The 90-Day and First-Year Plan

The headline of the entire AI moment is an uncomfortable gap: everyone has AI; almost nobody has AI value yet. Around 88% of organizations now use AI in at least one function, yet only roughly 39% report enterprise-level EBIT impact, and most of those attribute under 5% of EBIT to it (McKinsey, State of AI, 2025). A separate study of enterprise pilots found about 95% delivered no measurable P&L return — and named the root cause not as model quality, infrastructure, or talent, but as a learning gap: the work of redesigning workflows and integrating AI into how the business actually runs (MIT NANDA, State of AI in Business, 2025).

The lesson for you is precise: scattered pilots are exactly how value fails to materialize. What separates the few firms capturing real returns is sequence and focus — building fluency, picking a small number of high-value problems, redesigning the work around them, measuring honestly, and scaling only what moves a number. This lesson gives you that sequence as a four-phase plan you personally can run.

The bottleneck to AI value is not technology. It is leadership, learning, and organizational change — which is precisely the part executives own.

Think of the journey as a leadership arc, not a project plan. Each phase has one dominant job, and you should not advance until its exit criterion is met.

Notice the shape: each phase narrows from broad orientation toward concentrated, measured value. This is deliberate — the firms that win pick 2-3 high-value vertical use cases and go deep, rather than running many shallow experiments (McKinsey, Seizing the Agentic AI Advantage, 2025).

The first month is about you and an honest mirror. Four moves:

1. Build personal fluency. Use the tools yourself, weekly. This is not optional polish — BCG's AI Radar 2026 found that deeply engaged leadership teams are roughly 12 times more likely to be top-5% AI value creators, and that around 72% of CEOs now identify as the primary AI decision-maker, about double the prior year. You cannot champion what you do not personally use. Fluency here means conceptual fluency: for any AI term, you can place it in the hierarchy (AI to ML to GenAI to LLMs to agents), say what business problem it addresses, and name its main risk. Knowing the math is not required.

2. Inventory current and shadow AI. You cannot govern what you cannot see. Map current pilots, spend, data readiness — and especially shadow AI, the unsanctioned tools employees already use. Roughly half of workers report using unsanctioned AI tools, and executives are often the worst offenders (CIO, 2025). Banning it drives it underground; the fix is a sanctioned, easy, safe alternative plus a clear policy. The inventory is your starting line.

3. Pick a business problem, not 'do AI.' There is no standalone AI strategy — only business strategy strengthened by AI (Wharton). Leaders who make 'doing AI' the goal generate diffuse pilots with no P&L impact. Name one concrete problem tied to growth, cost, customer experience, or risk.

4. Name an executive sponsor. Assign a single accountable owner for AI strategy and governance. CEO oversight of AI governance ranked as the attribute most correlated with EBIT impact among ~25 tested (McKinsey, State of AI, 2025) — yet only about 28% of organizations say the CEO owns AI governance and 17% say the board does. Diffuse accountability is the clearest 'leaders get it wrong' signal.

Month two converts orientation into a focused portfolio. Four moves:

1. Select 2-3 high-value vertical use cases on a value-by-feasibility matrix. Plot candidates by business value against feasibility (effort, data, risk). Bias toward the Reshape ambition level — re-engineering whole workflows, where BCG estimates the largest efficiency gains sit — and toward unglamorous high-ROI wins, not just marketing demos. Sequence quick wins first to build credibility, then transformational bets.

2. Assess data readiness. Proprietary data is the one moat rivals cannot buy, but siloed, ungoverned data is a liability: an AI agent cannot query what it cannot find. For each shortlisted use case ask: is the data it needs clean, governed, and accessible? Roughly 80% of the real AI work is data, governance, and workflow integration (MIT Sloan).

3. Stand up lean governance and an AI policy. Early stage, that is an AI Governance Committee — chaired by a CDO, CTO, or CRO with Legal, Compliance, Risk, Security, HR, and business leads, reporting to the CEO. Adopt the NIST AI RMF vocabulary (Govern, Map, Measure, Manage) as your shared language, and map to the EU AI Act and ISO/IEC 42001 where in scope. The non-negotiables: an AI policy that kills shadow AI, risk-based tiering (a chatbot is not a loan-approval model), and a system inventory.

4. Default to buy or partner. Buying or partnering succeeded roughly 67% of the time in one large study, versus about a third of that rate for internal builds (MIT NANDA, 2025). Decide per use case on two axes — value potential versus competitors, and differentiated data access versus vendors. Build only where AI is a true competitive differentiator; otherwise buy for speed and odds.

Month three is where most programs quietly fail — by bolting AI onto a broken process and measuring the wrong thing. Avoid both. Four moves:

1. Run a few end-to-end pilots WITH workflow redesign. This is the single biggest lever. Workflow redesign has the largest effect on EBIT impact from AI, yet only about a fifth of firms have fundamentally redesigned workflows; most simply bolt AI onto legacy processes (McKinsey, Rewiring to Capture Value, 2025). 'End-to-end' means you reimagine the whole task, not insert a chatbot mid-stream.

2. Design human-in-the-loop. The governing principle is automation for execution, humans for judgment. Decide where a person stays in the loop (approving each step) versus on the loop (an escalation point). For anything where a wrong answer is load-bearing — legal, financial, safety — a human check must sit before the AI acts.

3. Set a three-tier KPI baseline BEFORE you start. Measure across three tiers: (1) adoption/usage, (2) workflow efficiency (time saved, throughput, error reduction), and (3) business/P&L impact (revenue, cost, EBIT). The classic failure is stopping at tier 1 — and worse, mistaking vanity metrics (logins, seats, token counts, suggestion-acceptance) for value. Track quality, not just volume. Capture the baseline before go-live or you will never prove the lift.

4. Empower middle managers. Executives see strategic advantage while middle managers confront AI's flaws unsupported, which stalls adoption and costs money (HBR, 2026). Involve them in planning, reduce their admin burden, and build feedback channels — they see the workflow reality you do not.

The first year is about turning proven pilots into operating reality and killing what does not move a number. Five moves:

1. Redesign workflows around the winners. Move deliberately from Deploy (off-the-shelf tools, modest productivity gains) to Reshape (re-engineered functions, where most value sits) to Invent (new AI-enabled revenue) — BCG's three ambition levels. Double down on use cases that proved out; do not let proven pilots languish.

2. Reskill broadly. The near-term story is augmentation, not replacement. The WEF Future of Jobs Report 2025 projects a net positive jobs picture by 2030 with substantial churn, and finds that around 59 of every 100 workers will need reskilling or upskilling. Close the gap between trained leaders and the frontline — disengaged frontlines stall adoption.

3. Operationalize governance. Make it routine: a full system inventory, risk-based tiering, lifecycle gates with halt/rollback, and a cadence of quarterly risk reviews, dataset-lineage records, and serious-incident reporting with one accountable executive who can explain it to a regulator. Around 51% of organizations reported at least one negative AI incident in the prior year (McKinsey, 2025) — governance is a value-enabler that lets you ship faster because trust is engineered in, not a brake.

4. Set EBIT-linked targets. Graduate from activity metrics to the metric that separates high performers: % of EBIT attributable to AI. Expect patience — most organizations reach satisfactory ROI in two to four years, not months (Deloitte). Kill or reshape any pilot that cannot show a path to a number; project-abandonment rates have been climbing precisely because firms fund on FOMO without a hypothesis.

5. Role-model use as the champion — throughout. This is not a phase; it runs across all four. High performers are about three times more likely to have senior leaders actively championing and role-modeling AI use (McKinsey, 2025). Your visible, weekly, hands-on use is itself a lever.

You will not personally evaluate models or write policy, but you must ask the questions that expose whether the work is real. Keep these on a card:

The hardest and most valuable of these is the first: never accept a confident accuracy claim at face value. Watch for automation bias — over-trusting fluent output. Final judgment and accountability stay human.