Six Questions for 2027

Fifty people arrive on Monday morning and offer to work for nothing. Most organizations would hesitate, and with reason: nobody has the time to supervise fifty newcomers, let alone decide what they should be allowed to see and who answers for mistakes made in the company's name.

The fifty arrived months ago with a software update, and in most organizations they are still waiting. What jobs are worth giving them? Try training fifty people for fifty roles at once, and your day disappears into a hundred tiny questions. But take one aside and show it how you do the work. Look at what comes back. Correct the mistakes. Let it try again. A few weeks later, it starts producing useful results.

A human role bundles judgment, context and dozens of small tasks. No agent can take on that whole package reliably yet. What works is a single checkable assignment: prepare a briefing before every client meeting, turn incoming requests into a first draft, or monitor a regulation and flag relevant changes. An agent that does one of those reliably can be given a second.

Once you stop thinking in roles and start thinking in small, checkable assignments, the fifty become much easier to place. Near-zero-cost labour also changes what is worth doing: you can spend it on work that previously had no business case. Some of the fifty can work on revenue. Some can work on quality, reviewing and improving what the company produces. Some can work on time, removing the repetitive administrative tasks that fill people's days.

That last category may be the best place to start, because the benefit is immediate and personal. When an agent gives someone three or four hours back every week, the user becomes its advocate. Adoption starts to spread through usefulness rather than through a programme.

Pushed far enough, this produces companies that look impossible. The American telehealth company Medvi reports a run rate of 1.8 billion dollars this year with just two employees - two brothers. The real story, though, is not that two people do everything. Beneath them sits an operating network of physicians, pharmacies, fulfilment companies, compliance providers and contractors, with AI handling parts of the work between them. The model is powerful, but not frictionless: it has also faced regulatory scrutiny, advertising concerns and a data breach involving an infrastructure partner.

The limiting factor, however, is not intelligence. It is access. An agent that cannot see your systems may sound intelligent, but it cannot do much - it is still standing outside the building. Real work starts when it can enter the offer archive, open the project files, read the register or act on your inbox.

And that is where the hard questions begin: what can it read? What can it write? Where can it act? Which systems may it touch? What happens when a connection breaks or a system changes?

The work itself may cost almost nothing to run, but reliability has a price: access, testing, exceptions, monitoring and the attention of the person responsible. The fifty are free in the same way an intern is free before you count the supervision. Most attempts do not survive that step: Gartner expects more than four in ten agentic projects to be cancelled by 2027, usually for cost, unclear value or missing controls rather than for lack of capability.

Which leaves the most important question: what are the fifty for? The most useful starting point is not replacing your people, but giving them time back. How those hours are used - to serve more clients, build something new, improve the work, or let people go home earlier - is ultimately a management decision, not an IT project.

SourcesMEDVi run-rate and two-employee model — ARR Club ↗ · MEDVi advertising and regulatory context — Business Insider ↗ · OpenLoop Health data-security incident — PR Newswire ↗ · MEDVi / OpenLoop relationship — MEDVi privacy policy ↗ · Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 — Gartner press release, 25 June 2025 ↗

Ask an experienced executive what only they can do, and they will often point to relationships, judgment and twenty years of knowing what good looks like in their field. Those answers may be right, but they are still abstractions. The sharper question is what someone else could reproduce - and what would resist being copied.

Someone may run a version of this test on your company before you run it on yourself. Since 2023, Bain & Company has been using AI to recreate parts of software businesses its private equity clients are considering buying. The point is not to build a perfect copy. It is to find out what is easy to copy. Hundreds of prototypes have turned that into a new diligence question: if the product can be rebuilt cheaply, where does the company's advantage actually live? At least one investor abandoned a deal after the answer proved uncomfortable. The moat, if there is one, may lie elsewhere - in customer relationships, proprietary data, trust or deeply embedded ways of working.

Run the same test from the inside. Assume that anything reproducible eventually will be reproduced: your documents, your processes, your software, soon your standard analyses. Then look for what resists - reading a client, sensing risk, earning trust, hearing what is not being said, knowing when to say no. A week of actual work is more revealing than asking people to describe what makes them unique, so go through it entry by entry and ask two questions: could an agent have prepared this? Could an agent have done it? What survives is often much smaller than the job description, and that remainder is where the company's hardest-to-copy value sits.

There is one catch: the boundary keeps moving. Every generation has assumed machines would take the routine work while people kept the meaningful part, and every generation has watched that human territory shrink. Models already draft difficult emails, prepare risk assessments and extract meaning from meeting transcripts. What survives your test today may not survive it next year. The durable skill, then, is not drawing the boundary once but redrawing it continuously - knowing where, right now, human judgment still changes the outcome, and noticing when that is no longer true.

The remainder can disappear from the inside too. In a Boston Consulting Group survey of seventy senior executives this year, half were already seeing de-skilling in their organizations, and more than six in ten expect it to become a serious problem within three to five years. The capabilities most at risk are the same ones that look hardest to reproduce: judgment, problem framing, creative thinking. If AI does more of the thinking, people get fewer repetitions. Fewer repetitions eventually mean weaker capability. So identifying the remainder is not enough. You have to keep exercising it.

The capability boundary may keep moving, but another one does not: accountability. Anything you hand to an agent still needs a human owner. Agents can collect, compare, draft, recommend and act. But when something goes wrong, someone human still has to answer for it. "The agent did it" is not an answer to a client, a board or a regulator.

Once the current line is visible, the calendar and the budget can follow. Give the fifty the preparatory work they can do reliably, and protect more human time for what remains difficult to reproduce. Invest there too - in relationships, judgment, trust and the conditions that strengthen them. If time with clients is part of the moat, put it into the calendar first. Let the fifty work around it.

SourcesBain AI replicas and vibecoding in software due diligence — Financial Times ↗ · De-skilling research — BCG ↗

Once the boundary from chapter two is visible, the next question is how work changes around it.

Senior people already receive work prepared by others. The difference with agents is not only who prepared it, but how much confidence you can place in the explanation afterwards. A human analyst can usually tell you why they chose a comparison. An agent can sound just as convincing, but its explanation may be generated after the fact rather than reflect how the result was actually produced.

Further down the organization, the nature of the job changes more dramatically, because this is where the work used to be made. Research, first drafts, reconciliations and preparation filled the day. Now much of that routine work is done before people arrive. What remains is the work that requires judgment.

The work moves rather than disappears. Production becomes verification: checking what came back, finding what is wrong, deciding what is good enough to use. That is a different job with different skills, and most people have never been trained for it. Everyone becomes the manager of a small overnight team that works fast, never gets tired, and does not know when it has misunderstood the brief. A good producer is not automatically a good verifier. Some people will need practice, some support, and a few a different role. And checking the output is work in its own right - it needs time, attention and space in the day.

Managing the overnight shift requires a new management skill: delegating to a machine. A vague brief produces confident improvisation, and the morning disappears into corrections. A good one is explicit - goal, context, format, constraints, and what done looks like. Without that last part, verification becomes guesswork. You can only judge an output against a standard that existed before the work began. German-speaking managers have a word for the complete version: the Auftrag. The inventory from chapter two decides what may be delegated to an agent. The Auftrag decides whether it comes back usable.

When much of the routine layer arrives already finished, the day gets heavier. The simpler tasks that once created breathing room between periods of judgment and decision-making are gone. So the flow of agent output matters too - when it arrives, how much arrives at once, and how much space comes after. The working day itself becomes something to design.

A deeper problem appears with juniors. They used to build judgment by doing the work - the thousand drafts, reconciliations and analyses that slowly taught them what good looks like. Agents now absorb much of that practice. Juniors are pushed toward verification before they have accumulated the experience that makes it reliable. Some production may therefore need to stay human, much like pilots retain flying hours - not for the output itself, but for the judgment it develops.

This is no longer only a hypothetical problem. In Switzerland, junior vacancies in office and knowledge work - IT, marketing, administration and finance - have fallen by roughly a third since generative AI became widely used. That does not make AI the sole cause; hiring also follows the economic cycle. But many of the tasks disappearing first - research, documentation, first drafts - are exactly the tasks juniors used to learn on. Some companies are protecting that learning path deliberately. Zuehlke increased its junior intake from two to sixteen. BDO teaches apprentices the accounting fundamentals before introducing AI. Thjnk replaced internships whose tasks had vanished with paid entry-level roles.

Middle management may provide part of the answer. The role shifts from control toward coaching. Agents can collect status, check compliance and consolidate reports around the clock, freeing managers from work that once consumed their time. Those hours can move toward juniors - reviewing their reasoning, giving feedback and helping them build judgment. Some of the repetitions may disappear, but managers can spend more time making the remaining ones count.

Sourcesjobs.ch KI Report 2026 · 7.3m job ads · Data: 2025 · Published 24 June 2026 ↗ · Coding Kanban — Tom Tunguz ↗

Many Swiss companies operate in markets where competitive positions have been stable for years. Regulation, reference lists and switching costs protect incumbents. But stability can hide a widening difference. A rival that has adapted well to AI may look unchanged from the outside - same headcount, same products, same prices - while inside, offers move faster, administrative work shrinks and margins widen. That difference can remain invisible until it reaches customers, margins or market share. By then, it may already be structural.

The reason is that the learning compounds. Any competitor can buy the same model tomorrow, likely cheaper and better than the one you started with. What takes time to build is the company-specific layer: agents that have been briefed, tested, corrected and trusted with real work. Each correction adds a piece of knowledge - which exception matters, which client tolerates what, which number deserves a second look. Captured in workflows, instructions and operating rules, that knowledge becomes organizational rather than personal. And each agent you teach makes the next one faster to build.

Not all of that learning lasts. Some instructions simply compensate for today's model - formatting rules, tone corrections, reminders not to invent sources - and better models may make them unnecessary. What survives is the part no model upgrade can supply on its own: the exceptions, standards and ways of working specific to your company. That is the knowledge worth documenting, assigning an owner to and keeping up to date.

The advantage can compound through hiring too. Candidates can increasingly tell the difference between a company experimenting with AI and one that has changed how work actually gets done. An organization that has embedded AI into everyday work can offer better tools, less routine production and more time for higher-value work. One that has not may increasingly lose strong candidates to competitors that have. The productivity advantage can therefore reinforce itself through talent.

Most of this happens before the financials show it. Agents move into permanent work, the next ones become faster to teach, company knowledge accumulates and candidates begin choosing differently. Margins and market share are the late signals. By the time they move, catching up may already take years.

The fifty did not arrive only in your organization. Your customers, applicants, suppliers and the neighbours in a planning dispute got them too, and they are already using them in their dealings with you. That changes how they find and choose you, what they still need from you, and how much work they can send your way.

The buying side changes first, and almost invisibly. Machine activity is already pervasive online: more than half of internet traffic is non-human. Purchasing is beginning to follow. Agents search, compare and order without a website visit or sales conversation. The volumes are still tiny, but growth is rapid. And while forecasts differ on timing, the direction is clear: more of the buying process is moving to machines.

The exact timing matters less than the practical consequence: an agent can only compare what it can read. Products, prices, availability, terms and evidence need to be visible to it. Anything buried in a PDF, behind a contact form or in a salesperson's head may never enter the decision. Relationships and trust still matter, but first the agent has to find you and put you on the shortlist. Most companies have built their sales process for human buyers; increasingly, it also needs to work for machines. That also means deciding which agents you accept, and on whose authority they are buying.

When everyone has access to the same AI workforce, it changes not only how clients buy, but what they still need from you. First drafts, standard analyses and routine reports can increasingly be produced in-house, so the contract may remain while the scope quietly shrinks, renewal by renewal. That leaves every provider with a harder question: once clients have their own fifty, what are they still willing to pay us for? What remains valuable is the work they cannot easily reproduce themselves - judgment, accountability, and tasks they cannot or may not do alone.

The pressure does not only come from clients buying less. Agents also make it much easier for outsiders to send you more work. A complaint, application, claim or formal objection once required time, persistence and some expertise, so friction itself limited the volume. Agents remove much of that friction. The effect is already visible in formal procedures, where the numbers are easiest to measure. In Britain, applications under one obscure employment-law provision reportedly jumped from around twenty a year nationwide to roughly twenty a month in each of twelve regional offices. Employer claims rose 39 percent in a year, the backlog grew 55 percent, and some cases filed today may not be heard before 2030. A Berlin researcher has documented more than ninety examples of what he calls agentic flooding across eleven countries.

The submissions do not just multiply; they can become more elaborate too - better researched, more heavily documented, sometimes padded into hundreds of pages that still have to be read because a legitimate issue may sit inside the noise. The arithmetic is unforgiving. Whatever you save by automating your own work can disappear if inbound volume triples. Some of that increase will require more people, but matching automated inflow with human headcount is not a sustainable answer.

Outsiders can also reach your agents directly. Agents read what you put in front of them: emails, documents, web pages, the contents of a form. If someone hides an instruction inside that content, an agent may follow it — using your access, your credentials, your name. Most organisations running agents have already seen an attempt. The defence is not a better filter. Assume the instruction gets through, and make sure an agent that follows it cannot do much: narrow permissions, no unattended payments, a log someone actually reads.

Any process whose volume was once kept in check by the effort required from the other side needs to be reconsidered. The fifty in your lobby are only half the story. Everyone who deals with you has their own fifty now, and they are already at work.

SourcesBad Bot Report 2026: automated share of web traffic — Imperva ↗ · Tribunal statistics quarterly, January to March 2026 — GOV.UK ↗ · AI-generated employment claims and interim relief — The Times ↗ · AI-assisted employment claims — Financial Times ↗ · AI-generated objections overwhelm Dutch municipalities — NL Times ↗ · State of Agentic AI Security and Governance 2.01 — OWASP GenAI Security Project, 1 June 2026 ↗ · State of AI Agent Security Report 2026 — Gravitee, April 2026 ↗ · Trusted Agent Protocol — Visa, October 2025 ↗ · Agent Payments Protocol (AP2) — Google ↗

Agents can turn an instruction into action with remarkable speed and scale. That means they can execute a bad objective extremely well without ever questioning whether it was worth pursuing in the first place.

Speed removes a safeguard we rarely noticed: friction. A brief once had to pass through analysts, reviewers and colleagues before it became finished work, giving someone along the way a chance to challenge the premise. Agents compress that path. By morning, a flawed brief can become forty polished pages; by Friday, a running process. The mistake has not been caught. It has been scaled.

The employment-tribunal example from the previous chapter shows a second risk. A weak grievance goes into a chatbot and comes back as a polished claim, complete with arguments and anticipated defences. When the Economist tested one with a complaint that had no clear legal basis, the model still found arguments to build on. In one real case, a claimant filed 67 grievances across 282 pages, then told the judge he would rely on only about a tenth of them. The system had not failed. It had turned a shaky premise into a much larger problem.

The same thing happens inside a company. Ask an agent to justify a project and it will find the arguments. Ask it to improve a metric and it will find ways to move it. But neither instruction asks the question that matters first: should we be doing this at all?

That changes what verification needs to do. It is not enough to check the output for typos, broken numbers or invented sources. The assumptions behind the work need checking too: whether the target still makes sense, whether the market assumption still holds, and whether the brief still reflects what the organization is trying to achieve.

The problem gets bigger when many agents work from the same brief. They can share the same blind spot and reproduce it at scale. Where errors are costly, the second view has to come from somewhere genuinely different - a genuinely different model, a human reviewing real samples, or an external test. And one rule should be fixed: the agent that produces the work does not approve it.

So assumptions need owners too. Every standing assignment should make clear who owns it, what assumptions it depends on, and when they will be revisited. And before a new brief is handed to the fifty, someone should have the explicit job of challenging it: why are we doing this at all?

SourcesAI-generated employment claims — The Times ↗ · Tribunal statistics quarterly, January to March 2026 — GOV.UK ↗

If you were starting your company today, you would probably build it differently. Until recently, that was mostly a thought experiment. Now it is a competitive question. A new entrant can build around AI from the start, with a smaller core team and none of the structures your company has accumulated over time.

The founding test asks a simple question: if you were building the company today, what work would still belong to people, what would move to agents, and what would disappear altogether? Only then would you design the roles and processes around what remains.

That changes how work is bundled into jobs. Many roles exist because a collection of tasks has to add up to one full-time position. As agents take on part of that work, those bundles begin to come apart. Roles can become more focused, teams smaller, and work that once required a permanent department may move to a shared service, an automated workflow or a temporary team.

The organization chart then becomes less a map of who performs each task and more a map of where responsibility sits: who makes the decisions, who sets the standards, who owns the client relationship, and who ultimately answers for the outcome.

A new company can design around this logic from the start. An incumbent has to reshape an organization built for a different way of working. Roles, processes and reporting lines often remain even when the work that justified them has changed or disappeared. The harder task is therefore not adding AI, but deciding what the organization no longer needs. If nothing old is removed, AI adds another layer of complexity instead of simplifying the company.

The point is not to rebuild the company from scratch. It is to use the company you would build today as a benchmark for the one you already have. The gap between them becomes the agenda: what should change first, and what can disappear as a result? A year from now, the clearest sign of progress is not how many agents you deployed, but what the organization no longer needs to do.