Skip to main content

The first decision: your 90-day AI priorities memo for your own organization (and why you will revise it in shame later)

The short answer

The memo is your first command act

A 90-day AI priorities memo turns your AI systems inventory into a decision you own. It is the moment you stop collecting information and start being accountable for a choice, which is what Topic 0.1 said makes the expert. (see Topic 0.1)

What you will be able to do

  • Write a 90-day AI priorities memo for your own organization that names at most three priorities, each with a decision, a named owner, a deadline, and a falsifiable success criterion.
  • Distinguish a priority (a choice made among competing goods) from a wish list (everything anyone wants), and defend why your memo has a short de-prioritization list.
  • Build the memo directly from the AI systems inventory you produced in Topic 0.2, choosing which system to govern first by consequence, not by convenience. (see Topic 0.2)
  • State for each priority the single result that would prove the decision wrong, so the memo can be tested rather than admired.
  • Classify each decision as a reversible ("two-way door") bet or a hard-to-reverse ("one-way door") commitment, and explain why a 90-day memo should favor reversible bets.
  • Assign clear accountability: every priority in the memo has one named owner and a date, because a decision with no owner and no date is a hope, not a decision.
  • Anticipate the attack: predict which line of your own memo a hostile reviewer will hit first, and pre-answer it in the document.
  • Diagnose a weak memo by recognizing the five failure patterns (the everything memo, the mood memo, the orphan memo, the undated memo, and the silent-tradeoff memo) and naming the fix for each.
  • Connect the memo to your accumulating dossier: explain where this artifact is attacked (Module 11) and revised (Module 13), so you write it today as a first-class record, not a throwaway exercise.

The ideas, one by one

Two or three priorities, never ten

A priority is a choice made among competing goods. A list of ten priorities is a confession that nothing was chosen. If the memo does not fit on one page, you have written a wish list, not a decision.

Every priority carries three marks

Falsifiable (it names the result that proves it wrong), owned (one named person, never a committee), and dated (a real date within ninety days). A line missing any mark is a mood, not a decision.

Rank by consequence, not by volume

The first system you govern should be the one whose failure reaches real people and real law, even if leadership is louder about a lower-risk system like a chatbot. Consequence over convenience. (see Topic 5.3)

The de-prioritization list is the real memo

Naming what you will not do this quarter, with the risk you accept by waiting, is what makes your priorities real and your choices defensible when they are attacked. An empty "not doing" list means you have not chosen.

Favor two-way doors

In ninety days you are still learning your AI estate. Prefer reversible bets you can undo cheaply over one-way-door commitments (long contracts, public promises) that lock you in before you understand them.

Expose the memo to attack before a critic does

Predict the sharpest objection to your own memo and pre-answer it on the page. The line you judge weakest is the line Module 11 is most likely to hit. (see Topic 11.1)

Revision is the goal, not the shame

A real decision commits you, then teaches you. IBM's 2023 projection was revised by 2025 reality, and the organization was better informed for it. The mild wince of rereading your first memo is the felt sensation of having learned something checkable.

The artifact persists

This memo is not an exercise you discard. It comes back in Module 11 to be attacked and in Module 13 to be revised as the opening page of your command dossier. Keep it exactly as written today. (see Topic 13.1)

A falsifier is confidence, not doubt

Naming the result that would prove you wrong does not make you look uncommitted; it makes you trustworthy, because only someone sure of their reasoning is willing to say in advance how they would know they were wrong. The memo that hides its failure conditions is the timid one.

Never carry a vendor's number as your own fact

If a priority rests on a supplier's savings or accuracy claim, make verifying that claim the falsifier itself, and measure real success against your own data once the tool is running. Borrowed numbers are how defensible memos become indefensible under scrutiny.

Listen to it

The same lesson, talked through by two hosts as episode 4 of the podcast.

Real cases

These examples show real organizations making, or failing to make, a first public AI-priorities decision, with the reasoning stated plainly. The centerpiece is IBM; the others are referenced from their owner topics for contrast.

Example 1: IBM's 2023 hiring-pause decision (the anchor). In a Bloomberg interview on 1 May 2023, IBM chief executive Arvind Krishna said the company would pause or slow hiring for back-office, non-customer-facing roles that AI was likely to absorb, and that he could "easily see" about 30 percent of those roughly 26,000 roles, near 7,800 jobs, replaced by AI and automation over five years, largely through not backfilling departures rather than through layoffs (Bloomberg, 1 May 2023). Read it as a memo. The decision was concrete (pause and slow specific hiring). The scope was named (back-office, non-customer-facing). The timeframe was dated (five years). The claim was falsifiable (a checkable percentage). And it carried an implicit de-prioritization and reinvestment: pulling back in back-office meant pushing elsewhere. Note the epistemic honesty of the original framing: "could easily see" is a projection, not a guarantee, and the program marks forward-looking claims that way.

Example 2: The same decision, revised (why "in shame later" is a feature). By its May 2025 Think conference, IBM described the outcome, and it was messier and more instructive than the 2023 headline. Krishna said AI, through the company's AskHR agent, had automated about 94 percent of routine HR tasks and replaced "a couple hundred" HR roles, far fewer than the 7,800 the headline implied, while total IBM employment actually rose, because the resources AI freed were spent hiring more programmers and salespeople; IBM also cited about USD 3.5 billion in productivity improvement across more than 70 areas (Krishna, WSJ-reported, May 2025). The direction of the 2023 decision held; the specifics were revised by reality. This is the model in Section 3G, beat 4: a real decision, publicly made, honestly revised. The 2023 memo was not wrong to write; it was right to write and then right to update.

Example 3: A decision that failed for lack of a falsifier and an owner (contrast). When an organization deploys AI with no dated success criterion and no single owner, there is no moment at which anyone must say "this is not working." The Michigan MiDAS unemployment system ran automated fraud decisions for years with inadequate human oversight before the scale of wrongful accusations forced a reckoning; a clear falsifier ("if human review overturns more than a set share of flags, we stop") and a named owner would have created a stopping point long before the harm compounded. That case is examined in depth as the build-buy-with-real-budget decision. (see Topic 3.2) The lesson here is narrow: a priority with no falsifier and no owner cannot fail on schedule, so it fails slowly and expensively instead.

Example 4: Choosing which system to govern first (consequence over convenience). Regulators reward organizations that put their highest-risk AI at the front of the queue. Under the EU AI Act (Regulation (EU) 2024/1689), certain uses such as AI in hiring and worker management are treated as high-risk and carry the heaviest obligations; a memo that governs a flashy customer chatbot first while leaving a high-risk hiring tool ungoverned has ranked by visibility, not by consequence. Risk classification is the owned subject of a later topic. (see Topic 5.3) For your memo, the takeaway is the ranking rule: the first system you govern should be the one whose failure reaches real people and real law, even if it is not the one your leadership finds most exciting.

Example 5: The honest-ROI check behind a headline number (from the money module). Bold public AI numbers invite scrutiny of the economics behind them. When a company claims a large cost cut from replacing staff with AI, the defensible version of that claim shows the real, fully-loaded costs, including the human supervision the AI still needs. The honest reckoning behind such claims is the owned subject of the ROI topic. (see Topic 8.1) For your 90-day memo, borrow one habit: if a priority rests on a savings or productivity number, write the number you expect and the number that would prove you wrong, from your own accounts, not a vendor's slide.

Example 6: The workforce priority as a decision, not a slogan (from the humans module). IBM's decision was, at heart, a workforce decision: which roles change, which grow, which end. The defensible way to make a workforce-AI priority is the same as any other: name the change, own it, date it, and state what would prove it wrong, rather than issuing a vague "we will reskill our people" that no one can check. The deep work of mapping which roles change belongs to a later topic. (see Topic 9.1) The borrow for your memo: if one of your priorities touches people's jobs, treat it with more falsifiability, not less, because the human stakes make a vague promise both crueler and less defensible.

Example 7: Sequencing under a real deadline. Priorities also have to respect the clock the outside world sets. Under the EU AI Act (Regulation (EU) 2024/1689), the timeline was itself revised: the Council of the EU's "Digital Omnibus on AI" simplification package, formally adopted 29 June 2026, deferred application of the stand-alone high-risk obligations to 2 December 2027 (Council of the EU press release, "Artificial intelligence: Council gives final green light to simplify and streamline rules," 29 June 2026). That is a live reminder that even the deadlines you plan against can move, weeks before you expected. A memo that sequences its priorities against a fixed external date should note that the date is current-as-of-today and may shift, and should prefer reversible first steps so a moved deadline does not strand an irreversible commitment. The specific current dates and their full implications are owned by the law modules; here the lesson is only that a good memo dates its assumptions, not just its tasks.

Where people go wrong

  • "A priorities memo should be comprehensive." Wrong, and the most common failure. A memo that lists ten priorities has made zero choices. Comprehensiveness is the enemy of a priorities memo; its whole value is the ruthless cut to two or three. If it does not fit on a page, you have written an inventory or a wish list, not a priorities memo.
  • "We should wait until we understand AI fully before committing to priorities." This is the information trap Topic 0.1 warned against. (see Topic 0.1) You will never understand AI "fully"; the field moves faster than any study. The memo is deliberately short-horizon (ninety days) and reversible precisely so you can decide before you are certain and learn from the decision. Waiting for certainty is a decision too, usually a bad one.
  • "A decision without a named owner is fine if the team is aligned." No. Shared ownership is no ownership. "The AI committee owns it" means that in ninety days, when you ask what happened, everyone points at everyone. One named owner per priority is not bureaucracy; it is the difference between a result that someone is answerable for and a result that quietly does not happen.
  • "The success criterion can be qualitative, like 'improved efficiency.'" A criterion you cannot measure cannot be failed, and a decision that cannot fail cannot teach. Use a number from your own organization's data and pair it with the result that would prove you wrong. "Improved efficiency" is a mood; "median handle time drops by our target by day 60, or we stop" is a decision.
  • "Revising the memo later means the first one was a failure." Backwards. Revision is the goal, not the embarrassment. IBM's 2023 projection was revised by 2025 actuals, and the organization was better for it (Bloomberg 2023; WSJ-reported 2025). A memo you never revise is a memo you never checked against reality. The mild wince of rereading your first draft is the feeling of having learned something checkable.
  • "Put the most visible or most requested system first." Rank by consequence, not by volume of requests. The system your leadership asks about most is often the least risky (a chatbot), while the real exposure sits in a quiet high-risk tool (hiring, credit, fraud). Governing the loud one first feels responsive and is often a misallocation. (see Topic 5.3)
  • "The de-prioritization list is optional." It is the most important part. Without it, every priority is under constant pressure to expand, and every choice not to do something is undocumented and therefore indefensible when attacked. If your "not doing" list is empty or contains only things nobody wanted, you have not actually chosen.
  • "A public, bold statement is a stronger decision than an internal one." Not necessarily, because public statements are closer to one-way doors. A public AI promise is hard to unsay when reality shifts, as IBM found. For a 90-day memo, prefer reversible internal bets you can update quietly. Save the public commitment for a decision you have already tested and are ready to defend.
  • "The memo replaces the strategy, the policy, and the risk register." No. The memo sits above those and points to what comes first; it does not do their job. Your policies, risk assessments, and conformity files are longer, later artifacts. The memo's job is only to decide where your scarce attention goes this quarter. Trying to make it comprehensive turns it back into the wish list it was meant to cut through.
  • "If I write a proves-wrong condition, I am promising to fail." A falsifier is not a prediction of failure; it is a pre-commitment to honesty about the result. Most priorities with good falsifiers succeed. The falsifier simply removes your future self's ability to rationalize a bad result into a "qualitative win." You are promising to look, not promising to lose.
  • "Once signed, the memo should not change for ninety days." The signature dates a decision; it does not freeze the world. If a serious incident or a legal change lands on day 20, you update the memo and note why, the same way a navigator corrects course mid-voyage. What you do not do is quietly relax a falsifier because a priority is going badly; that is editing the scorecard, not the plan.

Questions people ask

What is 90-day AI priorities memo?
A one to two page decision document, written by a governance leader for their own organization, that names the two or three AI priorities for the next ninety days, and for each states the decision, a single named owner, a date within the ninety days, and a falsifiable proves-wrong condition, plus an explicit de-prioritization list and a dated signature. It is the artifact this topic produces; it is attacked in Module 11 and revised in Module 13.
What is priority?
A choice made among competing goods, ranking one good above another with a stated reason. A true priority names what it beats. A list of many "priorities" with nothing given up is a wish list, not a set of priorities.
What is decision (versus a mood)?
A commitment that can be checked. A decision carries three marks: it is falsifiable (it names the result that would prove it wrong), owned (one named person or role is accountable), and dated (it has a real deadline). A statement missing any mark is a mood: it sounds like a commitment but cannot be tested, owned, or timed.
What is falsifiable (proves-wrong condition)?
The single, measurable result that would show a decision was a mistake. A priority without a falsifier can never fail and therefore can never teach; the falsifier is what turns a plan into a test.
What is owner?
The one named person or role accountable for a priority. Committees and "the team" are not owners, because shared ownership means no one is answerable at review time. Ownership is about having a single person to ask "what happened?", not about assigning blame. More on Owner

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →