Skip to content

No owner, no kill switch, no credential

Teams assume someone else owns agents on shared accounts. A BC tribunal held a company responsible for its chatbot. Identity, authority limits and a tested stop come before access.

In brief

  • A Cloud Security Alliance survey of 228 IT and security professionals, commissioned by identity vendor Aembit, found 68% could not clearly distinguish agent activity from human activity.
  • Identity programmes were built for people and fixed services, and agents are neither: business teams create them in an afternoon and nobody offboards them.
  • Before an agent gets a credential, give it a named owner, its own scoped identity, an authority limit, a budget, an expiry date and a kill switch someone has tested.

The usual question about AI agents is whether they can be trusted to act. The more pressing one is who answers when they do. Take an agent that reconciles invoices under a shared service account someone created during the pilot. Security assumes engineering owns it, and engineering assumes IT does. When it does something nobody intended, the first hour goes to working out who can switch it off.

That agent is typical. A Cloud Security Alliance survey of 228 IT and security professionals, commissioned and financed by identity vendor Aembit, found 85% running agents in production. Ownership split four ways, and security, the most common owner, held it in 28% of organisations.

Accountability itself is settled; the plumbing is what is missing. No agent should get a credential until it has a named human owner, its own scoped identity, an authority limit, a budget and a kill switch someone has tested.

The business owner answers for the agent, and security writes the rules

Ownership has to be assigned deliberately, and a workable division follows the benefit: the team that gains from an agent owns its purpose and outcomes, the platform team runs the machinery, and security sets the rules.

The team that benefits from an agent should own it, while security sets policy rather than owning every agent
RoleOwnsDoes not own
Business ownerPurpose, outcomes, approval to runCredential mechanics
Platform or identity and access management (IAM) teamRegistry, identities, gateway, kill switchWhether the agent should exist
SecurityPolicy, monitoring, incident responseDay-to-day ownership of every agent

Source: Wennovate framework.

The pattern to avoid is security as the default owner. It becomes the bottleneck for every new agent and the scapegoat for every incident, while the business team that benefits carries no accountability.

Agents fit neither kind of actor identity programmes were built for

Identity programmes were built for two kinds of actor. People join, change roles and leave, and HR events drive their access. Services are deployed by engineers and rarely change what they do. Agents are neither. Business teams can create them in an afternoon, they choose their own steps toward a goal, and nobody offboards them.

The volume problem predates agents. An August 2026 O'Reilly Radar article cites estimates, from Rubrik Zero Labs and CyberArk respectively, of 45 to 82 nonhuman identities (credentials used by software rather than people) for every human in enterprise environments.

The same article cites World Economic Forum analysis finding 51% of organisations have no clear owner for AI identities.

Agents add speed. In a vendor-commissioned Harris Poll of 685 CIOs for Dataiku, 84% said employees build agents and applications faster than IT can govern them. 83% lacked standardised agent lifecycle management. Asked who is responsible when agents go wrong, the CIOs split almost evenly across shared teams, central IT, the data or AI team, and security.

The CSA data shows the consequence. 68% of its respondents could not clearly distinguish agent activity from human activity, and 43% said agents use shared service accounts.

Access is loose too. 74% said agents often get more access than necessary. 52% said agents at least sometimes inherit access meant for humans or other systems.

The law already names who is accountable; the mechanics are new

In an MIT Sloan Management Review and BCG survey of 1,221 executives, 69% agreed that holding agentic AI, software that pursues goals by taking actions on its own, accountable requires new management approaches. Novelty bites hardest where agents delegate to other agents or call tools on a user's behalf. A June 2026 CIO.com analysis argues that authority sprawl, the stacking of delegated permissions with each hand-off, not agent count, drives the accountability gap, until nobody can say who authorised the final action.

In the same survey, a solid minority of 25% disagreed and called the push for new approaches a form of AI exceptionalism. They hold that existing frameworks can be adapted, provided people keep clear accountability for design, behaviour and outcomes.

Canadian law sides with the minority on accountability. In Moffatt v. Air Canada (2024 BCCRT 149), a customer relied on a website chatbot's wrong answer about a fare policy. The airline argued, in effect, that it could not be liable for information from its agents, chatbot included. BC's Civil Resolution Tribunal called this "a remarkable submission". It held the company responsible for all information on its website, whether it came from a static page or a chatbot. The award was small; the principle is not. An agent's action is the organisation's action.

So accountability is not new; the mechanics are. A human employee who makes a mistake can be identified, questioned and stopped within minutes. An agent running under a shared account can be none of those things unless someone builds for it.

Five controls come before any credential

  1. A registry entry with two names. Record the agent's purpose, the systems and data classes it touches, and its expiry date. Name a business owner accountable for outcomes and a technical custodian who operates it. If nobody will put their name on it, it does not run.
  2. Its own identity. No shared service accounts and no borrowed human credentials. Issue a workload identity, a machine identity tied to the running task, with short-lived credentials scoped to that task, checked when the agent calls a tool rather than only at install.
  3. Tiered authority. Separate reading, reversible writes, and irreversible or external actions such as payments, deletions and customer messages. The last tier needs human approval or a hard limit. Make actions idempotent, so that a repeated request has the same effect as one, and a retry cannot pay an invoice twice.
  4. A budget. A spend cap, a step limit and a rate limit per agent. A runaway loop is both a cost problem and a security signal; Part 3 of this series covers the cost side.
  5. A tested kill switch. Decide in advance whether stopping means revoking tokens, terminating compute or changing policy, and who can do it at 2 a.m. In the CSA survey, 49% said their containment action would be disabling identities or revoking tokens, and 42% would terminate the compute environment. Choose yours, then rehearse it quarterly.

Alongside these, keep an action record outside the agent framework. Practitioners on Hacker News describe an execution layer between agents and business systems that checks authority and budget, then logs each action before it runs. Framework logs scattered across teams are not a system of record, and a single record is what an organisation will need to answer a regulator, or a tribunal, about what happened and why.

In the agent builds we have been part of, the registry entry is the step teams resist and the one they are most glad of after the first incident.

A retired agent can leave live credentials behind

When an agent is switched off, its credentials, scheduled jobs and tool connections often survive it. An orphaned token with write access is worse than a running agent, because nobody is watching it. Treat decommissioning as a checklist: revoke every credential, remove the registry entry's authority, archive the action record and confirm that nothing still calls the agent's endpoints.

In the Dataiku survey, 47% of CIOs had decommissioned more than 20 agents this year, so this is routine work. The registry's expiry date forces the question: at expiry, the business owner either renews the agent with a fresh review of what it can reach, or it is retired. That single field turns the O'Reilly article's fourth question, when an agent should die, from an afterthought into a calendar entry.

Alberta's commissioner wants the power to order a system stopped

Alberta's privacy commissioner has set out the most detailed proposal in the West. In a 2025 report summarised by Bennett Jones, the OIPC (Office of the Information and Privacy Commissioner) recommended amending PIPA, Alberta's private-sector privacy law, for automated decision-making. Proposals include plain-language notice before an automated decision, an explanation of the information and criteria used, and a route to contest a decision and request human reconsideration. Higher-risk systems would need impact assessments filed with the OIPC before use, and the commissioner would gain powers to audit them and to order an organisation to stop using one.

These are recommendations, not law. But a regulator able to order a system stopped presumes something every agent owner should already have: the ability to stop it, quickly and completely.

Deployment is not waiting for the rules. KPMG Canada reported in May 2026 that 77% of Canadian executives surveyed already use agents, and 66% are moving toward an integrated human and AI workforce.

Put it to work: Audit your agents' credentials and kill switches

  1. How many of your agents run under shared service accounts or a person's credentials today?
  2. For each agent in production, can you name its business owner, its technical custodian and its expiry date?
  3. When was the kill switch last pulled in a rehearsal, and who could pull it at 2 a.m.?

An agent's action is already the organisation's action

The costliest mistake is letting agents inherit human permissions. Borrowed credentials make an agent's actions indistinguishable from the person's and impossible to scope. Close behind is treating install-time vetting as enough, when authority should be checked at the moment of action, and making security the owner by default, which concentrates blame without concentrating control.

Arguing that the agent did it is no defence; a BC tribunal has already rejected that line. A kill switch nobody has pulled is a hope rather than a control, and an agent without an expiry date is one nobody will review.

Most organisations will not settle this year whether agents are a new kind of worker. Decide this quarter that no agent runs without a name beside it and a tested way to stop it.

Next in this series: Shadow AI is the approved stack's harshest product review

Running AI across your teams? Book a 30-minute discovery call with the Wennovate team. No pitch deck required. Start with a scoping call

The Studio Letter

One letter a month from the CEO: what moved, what stalled, what we learned.

Get the Studio Letter