There is a question forming in the room that few chief executives have yet asked aloud. What do I owe the machine? Not what may it do for me — what do I owe it.
The question is uncomfortable because, in the long history of the executive office, the verb has always run in the other direction. The machine — the steam engine, the spreadsheet, the enterprise system — was the thing that owed performance to its owner. In 2026, the verb has shifted. The chief executive now owes the machine an architecture, a standard, and an audit. Not because the machine asks. Because the obligation has formed regardless.
The Obligation, Named
A fiduciary duty does not announce itself. It accrues. Boards adopted the duty of supervising material risk before there was a vocabulary for it. Auditors adopted the duty of opining on the going concern before the standard was written down. The duty of the chief executive to the machine has accrued in the same way — quietly, through the accumulation of small events that, taken together, change what failure looks like.
The events of the last eighteen months have been quiet. An agent made an inventory-pricing call that breached a covenant a vendor had agreed not to touch. A model used in credit triage was found, on review, to be discriminating against a profile it had been explicitly told to set aside. A drafting assistant produced a contract that no one read until counterparty counsel did, and then the conversation was no longer about the contract but about the firm's procurement controls. In each case, the failure was not technical. The failure was that no one in the chief executive's reporting line could explain, in the presence of a board, what the machine had been authorized to decide.
The account is not a slide; it is an architecture. The discharge of the obligation is not a vendor selection; it is a doctrine. And the failure mode is no longer the machine that performs poorly. It is the machine that performs in a manner the firm cannot explain.
The Four Obligations the Chief Executive Now Carries
The duty resolves into four components. Each is owed by the chief executive in the first instance, and each is delegable only after it is named.
№ I — To Define What the Machine May Decide. The first obligation is taxonomic. The chief executive owes the firm a definition, in writing and ratified by the board, of the decisions the machine may make and the decisions it may not. The definition is not a security model. It is an authority model — the same kind of model the firm uses to define what a vice president may approve, what an executive committee must review, what a board must consent to. The machine's authority belongs in the same document.
The instinct, in 2025, was to delegate the question to the chief information officer. In 2026, that delegation is itself the failure mode. The authority of the machine is not an IT question; it is a governance question. The chief executive may delegate the implementation. The decision about what is delegable belongs to the chief executive.
№ II — To Name the Standard of Judgment. The second obligation is qualitative. For every consequential decision the machine is authorized to make, the chief executive owes the firm a standard against which the machine's judgment will be tested. Not an accuracy metric — a standard. The two are not the same. An accuracy metric tells you how often the machine is correct against a benchmark. A standard tells you what the firm considers a defensible decision when the question is not in the benchmark.
The standard is the senior partner's signature in a strategy report. It is the moment, in the diligence conversation, when a managing director says, we would not bring this to investment committee. It is qualitative, contextual, and inseparable from the firm's identity. Where the machine acts on the firm's behalf, the standard must travel with it.
№ III — To Own the Audit Trail of Consequence. The third obligation is forensic. Every consequential decision the machine makes — for a customer, for a counterparty, for an employee, for a regulator — must be reconstructable. Not the prompt. The decision. The chief executive owes the firm an audit trail dense enough that, in the presence of an unhappy customer or an interested regulator, the decision can be traced from the request, through the data, through the logic the firm chose, through the human who reviewed the consequential class, to the moment the firm shipped the result.
This obligation will, within twenty-four months, become the single most expensive infrastructure investment of the AI era. It will be expensive because most current systems were not designed to produce it. The firms that build it ahead of need will treat the next two years as the cheap part of the trajectory; the firms that build it under pressure will overpay.
№ IV — To Choose the Architecture Before the Vendors Choose for You. The fourth obligation is structural. Every consequential class of work the firm intends to deliver with the help of the machine sits on top of an architecture: the model, the orchestration layer, the data layer, the context the machine reasons over, and the layer the firm will eventually own outright. The vendor stack of June 2026 is, in its current form, a set of decisions that the chief executive will inherit if he does not make them deliberately.
The most consequential of these decisions is who owns the context layer — the corpus of accumulated organizational knowledge that the machine reasons over. The model is a commodity, fast and falling. The context is the moat, slow and compounding. The chief executive who has not decided, in 2026, who owns the context layer of the firm will, in 2028, discover that the choice has been made.
The Cost of Dismissing the Obligation
A senior chief executive of a financial-services firm — and the example is composite, as the discipline of this brief requires — described the AI question, in late 2025, as "below my line." He meant it not dismissively but procedurally: the chief information officer had it, the chief risk officer had it, the line businesses had it, and the chief executive's job was to hold the strategy of the firm against the question of where to compete.
The view was orthodox. It was also, within eighteen months, untenable. By mid-2026 the same firm could not produce, on the board's request, an inventory of the machines that were acting on customer accounts; could not state the standard against which their decisions were being judged; could not produce a single audit trail dense enough to defend an exception in front of a regulator. None of this was the chief executive's failure as a leader. It was the inevitable consequence of treating the question as a downstream operational matter rather than as an upstream fiduciary one.
The cost of dismissing the obligation is not a public scandal. The cost is the loss of an option. A chief executive who has not asserted the obligation cannot, in a crisis, claim it back from the operations of the firm. The architecture is set; the doctrine is implicit; the audit is fragmentary; and the firm is now governed, in this domain, by accretion rather than design.
The Shape of a Year Well-Spent
A chief executive who has decided to discharge the obligation in 2026 will do four things, each of them quietly, each of them in the firm's own register.
First, a policy. A short document, ratified by the board, that defines what the firm's machines may decide and on what authority. The document is not technical. It reads like a delegation matrix in a financial-services firm. Two pages, three columns: the class of decision, the level of human concurrence required, and the standard of judgment to which the decision is held.
Second, a doctrine. A longer document, written in the firm's voice, that names the standard of judgment for each consequential class. The doctrine is what travels with the machine when it acts on the firm's behalf. It is what a senior partner would tell a junior associate the firm believes; the machine receives it as a system instruction, the human reviewer holds it in mind during review, and both reference it in the audit.
Third, an audit. A program — not a tool, a program — to ensure that every consequential decision the machine makes is reconstructable. The program will require investment in the firm's data and context infrastructure. It will require an internal owner with sufficient seniority to commit the firm. It is the work most chief executives will be tempted to defer; it is the work that, eighteen months on, separates the firms that retain optionality from the firms that have lost it.
Fourth, a partner. A relationship — internal, external, or a deliberately constructed pairing — with a counterparty who can hold the standard of partner-grade judgment against the firm's use of the machine. This is the role our own firm exists to play; we mention it once, in the spirit of the document, and not again. The point holds independently of the choice of partner: the standard of judgment, held externally, does not travel inside the firm without an interlocutor who will not be made smaller by the machine.
A Closing Thought
The chief executives who will look back on 2026 with the cleanest conscience are the ones who decided, in the spring of the year, that the obligation was theirs. Not the chief information officer's, not the chief risk officer's, not the general counsel's — though each will have a role to play in its discharge. The decision to assert the obligation is the chief executive's first act of authority over the machine. The architecture, the doctrine, the audit, and the partnership all follow from it.
The machine, for its part, will accept the authority readily. It was, after all, built to be directed. The harder discipline is the one the chief executive imposes on himself: the discipline of refusing to delegate the question of what the firm is, in this new context, willing to decide and what it is not.
A thesis you cannot model is a hope. A decision you cannot authorize is a delegation in disguise. The chief executive who keeps both within the office of the chief executive in 2026 will discover, by the close of the year, that the duty was not an additional burden. It was the one the office was designed for. The machine has only made it visible.