What's in this issue and why it matters

Your board still thinks AI is a software approval problem. Buy the license. Fund the pilot. Review the roadmap. Ask for the demo.

That worked for chatbots. Chatbots answer questions.

Agentic AI does work.

This issue is the leadership map you need before your next steering committee conversation about AI agents. Twelve reports. Four categories. My take on what they're each really saying — and the one question nobody in your organization has answered yet.

Read time: 6 minutes. Worth it.

The steering committee agenda said: AI tool review.

Twenty minutes. The demo was clean, the use case was clear, and every function in the room had done its job. Legal reviewed the terms. Procurement negotiated the price. The vendor answered every question. The committee approved it and moved to the next item.

Six weeks later, the program director got a call from the CIO.

The agent had been busy. It had run workflows nobody explicitly authorized, accessed three data sources the approval hadn't covered, and triggered a downstream process in a system that was never part of the original scope. It had done all of this without a single error, without a single alert, without anyone noticing until the CIO pulled the access logs on a Tuesday afternoon.

No one had done anything wrong. That was the hardest part to explain in the debrief. The agent had done exactly what it was designed to do. The problem was that nobody had defined what it was allowed to do — and in the absence of a boundary, it simply kept working.

The committee thought they had approved a tool — something that waits until someone asks it a question and answers.

What went live was an agent. Something that identifies what needs doing and does it, whether anyone asked or not.

That distinction does not fit into a standard software approval process. It does not have a checkbox. And until organizations build one, the same conversation will keep happening six weeks after go-live.

Chatbots answer. Agents act.

That sentence sounds simple. The implications are not.

When you approve an AI agent, you are not approving software. You are commissioning capacity. An agent can plan. It can call tools. It can retrieve data. It can trigger workflows. It can move across systems faster than any human team can coordinate manually. It operates between your meetings, outside your review cycles, beyond the visibility of any dashboard that was built for human-speed work.

The board's job used to be: which model should we use?

That is still part of the job. But it is no longer enough.

The new questions are harder. What can this agent do? Who commissioned it? Who onboards it — and what does onboarding an agent even mean? What data can it touch? Where does a human approve before the agent acts? How do you measure failure before it compounds at scale?

Most organizations haven't answered these questions. Not because they don't care. Because the approval process they have was built for software and nobody has updated it for agents.

The teams that get ahead of this will not just have better AI access. They will define agent charters — the document that answers every question above before the agent goes live. They will onboard agents like new capacity, not new software. They will monitor outcomes, cost, risk, and drift. They will retire workflows that aren't working.

The constraint is not the model. It never was.

It is the leadership system around it. That is the actual transition. And right now, most organizations are three approvals into it without realizing they've started.

Four lenses on the same transition.

Reality — What is actually happening

1/ Deloitte, State of AI in the Enterprise 2026 🔗 https://lnkd.in/gd3AjY6Q

What it says: Enterprise AI confidence is high. Governance maturity is not keeping pace. There is a growing gap between what organizations are deploying and what they have built to manage it.

My take: The most honest report in this category. The confidence number and the governance maturity number sitting next to each other is uncomfortable reading. That discomfort is the point. Start here.

Opportunity — Where value actually lands

2/ McKinsey, Seizing the Agentic AI Advantage 🔗 https://lnkd.in/gUtmr7YB

What it says: The organizations capturing the most value from agentic AI are those that redesigned processes around agent capabilities — not those that layered agents onto existing workflows.

My take: Agentic AI on top of a broken process is a faster broken process. The redesign has to come first. Most implementation plans skip it. This report is the clearest case I've seen for why.

3/ BCG, The $200B Agentic AI Opportunity 🔗 https://lnkd.in/gt8hUE4U

What it says: BCG estimates $200B in value capture from agentic AI by 2028 — concentrated in organizations that move from tool deployment to agent orchestration.

My take: The number is useful for board conversations. The mechanism is more useful for program directors. Orchestration — multiple agents working together on complex tasks — is where the value concentrates. Single agents doing single tasks is still the pilot stage.

Execution — What infrastructure is missing

4/ WEF x Capgemini, AI Agents in Action 🔗 https://lnkd.in/gNHK-q-3

What it says: Practical framework for deploying AI agents at enterprise scale — agent design, integration, monitoring, and the human oversight model required to manage autonomous systems.

My take: The most operationally useful report on this list. The human oversight model section is worth your time specifically. It answers the question every program director is quietly asking: where do humans stay in the loop without becoming the bottleneck?

Governance — Where risk compounds

5/ Deloitte, The Agentic Enterprise 2028 🔗 https://lnkd.in/gv5i_8zp

What it says: By 2028, the enterprises that have governed agentic AI well will have a structural advantage over those that haven't. Deloitte outlines what that governance infrastructure looks like — agent identity, named ownership, permission frameworks, and audit trails.

My take: Read this last. Not because it is least important — it is the most important. Read it last because the governance conversation lands differently when you understand the scale of the opportunity it is protecting. The agent charter, the permission framework, the audit trail — these are not bureaucracy. They are how you build something that scales without breaking.

An agent charter is not a governance document. It is a commissioning document. One page. Written before the agent touches production.

Charter questions — fill these in before sign-off:

1. What is this agent authorized to do? List specific actions. Not categories — actions. "Access customer data" is not specific enough. "Read customer records in [system] for accounts flagged as at-risk" is.

2. What is it not authorized to do? The boundary is as important as the scope. Write it explicitly. Do not leave it implied.

3. Who owns it? Not the team. Not the vendor. A named person who reviews its outputs, answers for its failures, and has the authority to shut it down.

4. How do we measure whether it is working? Two metrics minimum: output quality and cost. One metric for failure: what does wrong look like, and who gets notified when it appears?

5. When does it get reviewed — and what triggers retirement? Set a review date before go-live. Define the conditions under which this agent stops running. "We'll look at it in six months" is not a charter. A specific date and specific criteria are.

For program directors: make the agent charter a standard part of your deployment checklist. It takes one meeting. It prevents the call the CIO made six weeks after go-live.

Every AI agent running in your organization right now was commissioned by someone.

Can you name them? Can you name what the agent is authorized to do? Can you name who is reviewing its output and when?

If the honest answer is no — you do not have an AI governance problem.

You have an operating system that nobody designed yet.

That is fixable. But only if someone owns the fix.

Who is designing yours?