# The future of work with AI: what Anthropic's experience can teach companies starting to adopt AI

> Anthropic now runs on AI agents. Most companies do not, and do not need to. What its 2026 data teaches about the future of work with AI, and how to get ready one workflow at a time.

Language: en · Category: industry · Author: Jonas Canton (Co-founder) · Published: 2026-10-23

## What does the future of work with AI look like for companies starting out?

Anthropic is an outlier, and that is a fair objection. It sells the models it uses, and by August 2026 its employees' agents logged six times as many hours as the employees themselves. Yet its report, [How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic), published on 16 October 2026, raises the questions we already meet in companies that are only starting to adopt AI: who checks the work, who owns it, and where the freed time goes.

The change did not stay with engineers. At Anthropic, coding-agent sessions among non-technical employees went [from 2 a month in August 2025 to 90 in May 2026](https://www.anthropic.com/research/employment-at-anthropic). A sales representative built their own dashboard in three to four hours; a finance manager built an agent to review purchase requests. In most companies, the same move starts with no announcement: a personal ChatGPT account, a first automated email draft.

**AI agents in the workplace are AI systems that carry out work on an employee's behalf (drafting, checking, scheduling, writing code) inside the tools a team already uses, within a framework of rules, permissions and supervision.**

Most companies are nowhere near Anthropic. Many still run on email, spreadsheets and a handful of business applications, with data spread across systems that do not talk to each other. That gap is no reason to look away. Agents do not need a fully digitalised company to change how a team works: one workflow, such as preparing a quote or answering a tender, is enough to shift who does what, who checks and where the time goes.

## What did Anthropic's report actually measure?

[The report](https://www.anthropic.com/research/employment-at-anthropic) combines internal data on employment, agent usage, output and hiring from 2025 to 2026 with 17 employee interviews. Four findings matter for any company, because each raises a question leadership will have to answer.

| What changed | Anthropic's figure (2026 report) | The question it raises |
| --- | --- | --- |
| Output per person | Code commits per engineer about 5 times higher than six months earlier; contracts per legal employee about 3 times higher | Is more output more value? |
| Supervision | Reading everything their agents produce would take over half the work week, even in non-technical teams | Who checks the work? |
| Work about the work | Agent sessions spent on specs, reviews and monitoring rose from 10% (autumn 2025) to 22% (spring 2026) | Who owns the agents' instructions? |
| Junior hiring | The junior share of technical hires fell from 27% to 9% | Where will tomorrow's experts learn? |

The authors state their limits: the company grew very fast over the period. Output went up, but output is not yet value.

## Why does human attention become the bottleneck?

Agents produce faster than people can read, check and decide. At Anthropic, [interactive agent sessions spend about 80% of their time](https://www.anthropic.com/research/employment-at-anthropic) waiting for a person to respond, and one engineer described facing "an army of Claudes, with every one asking for my attention." [Written standing instructions for agents](https://www.anthropic.com/research/employment-at-anthropic) grew from a few hundred in July 2025 to 8,500 a year later.

We hit the same wall inside skilder. One of our own roles, built to set up three marketing foundations in a workspace, created the first one and stopped without saying so. Only a check caught it, and the fix was a new version of the role with a verification step before it reports the work as done. More agents do not mean less checking: they mean the checking has to be designed.

So the time an agent saves is not the time a person gets back. A workflow that is 60% automatable does not free 60% of anyone's week: supervision, exceptions and approvals remain. The figure worth tracking is **freed capacity**, the net human time left once that supervision is counted.

Time is also only one indicator. Depending on the workflow, value shows in errors avoided, deadlines met or quality delivered. Choosing the indicator before building, and recording its starting point, is what separates value from activity, the core difficulty of [putting a value on agent skills](/en/blog/measuring-agent-skills-value/).

A 40-person consulting firm we work with put it plainly: the subscription was never the problem, the invisible return was. Its team now sets an estimated value per run on each skill. A code review skill, estimated at 20 CHF of review time per run, ran 237 times in a month. It is an estimate, not a measurement, but it gave the leadership team a figure to discuss instead of a feeling.

## Should a company automate everything AI can do?

No. The useful exercise is to match what AI can do with what teams want to hand over, and to keep some simple tasks human on purpose. Simple tasks are often where people learn the trade. Agentize all of them and the company loses the path from newcomer to expert.

The signal is visible beyond Anthropic. A [jobs.ch study published in June 2026](https://www.thestar.com.my/tech/tech-news/2026/06/24/fewer-job-offers-for-junior-roles-due-to-ai-swiss-study-shows) analysed 7.3 million job ads. It found 26% more offers for senior positions in AI-exposed occupations in 2025, while junior postings in Switzerland were 32% lower than before 2022.

Sandra Carmignani, organisational development and corporate culture consultant at Jobprofile, named the risk at Vision Entreprise 2026 (University of Lausanne and PME magazine, 8 October): "The risk is solving today's talent shortage while creating tomorrow's, because we will no longer have the experts able to question AI and challenge it."

## Who is responsible for an AI agent's work?

The business expert who designed it. An agentic workflow is a team of roles, not one agent that does everything. Every role needs an owner: the person who knows the job, decides on colleagues' change requests and reports on each new version to their manager.

The same consulting firm's first worry was not the model. It was a colleague editing a skill that worked and breaking it for everyone. Its answer was to keep the best skills already in use, give each one an owner and publish changes through a validation step. Ali Mahfouz, Area Service Lead Romandie at Microsoft Switzerland, framed the questions at Vision Entreprise 2026: "Who is responsible for the agent? What data can it access? How do you check what it did?"

An AI role is never finished. It lives through the **AI Capabilities Lifecycle**, six stages run by its owner:

1. **Author**: write the skills, add the tools, keep the versions.
2. **Test**: try the role before an agent uses it.
3. **Deploy**: make it available to the team.
4. **Control**: set the guardrails, what the role is allowed to do.
5. **Monitor**: see every use, from start to finish.
6. **Improve**: turn real usage and colleagues' requests into a new version.

Control and Monitor apply to the role, never to the people who use it. The role, not the person, is also the [permission boundary for AI agents](/en/blog/least-privilege-ai-agents/).

## Why do AI agents push organisations to work in the open?

Because an agent only knows what the organisation writes down. At Anthropic, decisions taken in unrecorded meetings leave agents working from stale context. Meetings have not disappeared either: [median meeting time rose 20% in a year](https://www.anthropic.com/research/employment-at-anthropic), because the issues agents surface still need people to debate and decide.

Employees use AI without saying so, a pattern known as [shadow AI](/en/blog/what-is-shadow-ai/). Those hidden uses are a starting point, not a fault. Acknowledging them without sanction, and writing the AI charter with the teams, brings them into the framework faster than a ban.

We see this in almost every first meeting. At a Swiss construction group, one business unit alone was using half of the company's ChatGPT allowance, uploading full tender submissions with no review. The first reaction was a monthly cap per employee. A cap limits the cost, but it says nothing about which uses create value. At a property management firm, management had Copilot licences while the rest of the team had no structured tool: the AI use that existed stayed personal and invisible.

Silence is already a message: without clear words, teams fill the gap with their fears. Transparency alone does not build trust, though. It needs a visible commitment to invest in people, through training and evolving jobs.

## Who decides what happens to the freed capacity?

Leadership does. Tools can show what AI takes over and how much capacity it frees. Deciding where that capacity goes, and committing to it in front of the teams, belongs to the leadership team. No vendor can make that promise on a company's behalf.

A construction company we work with in French-speaking Switzerland showed what this looks like. Several departments had started their own AI projects, with no common place to share them. The managing director took the topic to the board with a clear message: AI is not an IT subject. The team chose to take one more month to align leadership before launching. In our experience, that month costs far less than a programme nobody at the top has endorsed.

Anthropic made its own choices visible. It is building an early warning system for roles where work is shifting, and [long-serving employees were 50% more likely to move between departments](https://www.anthropic.com/research/employment-at-anthropic) each month in 2026 than in 2025.

The goal is not to do more with less. It is to do much more with the people the company already has, and to give them time for what humans do best.

## How can a company get ready for AI agents?

By starting with one workflow rather than a company-wide programme. A company does not need to be ready to start: it becomes ready by moving forward. Early results give teams a reason to tackle less attractive work such as data and documentation.

1. **Set a direction and a framework, in a few weeks.** Security, confidentiality and compliance rules belong to the framework: they are not traded for speed.
2. **Acknowledge the AI uses that already exist.** Map them without sanction, then write the charter with the teams.
3. **Pick one workflow precise enough to measure.** You don't agentize "acquire a customer". You agentize "prepare a sales proposal from a brief".
4. **Choose the indicator before building, and record its baseline.** Time, errors, lead times or quality: whichever this workflow must improve.
5. **Match what AI can do with what the team wants to hand over.** Mark the learning tasks that stay human.
6. **Let the people who live the problem design the roles.** The designer becomes the role's owner.
7. **Measure net value against the baseline, then share it.** Show the result to the team and to leadership, and decide where the freed capacity goes.

Once one workflow has proved itself, the same steps extend to the next team, in reversible waves. Making the tacit knowledge of the job explicit is often the most valuable part, as [knowledge vs. know-how](/en/blog/knowledge-vs-know-how/) explains. A [free 30-minute exchange with an expert](https://koalendar.com/e/meet-with-nicolas-9) can help choose the first workflow.

## How does skilder help companies prepare for AI agents?

skilder gives every AI role an owner, a framework and a lifecycle, inside the tools teams already use. Roles are configured from blueprints, templates adapted to the company's own processes, then served in Claude, ChatGPT or Copilot, or any MCP-compatible agent. We run our own company on it: our proposal role turns a one-line request into a proposal and a matching deck after three clarifying questions, and its owner improves it version after version.

- **Roles with owners.** Each role bundles skills, tools, context and permissions, and keeps the business expert who built it as its owner.
- **The AI Capabilities Lifecycle.** Write and version skills (Author), try a role in a real conversation before an agent uses it (Test, with automated evaluations coming soon), publish it with role-based permissions (Deploy), set guardrails that skilder enforces on every tool call (Control, in beta), and see every use from start to finish (Monitor).
- **A trace of every run**, attributable and exportable.
- **Value read against a baseline.** The role owner can set a value per use in CHF; read against the baseline recorded before building, it gives an estimate of net value, not a count of users.
- **Hosting in Switzerland and the European Union**, with Infomaniak by default.

skilder makes visible what AI takes over and the capacity it frees. The decisions stay with the company. For the governance side, see the [practical framework for enterprise AI governance](/en/blog/enterprise-ai-governance/).

**Takeaway:** agents will not wait for an organisation to be ready. Start with one workflow, name the owner of every role, measure what is left once supervision is counted, and decide in the open where the freed capacity goes.

To find the first workflow worth agentizing, [book a free 30-minute exchange with an expert](https://koalendar.com/e/meet-with-nicolas-9).
