AI Governance
One AI feature, one board page: what could go wrong, what it would cost, and whether it's handled
Mike Miller argues boards want three answers on AI: what could go wrong, what it would cost, and whether you have it handled. Here is how a HealthTech founder can answer all three for a single AI feature on one page.
November 12, 2026 · 6 min read

A board can act on three answers about AI, and a briefing on prompt injection gives it none of them. That is the opening claim of a recent X article by Mike Miller, a vCISO, who puts it this way: "Your board doesn't want to hear about prompt injection. They want to know three things: What could go wrong? What it would cost? Do you have it handled?"
The framework is his, and his article is written for security leaders briefing a board about AI across a whole company. I want to borrow it for a narrower job that HealthTech founders face constantly: explaining one AI feature, the one you are about to ship or the one a health system buyer is asking about, on a single page. The same three questions come from your own board, from investors, and from the governance committee on the customer side. A page that answers them once saves you from answering them badly five times.
What Miller actually argues
His article is titled "The CISO's AI Playbook," and its core point is that the job is translation: taking technical risk and turning it into business language a director with a finance or operations background can act on. He splits AI risk into two sides. One is attackers using AI, through deepfake fraud, cleaner phishing, and faster bug discovery. The other is a company's own AI use, through data pasted into unapproved tools and agents that get more access than their job needs.
He then gives four moves for the next board meeting: inventory where AI is used, write a policy people will follow, update incident response for AI-driven attacks, and report AI risk in dollars and plain English. His line on that last move is the one I keep coming back to. Drop "AI is a top risk" from the deck, he writes, because it tells the board nothing. Replace it with what could happen, what it would cost, and what you are doing about it.
Note the cost question. In his framing, "what it would cost" is the cost of the bad outcome, anchored on your own numbers: revenue, data volume, payment limits, industry. It is an exposure figure. I add a second cost line below, the cost to run the feature, because a founder's board will ask that too. That addition is mine, not his.
Why one feature at a time
Miller's inventory move asks a company to record, for every AI item, the data it touches and the actions it can take, and he says the second column is where agent risk hides. For a HealthTech company, the product itself is a large share of that inventory. Each AI feature has its own data, its own users, and its own failure mode.
A company-wide AI slide blurs those together. A per-feature page keeps them apart, and it maps directly onto how a health system reviews you. Their intake asks about one use case at a time. If your board page and their intake form describe the same feature with the same facts, you are not maintaining two stories.
The page, section by section
Here is the page I would write. One feature, one page, plain sentences.
- Feature in one line. What it does, who uses the output, and whether a person reviews every output, samples them, or none. Example: "Drafts patient message replies for nurse review. A nurse approves every reply before it sends."
- Data it touches. What it reads, what it writes, whether PHI is involved, and where it is processed. This is Miller's first inventory column, applied to one feature.
- Actions it can take. Read only, write, send, or trigger something in another system. This is his second column. If the honest answer is "more than it needs," say so and say when that changes.
- What could go wrong. Two or three concrete scenarios, written so a director can picture them. "A reply goes out with another patient's details" is a scenario. "Model hallucination risk" is a category, and boards cannot act on categories.
- What it would cost. For each scenario, an exposure estimate built from your own numbers, the way Miller suggests: how many records the feature can reach, which customer contracts carry notification duties, what an outage of the feature would do to a customer's workflow. Then the cost to run it: model spend, review time, monitoring. Give ranges, and say where the inputs came from.
- Whether it's handled. For each scenario, the control that reduces it, who owns that control, and how you would know it failed. A human approval step, a scope limit, a monitoring signal, a way to switch the feature off for one customer.
- What is still unknown. The gaps you have not closed, with a date. Boards trust a page that names its own gaps more than one that claims none.
- The ask. Miller ends every board update with a clear ask: budget, a policy approval, or acceptance of a named residual risk. Do the same. "Accept the risk that the feature is unavailable for up to a day during a model provider outage" is a decision a director can make.
Where founders usually get this wrong
The common failure is answering the third question first. Founders lead with the controls because that is the work they are proudest of, and the board never hears what the controls are for. Keep the order: scenario, then exposure, then control.
The second failure is a cost section with only operating cost. Model spend is easy to measure and easy to report, so it crowds out the harder estimate. A board that sees token spend and no exposure figure will assume the exposure is small, which may not be true for a feature that touches PHI.
The third is no residual risk. Every feature carries some. Naming it and asking the board to accept it is how you get a decision on record instead of a nod.
Start with the feature buyers ask about most
Pick the AI feature that comes up most often in health system security reviews and write this page for it this week. Keep it to one page, and reuse the facts from sections two and three in your next intake response. For the operating-cost line, the per-feature worksheet in Epic's AI bill is now variable, so every AI feature needs a value gate gives you the columns. Which of the three questions took you longest to answer for your own feature?