Boards have a dangerous blind spot on AI strategy. Decisions about which tools employees feed sensitive information into (and where company data actually goes once it leaves the building) are getting made at the procurement level and at individual desks. Boards aren’t in those meetings. By the time the resulting architecture reaches the boardroom, it’s presented in a deck as already set. As consequences are starting to show up, boards are increasing incentivized to act sooner than later.
What Gets Decided in Procurement Doesn’t Stay in Procurement
Many directors are still treating AI as a purely technology question for management to handle. But procurement-level AI choices are fiduciary choices. Customer trust will take the public hit when these calls go wrong. Vendor leverage and regulatory exposure risk are the longer tail. Procurement isn’t built to solve any of these.
In one headline example, “private” ChatGPT conversations turned up indexed in Google search results. Employees who pasted company content into the tool didn’t understand the privacy properties of what they were using (or didn’t really care, in the name of efficiency), and there is no reason to assume most of them ever read the data retention policies. Picture the same dynamic playing out at scale across an enterprise right now, which, of course, it is. Employees feed client data, contract drafts, internal financials and product road maps into systems whose memory and retention policies nobody has read or been warned about. Exposure compounds, quietly over months or years, before businesses will realize what’s been happening.
Boards, which are most concerned with how AI enables enterprises to scale faster and more cost-effectively, have either been slow to catch up or underinformed of the stakes. A Q4 2025 global survey of 772 board members and C-suite execs by Protiviti and BoardProspects found that only 26% of directors discuss AI at every board meeting. Two-thirds of directors in a recent Deloitte boardroom pulse survey said their boards still have limited or no AI knowledge or experience. Some of that is structural. Capability doubling that used to take 18 months on the chip side is now happening every three months on the model side, per METR. Refresh cycles and director education programs weren’t built for technology that moves at that rate.
The New Era of Duty of Care
I spent my operating career pushing for digital and direct-to-consumer adoption and transformation from the inside, in senior leadership roles at Petco and Wolverine Worldwide. From the operator chair, the pressure is always to move faster and to get the board’s air cover for doing it. So, I certainly have sympathy for what management wants from a board on technology decisions: mostly speed and trust that the team has thought it through. A board owns some of that, yes, but what it also owes back are the questions that aren’t getting asked in the room, starting with whether the architecture is quietly locking the company into one provider, and what actually happens if that provider changes terms next quarter.
Then, there’s the data itself. Directors don’t need a technical background to ask where the data physically sits as it moves through the company’s AI stack (or who has access along the way), or what gets retained about customers and operations after an AI query, and how long that retention lasts. Most director peers I’ve spoken to have never questioned any of it. But duty of care applies just as much to a Fortune 500 corporation as it does to a pet services franchise or any other business, and across my own board seats, the underlying questions have turned out to be the same in every room.
Underneath all of it is an even bigger question: Is the company running AI workloads with proprietary/sensitive data on shared public infrastructure that thousands of other firms (and competitors) are also using? Or is it operating inside a private environment it controls entirely? Either can be the right call. What matters more is that the decision is made deliberately, with the board in the room and the trade-offs visible. Most boards aren’t being shown the choice at all. These are architecture decisions with major ramifications (both good and risky) getting fast-tracked implicitly and through procurement, and procurement is being run on price and convenience, or worse yet, on naivete. The teams making these calls don’t know what they don’t know.
From my board seat at Iterate.ai, an AI platform focused on private deployment, I see a lot of enterprises come through that started having this conversation only after the architecture was already in place. In one example, a general counsel found out the contract review tool the legal team has been using for six months routes queries through a shared inference service. Or,in another case, a CFO found out the modeling assistant finance had been piloting kept queries for vendor training. Any one of these can be cleaned up, but across an organization with 30 or 40 AI touchpoints, the aggregate position adds up to an exposure profile the board never authorized. That lands AI squarely in the crosshairs of the board’s core enterprise risk management considerations, whether or not the ERM framework has caught up to it.
2028 Is Closer Than It Looks
I do believe the time to get governance into these decisions is running short, especially with AI agents now making decisions on their own and remembering things across sessions. Boards need to really understand what AI governance strategy is in place when you have AI pursuing goals after the user has logged off and wired into workflows nobody ever sat down to design. Quantum computing adds yet another layer to all of this, and the more credible forecasts place the erosion of current encryption somewhere between 2028 and 2032. AI architecture choices today will determine how exposed the company is when that day arrives.
Boards wading through this period in good shape will correlate to those asking more thorough architecture questions in 2026 (while there’s still room to change the answer). Those that wait for the first headline will find the architecture has already made the choice for them.