The Agentic Readiness Test

1. Set the Authority

Consider a customer-service agent that can recommend a remedy, update a customer record, and issue a credit. Leadership has several defensible options. The agent can remain advisory. It can act after human approval. It can receive bounded authority up to an agreed financial or customer threshold. The appropriate choice depends on the consequence, but leaving the choice implicit is not an option.

Before approval, ask the business sponsor to complete one sentence: “This agent may take these actions, it must stop at these boundaries, and I remain accountable for the business outcome.” If no executive is prepared to own that statement, the use case is not ready.

Klarna illustrates why this responsibility continues after deployment. In 2025, chief executive Sebastian Siemiatkowski acknowledged that the company’s pursuit of cost reduction in customer service had gone too far and said customers should again have a reliable option to speak with a person. The leadership lesson is the willingness to own the cost-quality trade-off, read the operating signals, and change direction.

• READINESS TEST:  Can the responsible executive describe what the agent may decide, what it may do, and where human judgement must take over?

2. Make Governance Operational

When a pilot requests production access, the useful questions are operational. Which forum can approve it? What evidence must the sponsor provide? When will performance and controls be reviewed? Where does an exception go? Who can pause the agent without waiting for a committee to assemble? A policy that cannot answer those questions has not yet become governance.

Singapore’s Model AI Governance Framework for Agentic AI reinforces this practical sequence: bound risks before deployment, make humans meaningfully accountable, implement technical controls and processes, and enable users to exercise responsibility. Gartner separately predicts that by 2027, 40 per cent of enterprises will demote or decommission autonomous agents after governance gaps emerge in production. That figure is a forecast, not an observed failure rate. Its useful implication is that controls should become stronger as authority, access, and potential consequence increase.

• READINESS TEST:  Can governance produce a timely approval, escalation, redesign, or stop decision, or does it mainly produce documents?

3. Equip People to Supervise

Picture an exception at 4:30 on a Friday afternoon. The agent has produced an answer that conflicts with policy, the customer is waiting, and the specialist who designed the workflow is unavailable. Can the supervisor see what happened? Does the employee know whether to override, escalate, or stop the transaction? Is someone available with the authority to decide? This is where organisational capability becomes real.

Middle-market research suggests that leaders should not assume this capacity is already in place. Netrio’s 2026 study of 401 US IT leaders found that 82 per cent had AI in production or widespread use, while only 26 per cent described it as scaled and governed enterprise-wide. The study covers general AI and comes from a commercial provider, so it is directional. RSM’s survey of 1,030 US and Canadian middle-market executives found that 86 per cent had integrated AI into operations, while 36 per cent had fully embedded it across core processes. It also found that 85 per cent saw leadership as more enthusiastic about AI than employees.

For a mid-sized organisation, readiness does not require a large committee structure. It may mean one accountable executive, a small number of risk tiers, common evidence requirements, a clear escalation route, and managers who have the time and information to intervene. That is right-sized governance, not reduced governance.

• READINESS TEST:  Can the people running the workflow handle a foreseeable exception without waiting for a small group of specialists to rescue the process?

4. Manage the Portfolio

Now consider three business units purchasing similar agents from different vendors. Each pilot may appear useful, yet together they duplicate cost, create inconsistent access, and divide the evidence needed for assurance. Leadership needs a portfolio view that shows the business owner, authority level, shared capabilities, material risks, expected value, and current decision for each initiative.

The decision is not always to scale. One agent may remain advisory. Another may justify controlled action with approval. A third may be stopped because the operating burden exceeds the value. Strategic commitment includes the willingness to concentrate resources, redesign weak initiatives, and end experiments that no longer justify attention.

• READINESS TEST:  Can the leadership team explain what it will scale, restrict, redesign, or stop, and what evidence will trigger each decision?

When the Failure Is Actually Leadership

Unready leadership rarely appears as a dramatic refusal to govern AI. It appears in ordinary operating patterns that leave important choices unresolved.

Authority is assumed. The agent is already operating, but the people around it cannot state who approved its limits or who may change them.

Business ownership disappears. Risk, legal, security, or technology is named as the “AI owner,” while the executive accountable for the customer, financial, or operating result remains at a distance.

Human oversight becomes ceremonial. A reviewer appears in the workflow but lacks the time, information, expertise, or authority to disagree with the system.

The board receives activity instead of assurance. Pilot counts and adoption figures show movement. They do not show agent authority, incident patterns, control performance, realised value, or unresolved decisions.

The 2024 Air Canada chatbot decision makes the accountability boundary concrete. The British Columbia Civil Resolution Tribunal rejected the argument that the chatbot was a separate legal entity responsible for its own actions and held the company responsible for information on its website. The case concerned a conventional chatbot rather than agentic AI and should not be treated as a universal legal rule. Its executive relevance is direct: accountability does not move outside the organisation because a digital system generated the interaction.

Use the Six-Question Test Before Scaling

The four-step chain at the start of this article is a pre-deployment sequence for clarifying one agent’s operating authority. The six-question test below serves a different purpose: it checks whether the leadership and operating system is ready to scale that authority. Choose one agent that is live or approaching production. Put its workflow, permissions, and business outcome in front of the leadership team, then work through the questions below.