The First 90 Days as Head of AI: A Week-by-Week Operating Plan

Monday morning. New title — VP of AI, Chief AI Officer, or some variation of "please make us AI native" stapled to your previous job — and an empty calendar. The CEO gave a speech at the all-hands. The board slides from last quarter promised "significant AI investment." Your inbox is already full of vendor introductions forwarded by the CRO. This is the highest-risk window of your entire tenure, and most people in this seat do not survive it.

The operating plan that works looks almost nothing like the one your predecessor probably ran. It starts with listening, not announcing. It builds a financial baseline before it picks a pilot. It finds one real shipped result instead of chasing the project that sounds best in a board deck. And it treats the CFO, the CISO, and the General Counsel as allies — because those three people will either defend your second year or end your first one.

Roughly 95% of AI pilots fail to reach production with measurable P&L impact.^[3] Between 70 and 85% of GenAI deployments miss their ROI target.^[4] The leaders running those programs were not incompetent. They made five specific, predictable mistakes inside the first 90 days, and by the time the mistakes surfaced, the political capital required to fix them had already been spent. IBM's 2026 CEO study found that 76% of organizations now have a CAIO — up from 26% a year earlier.^[9] The role is proliferating faster than the knowledge of how to survive it.

One diagnosis up front: this is not a technology problem. The tools exist. The models are good enough. What kills transformation mandates is organizational friction — misaligned incentives, missing financial baselines, and political capital burned on the wrong first bet. The technical decisions in the first 90 days are almost never what determines the outcome. The relationship and ownership decisions are.

Most first-year transformation failures do not announce themselves. They compound quietly — a vendor relationship that distorts your priorities, a pilot that burns political capital faster than it produces results, a financial conversation you keep deferring. By the time the problem is visible to the board, the goodwill required to fix it is already gone.

The five failure modes below are not hypothetical. They are the patterns that show up over and over when transformation leads get replaced in year one. Each one has a tell, a structural cause, and a countermove. Most of them are invisible at day 30 and obvious by month six — which is the trap. You do not get to learn from month six if the credibility you needed to survive it has already been spent.

Note where the failure originates: it is almost never the model, the tool, or the technical approach. It is the political and financial architecture around the work. IBM's research found that organizations with clear AI governance structure saw operating margin growth averaging 22% above market.^[9] Governance is not a constraint on transformation — it is the mechanism that makes transformation survivable.

The decisive stakeholders are the HRBP, the General Counsel, and the CFO.^[8] The CEO hired you and will lose interest in the details inside 60 days — that is not a criticism, it is how executive attention budgets work. The CIO may read you as a territorial threat, particularly if AI tooling previously sat under their remit. The CISO becomes your strongest governance ally if you treat them as a partner rather than a checkpoint, and your fastest-moving adversary if you do not. The General Counsel cares about IP ownership, vendor data clauses, and liability exposure from generated content. Most transformation leads meet GC in month four when a contract needs signing. Meet them in week one instead.

Meet every person on this list before you have anything to announce. The listening posture is not theater — it is intelligence collection that determines the entire plan. Each of the nine people below can independently kill the program. Understanding what they want and what they fear before you have positions to defend is the only way to design an approach that does not collide with all nine of them at once.

The temptation in week one is to announce a strategy. Resist it. You do not have enough information yet to announce anything credible. The mandate, the title, and a room full of people watching to see whether you are here to solve their problems or to add to them — that is what you have. An early strategy announcement without the listening tour behind it gets read instantly by the people who live in those workflows every day, and they disengage before the program starts.

The first 30 days produce three artifacts: a financial baseline that names what the organization is actually spending on AI today (including the parts that never hit the IT budget), a shadow AI inventory that maps the tools people are running without authorization, and a shortlist of five workflows that already have the shape of good quick wins. Everything else is secondary.

Thirty days feels short. It is short. The discipline is to resist the pull toward action — toward announcing, planning, hiring, launching — and stay in collection mode until you can be specifically right about something. The leads who fail at this stage usually fail not because they were impatient but because the organization pressed them to perform. Hold the line.

On shadow AI specifically: roughly half of employees at most organizations are already using unsanctioned AI tools.^[11][12] That is not a compliance problem you need to solve in month one. It is a market signal — the clearest data you will get about which workflows people want automated and which tools they trust enough to pay for themselves. Treat it accordingly.

The single most important thing between day 31 and day 60 is to ship something real people use. Not a demo. Not a pilot that requires a dedicated team to operate. A workflow improvement that is live in production, used by at least 20 people, with a number attached to it.

The moonshot is the trap. The leader who pitches an end-to-end AI customer experience as their first move gets replaced before it launches. The leader who ships an AI search tool over the knowledge base in week six, reports four hours saved per rep per week, and uses that number in every subsequent conversation gets a second 90 days. Credibility compounds. Pick the win you can actually ship.

The team you build inside this window matters as much as the pilot. By day 60, you need at least one person who can write and deploy code — not just manage a vendor. If your entire team is program managers and strategists, you do not have an AI program; you have a consulting engagement. Research from MIT found that pilots blending internal AI specialists with external expertise achieved a 67% success rate versus 22% for IT-only builds. The platform engineer is the hire that creates the conditions for that ratio. Internal transfer, external hire, or a two-month secondment from engineering — get them in seat before you commit to pilots two and three.

On instrumentation: Google's internal AI deployments demonstrate what disciplined measurement produces. Their sales intelligence AI achieved a 14% increase in lead-to-opportunity conversion in six weeks; their marketing campaign agent saved 18,000 hours in 2025.^[10] Both results were measurable because instrumentation was built before launch, not retrofitted afterward. Build it before you ship.

The difference between a pilot that earns a second cycle and one that dies quietly is almost always measurement clarity — not model quality. You need three numbers before you ship anything: a pre-intervention baseline, a post-intervention measurement, and a sample size large enough to be credible. Everything else is commentary.

Here is the minimum instrumentation you can stand up in a day for a knowledge-base search pilot, a meeting summary tool, or a contract extraction workflow. The code is intentionally simple. The goal is shipping measurement, not a perfect analytics stack.

By day 61, you have something you did not have on day one: a number. One thing deployed, one metric to point at, at least one stakeholder publicly endorsing what shipped. That is the moment to publish policy. Not before. A policy published without a track record reads as the no-department arriving. A policy published after a visible win reads as the organization growing up responsibly.^[6][7]

Good AI policy in 2025 and 2026 is not primarily a prohibition list. It is an enablement document — what employees can do, what data they can use, how to try something new through a legitimate sandbox, what to do when something breaks. The organizations getting this right ship policy under five pages with an approved tools list and a lightweight process for adding new ones. The organizations getting it wrong publish 40-page frameworks that nobody reads, and an informal routing-around economy emerges in every business unit.

Only 37% of organizations have formal AI governance policies in place.^[11] That gap is not a competitive advantage for the ones without one — it is a liability that surfaces the moment a shadow AI tool causes a data incident. IBM's research found that companies with clear AI governance structure saw operating margin growth averaging 22% above market.^[9] Governance is the program's durability mechanism.

The cadence you set in days 61–90 — the quarterly review, the weekly dashboard, the board cadence — is what converts a 90-day sprint into a durable program. Most mandates fail at this transition because the energy of the first 90 days does not naturally convert into governance discipline. Build the structure explicitly, before the sprint energy fades.

The right quick win has three properties that have nothing to do with how technically interesting it is. It produces a measurable result — a specific number, not a 'positive user response' — inside 30 days of going live. It sits in a workflow real people use every day, not a workflow that exists to feed a demo. And it generates a story you can repeat: 'we shipped X, it saved Y hours per week, Z people use it.' That story structure is the political infrastructure of a working transformation program.

The bad quick wins below fail for a consistent reason: they touch the organization's edges instead of its daily work. An AI chatbot for HR sounds high-impact. The HR workflows employees actually interact with — benefits questions, policy lookups, onboarding paperwork — are low-frequency and high-stakes enough that errors erode trust faster than the tool builds it. The good quick wins are boring by comparison. AI search across an internal knowledge base. Meeting summaries. Contract clause extraction. Boring tools used every day generate more political capital than impressive tools used when someone remembers to.

What we got wrong initially: we picked quick wins on ease of implementation. The criterion that actually matters is ease of measurement. A technically harder tool with an obvious metric (time-per-contract before and after) beats a technically simple tool where nobody agrees what success looks like. Measurement clarity is what converts a shipped tool into political capital. Without it, the CFO shrugs and asks what it cost.

Most first board briefs on AI are 20 slides of market context, competitive benchmarking, and technology diagrams. The board has seen that deck from three other executives this year. What earns trust at day 90 is specificity and honesty about what is not working.

Slide 1 — Where We Stand: one financial baseline number (current AI spend, normalized). One shipped result (the quick win, with the metric). One organizational health signal (adoption rate of the first pilot). No projections yet. The discipline of refusing to project at day 90 is counterintuitive — every instinct says show ambition. Projections at day 90 are guesses dressed in numbers. The board knows this. Real data from a shipped result carries more weight than a three-year revenue model built on assumptions.

Slide 2 — The Three Bets: three pilots, each with a named BU owner, a committed timeline (specific dates, not quarters), and a success metric that can be verified externally. Resist the urge to list six. Three focused bets with named owners are more credible than six ambitious ones without ownership. The named owner is non-negotiable — a bet without an internal champion is a consulting engagement, not a transformation.

Slide 3 — The 12-Month Risk Map: three risks, ranked by likelihood and impact. At least one of them should be something the board did not already know. Regulatory exposure is expected. The risk that surprises them — a specific vendor dependency, a data quality problem, a skills gap surfaced in the listening tour — is the signal that you have an accurate view of the program. Boards that hear only good news stop trusting the person delivering it. Boards that hear one real risk they had not considered start paying attention differently.

Every CAIO faces this decision in the first 60 days: build a central AI team that owns all projects, or stand up a hub-and-spoke model where each BU has an embedded AI lead with the central team providing standards, tooling, and governance. The answer is almost always hub-and-spoke — but the reasoning matters more than the label.

A fully centralized team is faster to spin up and easier to govern. It is also the model that loses BU buy-in the fastest. When a BU head has to queue a request through your central team and wait six weeks for a pilot that addresses their specific workflow, they start their own program. That parallel program becomes the shadow AI problem you are now managing instead of the transformation you were hired to run.

IBM's 2026 CEO study found that hub-and-spoke AI operating models yield 36% higher ROI than fully centralized approaches.^[9] The mechanism is not mysterious: BU-embedded leads understand the domain. They catch the workflow edge cases that a central team — however talented — discovers only after a painful production incident.

The minimum viable hub-and-spoke in the first 90 days does not require headcount you do not have. Identify one person in each of the three largest BUs who is already obsessive about AI tools. Formalize the relationship: they are your AI council. They get early access to new tools, they bring you pilot candidates, and they own adoption for their BU. In return, they get the central team's tooling, governance cover, and a line into the vendor conversations. That is enough to start.

The first 90 days do not decide whether you win. They decide whether you get a second 90 days — and a second, and a third, until enough compounding wins exist to survive the inevitable quarter where something important does not ship on time. The math is straightforward: most failed transformations did not fail because the technology was wrong. They failed because the credibility required to make hard decisions in months four through nine was never built in months one through three.

The leaders who lose this role do not lose it because AI is hard. They lose it because they confuse the mandate with the trust. The mandate arrives on day one. The trust is earned stakeholder by stakeholder, week by week, number by number. The CFO who understood the financial baseline from week one becomes the person who defends your budget in the room you are not in. The CISO who co-authored the policy becomes the person who says yes to the tool that would otherwise have taken three months of review. The BU head who co-owned pilot one becomes the person who brings you pilot four.

Build the financial baseline first. Everything else follows from that.