks keep migrating from assisted use toward more automated workflows. That is usage data from one vendor’s product, so it supports the developer row as a direction and says nothing about job counts. The detail of that shift is in our piece on developers becoming AI product managers.

For support and operations, the pattern is easier to picture than to measure. A support agent who used to answer forty routine tickets a day now reviews the five the agent could not close and writes down why. That is a different job, with a different skill mix, and it needs different training.

Two roles are new enough to be worth naming. The first is the person who owns the harness around the model. The second is the person who owns evaluation. In a small firm both can sit with one individual, and the combination is close to what the market calls a forward deployed engineer. Posting counts for that title have risen steeply from a small base, but the reported growth figures conflict, so do not read a precise number into them.

A Headcount Rule for an AI-First Company of 20 to 200 People

Instead of a gate that asks teams to prove a negative, use a rule built around evidence, owners and a time limit. This is a suggestion from our practice and has not been tested in a published study.

  1. Ask for a written workflow test before any hire. The requesting team documents the task, runs an agent against a sample of real cases, and records accuracy, time and exceptions.
  2. Name an owner for the result. One person is responsible for the agent’s output, whether the answer is yes or no.
  3. Approve the hire when the test shows a gap. A failed test is information. A hire that follows a documented gap is easier to defend than one made on instinct.
  4. Redeploy before you reduce. If an agent absorbs work, move people to the exception and review roles first. Reduce only through normal turnover, and give it a quarter before judging.
  5. Protect the entry-level pipeline. If Stanford’s early-career signal holds, firms that stop hiring juniors lose the people who become next year’s reviewers. Keep a small number of junior roles built around reviewing agent output.
  6. Track quality alongside cost. Complaint rate, rework, and time to resolution belong on the same page as savings.

An illustrative case: a 40-person distributor automates order-confirmation emails and cuts the work of two part-time administrators by half. The sensible move is to give those two people the exceptions queue and supplier follow-up that nobody had time for, and to leave both positions in place. The savings show up as capacity.

Hiring, Buying or Embedding AI Skills in Switzerland

Swiss SMEs mostly expect efficiency gains and few expect smaller teams. In AXA and Sotomo’s 2026 SME labour-market study, with 336 SMEs in German- and French-speaking Switzerland, only one in nine expected job cuts due to AI. Among SMEs with AI experience, 12% expected to need fewer staff and 47% expected time savings without a change in headcount. AXA is an insurer with an interest in the SME market, so read it as a sentiment indicator.

On pay, the SwissICT salary study 2026, as reported by Netzwoche, gives an overall ICT median of CHF 130,200 and a senior-expert median of CHF 177,938, based on 37,112 salaries from 214 companies. These are salaries for ICT roles broadly rather than for AI engineers specifically, and the report notes scarcity in areas including AI, cloud, cybersecurity and data. Employer cost is higher than base pay.

Your options are roughly three:

  • Hire. Fits when AI work is continuous and you can attract and keep a senior person. The hardest part is finding someone who has run agents in production and has done more than prototype them.
  • Buy. A platform or vendor fits well-defined, standard tasks. The trade-off is less control over context, permissions and evaluation.
  • Embed. An engineer working inside your systems for a defined period fits when you need the first workflows built and a team that can run them afterwards. The figures are in what an embedded AI engineer costs.

One legal point deserves a call before an announcement. Swiss law (Art. 335d ff. of the Code of Obligations) sets consultation duties for mass dismissals, and the thresholds depend on establishment size and the number of dismissals within a 30-day window. Commonly cited figures start at ten dismissals for establishments with more than 20 and fewer than 100 employees. Check the current text and your contracts with an employment lawyer, because we cannot give legal advice and the details matter.

Where This Leaves the AI-First Company

The memos agree on one thing. AI use is now a management expectation, and managers will be asked to justify capacity requests against it. The evidence agrees on another. So far, AI has changed what people do more than how many people firms employ, with the sharpest signal sitting among the youngest workers in the most exposed jobs.

For an SME this suggests a plain order of operations. Redesign a process, move people into the roles that the redesign creates, measure quality next to cost, and let headcount follow from the result. Companies that start from the headcount number tend to rediscover, in public and at their own expense, what a customer will tolerate.

Frequently asked questions

What is an AI-first company?

It is a company that treats AI use as a default expectation in daily work and checks whether AI can handle a task before adding people or contractors. The label comes from public memos at Shopify and Duolingo in 2025. It describes a management stance, so the useful question is what the team actually does differently.

Should an SME require teams to prove AI cannot do the work before hiring?

Use it as a review step and never as a veto. Asking for a short written test of the AI route forces teams to describe the process, which often surfaces waste. A hard rule that blocks hiring regardless of the result can push quality problems onto customers, as Klarna's reversal in 2025 suggests.

Does AI reduce headcount in practice?

The best current evidence shows little net effect so far. An NBER survey of about 9,000 executives found essentially no net employment change, while Stanford found a 13% relative employment decline for 22 to 25 year olds in the most exposed jobs. Effects differ by role and age, so measure your own workflows.

Which roles change most when a company adopts AI?

Developers move toward reviewing and orchestrating agent output, support agents toward handling exceptions and escalations, and operations staff toward owning a workflow end to end. New responsibilities appear around evaluation and the agent harness. These shifts are directional, since no neutral dataset tracks job titles inside firms.

What should a Swiss SME check before cutting staff because of AI?

Check the employment contracts and the mass-dismissal consultation rules in Art. 335d ff. of the Swiss Code of Obligations with an employment lawyer before any announcement. Thresholds depend on company size and the number of dismissals within 30 days. Also confirm that the AI workflow has run stably for months, since rehiring is costly.

Insights

Put these ideas to work

A 30-minute call is enough to find out whether an AI agent fits your workflow — and what it would return.