I run a lot of AI strategy sessions with non-technical CEOs. The pattern is predictable: someone — usually the board — has asked them what their "AI strategy" is, they bought ChatGPT Enterprise for the company, and they're now trying to figure out what comes next. The deck-building consultants will happily sell them a 60-page strategy document. I think that's the wrong artifact.
The right artifact for an AI strategy in 2026 is a 90-day pilot with a measurable outcome and a budget line. Strategy decks describe; pilots prove. Here's the framework I use to get from "we should do AI" to a shipped result.
First: where AI actually moves the needle vs where it's hype
Not every problem benefits from AI. The categories where AI actually delivers ROI in 2026 are surprisingly narrow:
- Knowledge retrieval at scale — RAG (retrieval-augmented generation) over a body of internal documents. Customer support, internal knowledge bases, legal contract review. ROI: 40–60% reduction in time-to-answer, measurable in ticket-resolution time.
- Coding assistance — Copilot, Cursor, Claude Code. Engineering velocity gains of 15–30%, with the gains concentrated in boilerplate and routine refactors, not novel architecture work.
- Document generation — first drafts of proposals, summaries, reports. Cuts 50–70% of drafting time but still requires human review.
- Pattern detection in unstructured data — analyzing customer feedback, call transcripts, support tickets for trends. Replaces what used to be manual sampling with full-coverage analysis.
- Personalization at scale — email subject lines, product recommendations, onboarding sequences. Lift in conversion of 10–25%.
Where AI does not deliver in 2026 (despite vendor claims):
- "AI sales agents" that close deals. They generate leads; humans still close.
- Fully autonomous agentic workflows. The infrastructure works in demos, falls apart in production with messy real-world data.
- Replacing strategic decision-making. AI doesn't strategize; it executes patterns from training data.
If your AI strategy is built around the second list, your pilots will fail and you'll spend a year explaining to the board why "AI didn't work for us." Pick from the first list.
The 90-day pilot framework
The pilot has three phases, each 30 days. The point of the structure is to make a kill/scale decision at day 90 with real data, not vibes.
Days 0–30: scope and de-risk
Pick one use case from the high-ROI list above. Resist the urge to pick two. The discipline of one pilot at a time is what separates companies that ship AI from companies that talk about it.
- Define the measurable outcome in writing. "Reduce average customer-support ticket resolution time from 18 minutes to under 12 minutes" is measurable. "Improve customer experience" is not.
- Identify the incumbent process the AI will partially or fully replace. Document the current workflow, time-to-output, and cost.
- Pick build or buy. For 80% of pilots in 2026, the answer is buy — the off-the-shelf tools (Glean, Sierra, Dust, custom-RAG-on-Claude) are good enough and ship in days. Build only if you have a defensible data advantage (a proprietary dataset competitors don't have).
- Set the kill criteria in writing before starting. "If we haven't hit X by day 60, we kill it."
Days 30–60: ship to a small audience
The biggest mistake at this stage: launching to the whole company on day one. The infrastructure isn't ready, the prompts aren't tuned, the failure modes are unknown.
Instead:
- Pick 5–10 internal pilot users who'll use the tool daily.
- Instrument everything — usage, latency, satisfaction, output quality.
- Hold a weekly 30-minute feedback session with the pilot users.
- Iterate fast on prompts, retrieval logic, and UI. Most pilots that fail in this phase fail because the team treated the AI tool as a one-shot deploy instead of an iterative product.
Days 60–90: measure against the kill criteria
By day 60, you should have enough usage data to evaluate against the original metric. By day 90, you've either:
- Hit the metric — scale to a broader audience. Define the next 90-day expansion plan.
- Missed the metric but learned something specific — adjust scope, run a second 90-day pilot.
- Missed the metric and found nothing actionable — kill the pilot. Write up what you learned. Move on.
The kill option is non-negotiable. The companies that won at AI in 2024–2026 killed twice as many pilots as the companies that lost. The losers kept half-funded pilots running for 18 months because nobody wanted to admit defeat.
Build vs buy: the framework that actually works
Non-technical CEOs over-invest in build-vs-buy debates. Here's the decision rule that holds up:
Buy when:
- The use case is well-served by an existing vendor (most are in 2026).
- Your team doesn't have a 2+ engineer ML/infra capacity to maintain a custom build.
- The data you're using isn't a competitive moat.
Build when:
- You have proprietary data that constitutes a moat.
- Your use case has unusual constraints (regulatory, latency, format) vendors can't accommodate.
- You're at scale where the per-seat vendor cost exceeds 2–3 engineering salaries annually.
In practice, 80% of AI use cases at companies under 1,000 employees should be bought. Build is romantic; buy ships.
The org-design question nobody asks
The hardest part of an AI rollout isn't the technology. It's the org. Three questions to answer before launching pilot #2:
- Who owns AI usage policy? (Legal? IT? CTO? Centralized AI team?) Pick one. Distributed ownership leads to incoherent policies.
- How do we measure AI productivity gains? If you don't establish a baseline before deployment, you'll never be able to prove ROI to the board.
- What happens to the time saved? Reinvested in higher-leverage work? Removed via attrition? Used for new initiatives? The CEO needs to answer this; otherwise the productivity gain disappears into background noise.
These questions are upstream of the technology decision. CEOs who answer them in advance run cleaner AI programs.
What to put in front of your board
Most AI strategy decks are 60 pages of futurism. Replace yours with one slide:
- Pilot in flight: name of use case | metric | day-60 status
- Pilots completed: count of shipped + scaled / count of killed
- Investment YTD: dollars spent on AI tools and consulting
- Next 90 days: next pilot, why it's next
That's the AI strategy that survives contact with reality. The deck is for investors who haven't operated; the slide is for boards that have.
The AI strategies that ship in 2026 look like operational programs, not strategy documents. Pick narrow problems, ship fast pilots, kill aggressively, and scale what works. The non-technical CEOs who succeed are the ones who treat AI like any other operational rollout — measurable, time-bounded, with clear ownership. The ones who lose are still polishing the strategy deck.
For the SEO companion to this — how to think about AI's impact on B2B discovery — see B2B SEO in 2026.
