Agentic organization:AI agents as employees,and who answers for them

I have worked with AI agents since they first appeared in software development. Today, I work with a whole fleet of them. At some point you notice you are no longer operating tools. You are answering for a position. So how far along are other businesses?

The latest German figures come from Bitkom, the German digital industry association: September 2026, representative, 603 companies with 20 or more employees. 57 percent of all companies now use AI. A year ago it was 36, two years ago 20. Of those that use or plan AI, 11 percent already run agents, 29 percent plan to, 31 percent are discussing it. For only 24 percent it is not on the table (Source: Bitkom, 2026). And zero percent say they get the full potential out of AI. 59 percent say: not at all.

Bitkom president Ralf Wintergerst puts it this way: AI agents today are roughly where AI as a whole was three years ago. The same release also says an agent that acts on its own needs reliable data, connected systems, and rules about what it may and may not do.

Startups look different. In a Bitkom survey of 102 tech startups from July, more a mood check than a representative sample, 48 percent use AI agents. 65 percent build their own AI applications, and not a single one goes without AI (Source: Bitkom, 2026).

McKinsey looked at this internationally. Large corporations are now scaling agents at 40 percent, up from 27 the year before. Smaller companies remain flat at 22 percent ( Source: McKinsey, 2026 ). The gap is growing, even though small businesses could benefit greatly. A single agent can already carry a noticeable share of their work. “Smaller” at McKinsey, though, means under a billion dollars in revenue. A twelve-person business barely shows up in that statistic. So translating the frameworks of the big players down to your size is up to you.

To be fair, the 11 and the 48 percent cannot be compared directly. One is representative of businesses with 20 or more employees, the other a snapshot of an industry built around software. The startups offer a glimpse of where the market is heading. Adoption could move quickly: Wintergerst expects agents to spread much like AI overall, which rose from 20 to 57 percent in two years.

This spring, I argued in “From Agentic Coding to Agentic Organization” that AI should flow through every part of a business like electricity, reaching far beyond development. This article picks up one step later: the agents are already at work, and someone has to lead them.

For a 2026 study, Deloitte asked executives about agents in the workplace. One finance executive says it plainly: “The machine will never be accountable” (Source: Deloitte, 2026). Accountability is the dimension most likely to be overlooked, Deloitte writes. Almost 70 percent of respondents name the inability to trust and govern agents as a major barrier (Source: Deloitte, 2026).

91 percent of AI investment goes into technology infrastructure, 7 percent into work and people (Source: Deloitte, 2026).

Since July, Microsoft recommends governing agent identities exactly like employee identities. They name three problems. Agents pile up access rights and keep them forever. When an agent acts on its own, nobody is accountable. And managing all of it by hand does not scale. Their answer: every agent gets a “named sponsor”, a person who answers for the agent’s access and lifecycle, renews it, or shuts it down (Source: Microsoft, 2026).

In April, Gartner published six steps against “agent sprawl”: governance rules, a central agent inventory, identity and lifecycle per agent, rules for data access, monitoring behavior, culture (Source: Gartner, 2026).

Gartner designed these steps for organizations running hundreds of agents. Applied to a small business, they fit into a single table. I take the term sponsor a little further than Microsoft does. There, the sponsor answers for access and lifecycle. In a small business, the same person is also accountable for what the agent does. Several people may work on an agent. One person carries responsibility: one name per agent. The inventory is the table itself, one row per agent. Identity means: the agent has its own account, not yours. Lifecycle means: a review date and a stand-in for when the person responsible goes on holiday or leaves the company. Data access lives in the “may and may not” column. Monitoring and culture do not fit into a column. They are what the next two sections are about.

An agent without a sponsor is a temp nobody hired. With access to the books.

Trust and AI use go together, and two studies show it independently. Both are correlations, not proof of cause: people who are productive may simply like AI more. In spring, Gartner surveyed 12,004 employees and managers in 40 countries. Employees with a positive outlook on AI are 3.4 times more likely to be highly productive (Source: Gartner, 2026). The main drivers: confidence in one’s own current and future role, and ongoing, open communication about how AI is used and what that means for jobs. Only 27 percent of executives there have a comprehensive AI strategy, only 20 percent think their workforce is truly ready.

Deloitte measures a different pair, with its own method, so the two multipliers cannot be compared: trust in one’s own organization and actual use. Employees with high trust in their organization are 3.8 times as likely to use agentic technology (Source: Deloitte, 2026).

In the Work Trend Index, 81 percent of leaders say they feel safe suggesting new ways of working with AI. Among employees it is 67 (Source: Microsoft Work Trend Index, 2026). “My manager makes room for experiments”: 78 to 59. “Trying new things is rewarded, whatever the outcome”: 21 to 10. Leadership loves the agents. Someone on the team waits to see whether they get the blame when the agent makes a mistake. The name in the table says who cleans up the mistake and teaches the agent what went wrong. Answering for an agent here means fixing and explaining. If an agent’s mistake turns into trouble in the stand-up, nobody will volunteer their name on the board for the next agent.

What happens when output flows into the team unchecked was measured by a team from BetterUp Labs and the Stanford Social Media Lab. 40 percent of 1,150 full-time workers in the US received “workslop” in the past month: AI-generated work that looks like good work but does not move the task forward (Source: BetterUp Labs and Stanford Social Media Lab via CNBC, 2025). 53 percent were annoyed by it, 38 confused, 22 offended. Half of them saw the sender as less capable afterwards. An agent that delivers into the team unchecked does the same thing, only faster. The reputation that takes the hit belongs to the sponsor.

Deloitte also advises leaving enough slack in processes and plans so people can absorb the unexpected. For me that means: whoever wears the controller hat gets hours for it in the weekly plan, not “on the side”. Otherwise friction on purpose turns into an approval click at 5:58 p.m.

Tell people what the agent is not allowed to do. That reassures more than the list of what it can do. In the stand-up, the log sits next to the result. And the person who answers for an agent gets to say when they want it switched off.