A credible enterprise AI strategy is a short set of decisions, not a vision statement. It says which use cases get funded and which get stopped, which data and platform choices are fixed for everyone, who answers for risk, how the money is released, and what happens in the first 90 days. If a sentence in the document cannot be turned into a decision someone can be held to, it does not belong there.
What a strategy actually has to decide
Most strategy documents fail because they describe an ambition and skip the choices that make it real. A useful test is whether the document would change a budget meeting.
Public guidance gives a reasonable picture of what a strategy covers. OMB Memorandum M-25-21, issued on April 3, 2025, tells US federal agencies what their AI strategies must include. The list is concrete: current and planned use cases that matter most to the mission, an assessment of present AI maturity, and plans for infrastructure across the AI lifecycle, access to quality data with traceability, enterprise capacity for AI, risk governance, workforce skills, and how future AI investment and procurement will be tracked. The memo is written for agencies, but the shape transfers to a company.
A company strategy can be reduced to six decisions:
- Where AI will be applied, and where it will not.
- Which data sources and platform components are shared across the company.
- Who is accountable for each AI system and for the portfolio as a whole.
- What level of risk the company will accept, and how that is checked.
- How funding is released and withdrawn.
- What the first 90 days produce.
What to leave out
A strategy gets longer when nobody is willing to decide, so the first editing pass is subtraction. Four things are worth cutting.
First, slogans. "AI-first" and "responsible by design" tell a team nothing about what to do on Monday. If a principle matters, attach a rule to it, such as a named approver for any system that affects customers or employees.
Second, a single maturity score. A score of 2.7 out of 5 does not tell anyone what to build next. A ranked list of use cases with the specific gap that blocks each one does.
Third, a technology roadmap detached from use cases. Platform spending that is not tied to at least two named use cases tends to drift toward whatever the loudest vendor is selling.
Fourth, a long list of ethical statements with no owners. The OECD AI Principles, adopted in May 2019 and updated in May 2024, are a sound reference point. They name inclusive growth, human rights and fairness, transparency and explainability, robustness and safety, and accountability. Citing them is easy. A strategy earns credit only when it says who checks each one, for which systems, and how often.
Build a use-case portfolio with kill criteria
A portfolio is different from a list of pilots. A list grows. A portfolio has a fixed capacity, so adding one item means removing or deferring another. That constraint is what forces prioritization.
Sort use cases into three groups. Run-the-business improvements change a process the company already performs, and their value can be measured against a baseline. Growth bets create a new product or channel and carry more uncertainty. Enablers build a shared capability, such as a retrieval layer over internal documents, that several other use cases depend on. Each group is judged differently, and mixing them in one ranking produces bad decisions.
Every entry needs five fields before it is funded:
- The business owner, who is a line executive and not the technology team.
- The decision or task the system changes, described in one sentence.
- The baseline measure today, and the target that would justify continuing.
- The data it needs, and whether the company may use that data for this purpose.
- The kill criteria, written before any money is spent.
Kill criteria are the part most strategies omit, and they are what separates a portfolio from a collection of hopes. They are conditions that end the work, agreed in advance. Examples include: the target metric has not moved by the end of a fixed pilot period; the cost per transaction in production exceeds the cost of the manual process; the review team overrides outputs more often than a stated threshold; a required data source cannot be licensed or cleaned within the planned window. Without them, the first review becomes a negotiation with a team that has already spent its budget.
The UK government's AI Playbook, published on February 10, 2025, makes a related point from the other direction. Its principle on using the right tool says teams should be open to the conclusion that AI is not the best solution, because a problem may be more easily solved with established technology. It also lists use cases to avoid, including fully automated decisions on significant matters such as health or safety. A portfolio needs the same discipline: a visible list of ideas that were considered and declined, with the reason.
Make the data and platform choices once
The expensive failures in AI programs usually come from each team making its own data and tooling decisions. A strategy should fix the decisions that are costly to reverse and leave the rest to teams.
For data, decide the following at company level:
- The small number of data domains, such as customers, products and contracts, that will be treated as shared assets with a named owner.
- The quality standard a source must meet before AI systems may use it, and who certifies it.
- The rule for traceability: for any production output, the team can show which data and which model version produced it.
- The legal test for reusing data collected for another purpose.
M-25-21 treats quality data and data traceability as an explicit part of an agency's plan, which is a useful signal. Traceability sounds administrative until the day an output is challenged and nobody can reconstruct how it was produced.
For the platform, decide the components that every AI system must pass through rather than the products you will buy. Typical examples are one place to log model calls, one method for evaluating output quality before release, one route for secrets and access control, and one way to switch a model provider without rewriting the application. NIST's AI Risk Management Framework (AI RMF 1.0, January 2023) asks for policies on third-party risk, including infringement of a third party's intellectual property, and for contingency processes for failures in third-party data or AI systems (GOVERN 6.1 and 6.2). A shared platform layer is where those policies get enforced in practice.
Govern by risk, and reuse what exists
Governance in a strategy should answer three questions: who is accountable, what is reviewed and by whom, and what happens when something goes wrong.
On accountability, NIST's framework states that executive leadership takes responsibility for decisions about risks from AI development and deployment (GOVERN 2.3). The UK Playbook asks for a senior responsible owner for each project. Both point to the same practice: a named executive per system, and a named executive for the portfolio.
On review, scale the effort to the risk. NIST notes that its framework does not prescribe risk tolerance and that tolerance is highly specific to the context and the use case (section 1.2.2). So the company has to set it. A workable approach is a short tiering rule. Systems that inform an internal task with a person checking the result get a light review. Systems that act on customers, employees or money get independent review before launch and scheduled review afterward. Systems that decide on someone's access to a job, credit, care or benefits are either excluded or run with a human who has the authority and the information to overrule.
Two small mechanisms carry a lot of weight. The first is an inventory of every AI system in use, with owner, purpose, data and risk tier. NIST lists an inventory mechanism (GOVERN 1.6) and the UK Playbook describes a live inventory that is kept current. The second is a decommissioning procedure. NIST includes processes for phasing out AI systems safely (GOVERN 1.7), and post-deployment plans that cover appeal, override, incident response and change management (MANAGE 4.1). Kill criteria only work if there is a defined way to switch something off.
Finally, do not build a new bureaucracy. M-25-21 tells agencies to use and scale existing governance tools and processes wherever possible to avoid creating additional bureaucracy. NIST similarly encourages treating AI risks alongside other critical risks such as cybersecurity and privacy. In most companies this means adding AI to the existing risk committee and the existing change process, with a clear threshold for escalation.
Fund in stages, and budget for running costs
Annual project funding suits work with predictable outcomes. AI use cases rarely have them. A better pattern is staged release. A small amount funds a time-boxed test against the baseline. A second tranche funds production build only if the kill criteria were not triggered. A third funds scale-up, and only after the run costs have been measured in real use.
Three budget lines are routinely missing:
- Run costs, including inference, monitoring, evaluation and the people who review outputs.
- Data work, which often exceeds model work and is rarely attached to the use case that needs it.
- Change management, meaning training and process redesign for the people whose work changes.
Fund the shared platform and data work in its own line. And keep a small reserve that the portfolio owner can move between use cases without a new approval cycle, because the best candidates will not be the ones you predicted.
The first 90 days
A strategy that does not specify the first 90 days will take six months to start. The UK Playbook observes that supporting structures do not need to be fully mature before a first project, and that experience from the first project will shape them. That is a sound instruction for private companies too.
A realistic sequence looks like this:
- Days 1 to 15: name the portfolio owner and a senior owner for each of the first use cases. Agree the risk tiers and the escalation route.
- Days 1 to 30: build the inventory of AI systems already in use, including tools staff adopted on their own. Rank the candidate use cases against the five fields above.
- Days 30 to 60: write kill criteria and baselines for the top three to five use cases. Settle the shared data rules and the evaluation method.
- Days 45 to 90: start staged pilots on the first two or three. Hold the first portfolio review and stop at least one item if the evidence supports it.
Where to start
Begin with a one-page decision record rather than a long document. Write down the six decisions listed earlier, and next to each one write the name of the person who owns it and the date it will be reviewed. Where you cannot name an owner, you have found the first thing to resolve.
Then run the use-case inventory. Ask every function to list the AI systems and pilots it has, including free tools used by staff, and capture owner, data, purpose and risk tier for each. This usually reveals duplicate spending and unreviewed systems within a few weeks.
Pick the first three portfolio entries and write their kill criteria before approving any budget. Do this first, because an agreed stopping rule is the clearest sign that a use case has been thought through. Then schedule the first portfolio review for day 90 and make it a real decision meeting, where at least one item can end.





