This year the vocabulary shifted from chatbots to agents: software that does not just answer but acts, across systems, over multiple steps, with less human involvement per step. I find some of the demos genuinely impressive. And in every leadership conversation I have about them, one question decides whether a company will benefit or produce incident reports: when an agent acts, who is accountable for the outcome?
The economics are catching up with the enthusiasm. McKinsey’s 2026 work on agentic AI warns that as companies chase spiralling agent costs, the operating model, not the agent: decides whether value appears[Q]. Its research adds a sharper point: when agentic systems run inside workflows that were not redesigned for them, errors propagate across the organisation at machine speed[O]. And the 2025 State of AI survey finds most organisations still in the early stages of scaling agents at all[S]. Autonomy is shipping faster than accountability.
Agents are operators, not employees
The language of digital coworkers is seductive and wrong. An employee carries responsibility. An agent cannot. It can execute, draft, sort, trigger and escalate, but accountability does not transfer to software, no matter how fluent it sounds. When vendors present agents as team members, they are describing an interface, not an organisational reality.
| An agent may prepare | An agent must not complete alone |
|---|---|
| Drafts, summaries, research | Customer-facing commitments |
| Sorting and routing | Payments and pricing decisions |
| First-pass analysis | Personnel decisions |
| Escalation with context | Anything nobody would own afterwards |
What this means for you
Before you scale agents, settle three questions in writing. Which decisions may an agent prepare, and which may it complete? What happens on the first material error, and who explains it to the customer? What pace of autonomy can your organisation absorb without losing control or trust?
Every agent in your operation needs a human owner, defined boundaries, and an escalation point for what it must never decide alone. If that sounds like bureaucracy, consider the alternative: an automated process nobody owns, making mistakes at machine speed with human consequences.
Human control before autonomy is not caution. It is management.
The correction is already priced in
The analysts have stopped being polite about this. Gartner predicts that more than 40 percent of agentic AI projects will be canceled by the end of 2027: escalating costs, unclear business value, inadequate risk controls, and estimates that of the thousands of vendors now selling agents, only around 130 are real, the rest engaged in what it calls agent washing[GA]. Deloitte finds the same gap from the buyer side: close to three quarters of companies plan to deploy agentic AI within two years, yet only 21 percent have a mature model for governing autonomous agents[DE]. Deployment is sprinting. Accountability is walking.
What I see in the field
In every agent discussion I join, the technical questions get answered quickly and one question survives to the end: who signs for the outcome when the agent acted? Organisations that answer it with a name deploy faster afterwards, not slower, because nobody has to hedge anymore.
Before your organisation gives software the authority to act, the accountability map has to exist: which decisions agents may take, which stay human, and who owns each outcome. That map is precisely what we build with leadership teams.
Personalised answer within 48 hours.
Sources
- McKinsey QuantumBlack, Is That AI Agent Worth It? Agentic Economics and the Modern Operating Model, 2026. mckinsey.com
- McKinsey & Company, The Operating Model Advantage: Why AI Winners Are Rewiring Their Organizations, July 2026. mckinsey.com
- McKinsey & Company, The State of AI in 2025: Agents, Innovation, and Transformation, November 2025. mckinsey.com
- Gartner, Press Release: Over 40 Percent of Agentic AI Projects Will Be Canceled by End of 2027, June 2025. gartner.com
- Deloitte, The State of AI in the Enterprise 2026 (survey of 3,235 leaders). deloitte.com
The AI Act Just Moved. Your Deadline Did Not.