How much may an AI agent decide on its own?
Witold Reichhart · witold.reichhart@capgemini.com
Before the map, the first slide puts what a firm lets AI do on a ladder and marks where an answer becomes an act. The second shows capability and control moving apart, with the incident count, the adoption rate, and the share of firms that govern their agents. The third says what control is made of in a firm, and how small the software part is. The fourth turns the argument round: the value of an agent grows with how much the firm can safely let it do, so the evidence a firm can show sets the ceiling on the value it can take.
About 6 minutes. → or the space bar moves on, and through the steps first where a slide builds in steps. Click a line on the left to go straight to that slide.
M opens the contents, F finds a word, and T switches between light and dark.
Then part one draws the map of what must be governed when an agent acts, at the enterprise, in the workflow, and on the platform, and says what the control plane is.
Would you let an AI system do this on its own? The answer changes as you climb the ladder, and so does the question the firm has to answer before it says yes.
Below the line the firm checks an answer. Above it the firm has delegated an act, and the question changes: what may it do, for whom, and up to what amount?
Capability moves with every model release, every few weeks. The firm’s control of delegated work moves with its policy cycle, once or twice a year. The gap between the two is where the incidents happen.
Up by more than half in one year. In the same year regular use of AI in at least one business function reached 88 percent of the organisations surveyed.
21 percent of the firms planning agents within two years said they had one, in a survey of 3,235 business and IT leaders in 24 countries. In Deloitte’s words, agents are scaling faster than their guardrails.
Probabilistic. The output varies from run to run, and its confidence can exceed its correctness.
Connected. The model retrieves data, calls tools, and triggers acts in other systems.
At machine scale. One weak control repeats across thousands of decisions before anyone reads one of them.
A model error becomes an enterprise event the moment it is connected to authority, data, and action. The ladder’s upper rungs connect all three.
The gap on the last slide is not closed by a product. Anyone who runs an organisation knows what control is made of: parts the firm already has, each owned by someone in this room, and only one of them runs on a server.
Regulation binds the firm; the firm decides what applies, and writes it into policy. For an agent that reading has to be written down, dated, and owned.
The risk appetite, the ceilings, the control standard. They exist today for people; an agent needs them as numbers and rules a machine can read.
How the work is done, step by step, and who may decide what. The mandate an agent works under is a procedure written tightly enough to be checked.
What was decided, on what basis, by whom. The three lines and the supervisor read it; an agent has to leave one that someone who was not there can rebuild.
The tacit part: how the unusual case is treated, what the procedure leaves unsaid. It reaches an agent only through the people who take its referrals.
Identity, access, the gateway, the logs, the model and its harness. The part that runs on a server, and the only part a product can sell.
Every one of the six is a piece of the control plane, and every owner above has a seat at this table. The map in part one gives each piece a box, at the altitude where its owner works.
The value of an agent grows with how much of the work the firm lets it do alone. Each step up needs a step of evidence the firm can show, and the set’s first hard rule holds at every step: autonomy may not rise above evidence.
The evidence a firm can show sets the ceiling on the value it may take from its agents. Raising that ceiling is the work of the control plane, and part one draws the map of it.