The Pilot Trap, Named Precisely
By 2027, almost every organization of size has run AI pilots, and a striking number are on their second or third round of them. The trap has a precise anatomy. Pilots are scoped as demonstrations: a use case is chosen for impressiveness, a vendor or internal team builds it, a demo lands well, and then nothing, because the pilot was never attached to a business outcome, an operating owner, or a measurement that survives the demo meeting. The organization concludes 'AI is promising but not ready for us,' re-runs the cycle next year, and calls the result a strategy. The technology matured through all of this; the governance did not.
The macro numbers make the micro pattern expensive. Adoption is mainstream, 78% of organizations use AI in at least one function per the Stanford AI Index, and investment is enormous, $109.1 billion of U.S. private AI investment in 2024 alone, yet enterprise surveys, including McKinsey's state-of-AI tracking, consistently find only a minority of companies scaling enterprise-level value from it. Gartner's companion prediction, over 40% of agentic AI projects cancelled by end of 2027 for cost, unclear value, or weak controls, is the pilot trap priced. The fix is not more technology ambition; it is portfolio discipline, and it fits on one page.
The Three-Horizon Portfolio
A workable AI strategy for business in 2027 is a portfolio across three horizons, funded and measured differently. Horizon one, assist, runs now: generative drafting, summarization, and research embedded in existing roles, cheap, low-risk, immediate, and its real product is organizational skill, the instruction-writing and checking habits every later horizon needs. Horizon two, automate, runs this year: agentic workflows on high-volume, verifiable tasks, document intake, status assembly, queue triage, each with an owner, a baseline, guardrails, and a weekly delta on minutes and errors. Horizon three, transform, is a shortlist, not a build: the two or three processes that would change the business if rebuilt AI-first, revisited quarterly with evidence from horizons one and two.
The portfolio rule that keeps strategies honest: no horizon-two item starts without a horizon-one skill base in the owning team, and no horizon-three commitment is made before horizon-two evidence exists. Sequencing beats ambition, the organizations cancelling agentic projects in 2027 are, overwhelmingly, the ones that inverted this order. The good news embedded in the rule: horizon one is nearly free and starts with training, not software, which means the first quarter of a serious AI strategy costs a rounding error and produces exactly the evidence the later quarters need.
- Horizon 1 (now): assist, generative drafts, summaries, research inside existing roles; product = skill
- Horizon 2 (this year): automate, agentic workflows on verifiable high-volume tasks; product = hours and errors
- Horizon 3 (shortlist): transform, AI-first rebuilds of 2-3 core processes; revisited quarterly, committed on evidence
- Sequencing rule: H2 requires H1 skills in the owning team; H3 requires H2 evidence
- Every item: one named owner, one baseline, one guardrail standard, one metric
The One-Page Strategy Outline
Strategy documents fail by accretion: forty slides, six priorities, no decisions. The outline that survives contact with a real organization fits on one page and forces exactly the decisions that matter. Outcome block: the two or three P&L-visible outcomes AI is expected to move this year, in the CFO's own units. Portfolio block: the three horizons with current items and one named owner each, if an item has no owner, it is not in the strategy, it is in the wishing well. Capability block: who gets trained, on what, starting when, with the free-first applied path as the default entry, this is the line item most strategies omit and the one that predicts success.
The remaining three blocks are the governance that keeps the page honest. Guardrail standard: the five first-run controls (scoped permissions, human checkpoint, audit logs, evaluation set, kill switch) applied to every deployment, no exceptions granted for enthusiasm. Measurement stack: baseline before deployment, delta after, reported in the same units as the outcome block, usage metrics are explicitly demoted. Decision cadence: quarterly expand / adjust / stop calls per item, with the stop call celebrated in the meeting, a strategy that cannot stop anything is not a strategy, it is a collection of hopes with a budget.
| Block | Contains | Failure it prevents |
|---|---|---|
| Outcomes | 2-3 P&L-visible targets in CFO units | Activity mistaken for progress |
| Portfolio | 3 horizons, one named owner per item | Ownerless pilots that never scale |
| Capability | Who is trained, on what, starting when | Platforms deployed onto untrained teams |
| Guardrails | Five controls, applied to every deployment | The incident that cancels the program |
| Measurement | Baseline then delta, P&L units only | Usage dashboards that convince nobody |
| Cadence | Quarterly expand / adjust / stop calls | Zombie pilots funded forever |
Source: Strategy outline from the adoption playbook
Build, Buy, or Train
The 2027 version of this decision has a clear default. Buy the platform: agentic AI is embedding into enterprise software at speed, Gartner's 33%-by-2028 projection, so generic capability arrives in the stack your organization already licenses, and building it in-house is a race against vendors you will lose. Build the thin layer: your company's specific workflows encoded as instructions, prompts, guardrails, and evaluation sets, this is where differentiation lives and it is cheap, measured in days of trained effort, not development quarters. Train the people: the chronically underfunded line that determines whether the first two pay off, and the cheapest of the three by an order of magnitude.
The train leg deserves its own arithmetic because it is where most strategies quietly fail. A platform deployed onto an untrained team produces demos and dashboards; the same platform on a team that can specify tasks, check outputs, and redesign workflows produces hours back every week. The cost side is trivial by enterprise standards, an applied program at $47 per person, one-time, anchored at $149, with a free 2-hour first day that proves the method before a single dollar of the full fee is committed. Against that, the typical alternative, unguided licenses plus hope, has produced the cancellation statistics cited above. Capability is the strategy; everything else is procurement.
Sources: Stanford AI Index 2025 · Gartner
Start Monday, Grounded in Reality
A strategy grounded in observed reality beats one grounded in vendor decks, and grounding is now a two-hour exercise. Day 1 of Leading AI Adoption is free, about 2 hours, no credit card: your managers and team leads register free, each maps one real workflow from their own team, and runs a guarded AI task against it, producing the raw material every block of the one-page strategy needs, real minutes, real error rates, real guardrail behavior. If the method fits, the 15-day program is $47 per person, one-time, anchored at $149, and produces the trained middle layer that the portfolio, guardrails, and cadence all assume.
Context that helps as you draft the page: the AI skills gap in 2027 analysis quantifies why the manager layer is the binding constraint, the AI change management playbook covers the human side of the rollout, and if you need the board version of the measurement story, the board presentation playbook covers framing numbers for that audience. Write the page after the first free day, not before, the order matters more than the template.
Frequently asked questions
▸Why do most corporate AI strategies stay stuck in pilots?
Because they are built as technology lists instead of business portfolios. A typical stalled strategy names fifteen use cases, funds none end to end, assigns no owner for the operating change, and measures nothing except 'usage.' By 2027 the distinction that matters is portfolio discipline: a small number of workflows, each with an owner, a baseline, a guardrail standard, and a P&L line it moves, spread across three horizons, assist now, automate next, transform later. Pilots are not the failure; staying in pilot is, and that is a governance choice, not a technology one.
▸What should a one-page AI strategy contain?
Six blocks: the two or three business outcomes AI is expected to move this year, stated in numbers the CFO already tracks; the workflow portfolio across three horizons with one named owner per item; the capability plan, who gets trained, on what, starting when (free and applied beats generic and expensive); the guardrail standard applied to every deployment; the measurement stack, baseline before, delta after; and the decision cadence, when expand / adjust / stop calls get made. If it does not fit on one page, it has not been decided yet.
▸Build, buy, or train, how do we decide in 2027?
The 2027 default is buy the platform, train the people, and build only the thin layer that encodes your company's specific workflows, prompts, instructions, guardrails, evaluation sets. Vendors are embedding agentic AI into a third of enterprise software by 2028 per Gartner, so building generic capability is a losing race. Training is the piece organizations chronically underfund even though it is the cheapest line and the one that determines whether the other two pay off: a $47-per-person applied program routinely outperforms six-figure platforms deployed onto untrained teams.
▸Where does a realistic AI strategy start on Monday?
With experience, not documents. Day 1 of Leading AI Adoption is free, about 2 hours: your managers register free, map one real workflow each, and run a guarded AI task, so the strategy conversation starts from observed reality, real minutes, real error rates, instead of vendor decks. If the method fits, the 15-day program is $47 per person, one-time, anchored at $149, and produces the trained middle layer every serious strategy assumes. A strategy written by leaders who have personally run the change is a different document.