Most companies do not have an AI strategy. They have an AI budget, an AI vendor list, and a handful of enthusiastic teams shipping disconnected tools. That is not the same thing, and the difference shows up about eighteen months later as a pile of half-maintained integrations nobody can retire.
A strategy answers three questions: where do we apply this, what do we own versus rent, and how do we know it is working? If you cannot answer all three, you are running experiments — which is fine, as long as you know that is what you are doing.
Why "We'll Adopt AI Opportunistically" Fails
Opportunistic adoption feels efficient. Each team picks the tool that solves its own problem. The costs are invisible at first and structural later:
- Fragmented data. Six tools each build their own partial view of your customer. None of them is authoritative.
- Duplicated spend. Three departments pay three vendors for overlapping capability.
- Unmanaged risk. Nobody can answer which systems touch customer PII, or what happens when a vendor changes their model.
- No compounding. Point solutions do not make the next project easier. A platform does.
The strategic question is not "should we use AI." It is "what capability are we accumulating?"
The Four Layers of a Real AI Strategy
Layer 1 — Data readiness
This is unglamorous and it is where most strategies actually fail. Before AI can help, you need to know: where your data lives, who is allowed to see it, whether it is accurate, and how a system can access it programmatically.
A useful diagnostic: how long would it take an engineer to get read access to your last 12 months of support tickets, joined to customer records? If the answer is measured in weeks, that is your real bottleneck — not model selection.
Layer 2 — Capability mapping
Score your candidate workflows on two axes: volume × cost of error.
| Low cost of error | High cost of error | |
|---|---|---|
| High volume | Automate aggressively | Automate with human exception handling |
| Low volume | Assist, do not automate | Leave alone for now |
Most organisations should start in the top-left and earn their way rightward. Starting in the high-risk quadrant is how you get a board-level incident and a two-year AI freeze.
Layer 3 — Build vs buy
The honest framing is not build-or-buy but where does differentiation live.
- Buy commodity capability: transcription, OCR, generic summarisation, standard coding assistants. You will not out-engineer a focused vendor on these.
- Build where your data or workflow is the moat. If the value comes from your proprietary process, your customer history, or a domain model your competitors do not have, a generic tool cannot capture it.
- Own the seams regardless. Even when you buy, own your data layer, your evaluation harness, and your abstraction over the model provider. These are what let you switch vendors without a rewrite.
We go deeper on the economics of ownership in The Hidden Cost of AI-Generated Software.
Layer 4 — Governance that does not strangle
Governance has a bad reputation because it is usually implemented as a review board that adds six weeks to everything. Effective AI governance is narrower and faster:
- A register of what AI systems exist, what data they touch, and who owns them
- A required evaluation set for anything customer-facing
- Clear rules on what must never be fully automated
- A rollback plan for every deployed system
Four artifacts. Not a committee.
Connecting Strategy to Numbers
An AI strategy that cannot be measured is a mission statement. For each initiative, define the metric before you build:
- Support: resolution rate, handle time, CSAT on AI-handled tickets
- Sales: pipeline touched per rep, response latency, conversion on AI-assisted sequences
- Engineering: cycle time, escaped defect rate, review turnaround
- Ops: exception rate, cost per transaction
Then hold the line on the second metric — quality — because the first one is easy to game. A support bot that closes tickets fast by frustrating customers into giving up will look excellent on a dashboard.
Sequencing: What to Do First
The order matters more than the individual choices.
- Audit your data access. Not your data quality — your access. Quality can be improved incrementally; access is often a hard blocker.
- Pick two workflows in the safe quadrant. One customer-facing, one internal. The contrast teaches you a lot.
- Build the evaluation harness once, reuse it everywhere. This is your compounding asset.
- Establish the four governance artifacts while the stakes are low, not after an incident.
- Only then consider the high-risk, high-value workflows.
The Strategic Risk of Waiting
There is a symmetric error to over-investing, and it is more common in mid-market businesses: waiting for the technology to settle. It will not settle. What compounds is not the model — it is your organisation's ability to deploy, evaluate, and safely retire AI systems. That capability takes 12–18 months to develop regardless of when you start.
Companies starting in 2027 will have access to better models than companies that started in 2025. They will also be two years behind on the thing that actually matters.
If you want help turning a list of AI ideas into a sequenced plan with real metrics attached, get in touch — we do this with engineering teams, not just slide decks.