How to build an AI strategy: decisions, not a tool list
An AI strategy is a set of decisions about workflows, ownership, evidence and stopping, and this guide walks through them in the order I would take them.
In McKinsey's 2026 global survey, 80 percent of respondents said AI had improved their individual productivity, while only 37 percent attributed any EBIT impact to their organization's use of it. Just 6 percent qualified as high performers. The distance between those numbers is the problem an AI strategy exists to solve.
An AI strategy is the set of decisions a company makes about which workflows it will change with AI, who owns the result, what evidence will count as success and when it will stop. A list of tools, vendors or use cases can sit inside a strategy. It is not the strategy.
I have written about each of these decisions separately. This essay puts them in the order I would take them, and links to the longer argument where one exists.
Why do most AI strategies produce adoption but not value?
Most AI strategies treat adoption as the unit of change, so they measure whether people touched the technology and not whether the work changed. Deloitte found that 48 percent of organizations had introduced AI without redesigning the workflows or roles around it, and only 12 percent reported redesign at scale. Stanford's Enterprise AI Playbook, which studied 51 successful deployments, found that 77 percent of the hardest implementation challenges were change management, data quality and process redesign, not model problems.
In Stop adopting AI I argued that the unit of change has to move from tools and tasks to workflows, decisions and the operating model. Employees are often blamed for the gap. Microsoft's 2026 Work Trend Index surveyed 20,000 AI users and found that only 26 percent said leadership was clearly and consistently aligned on AI, and only 13 percent said they were rewarded for reinventing work even when an experiment missed its targets. People are asked to change their work while being measured by the rules of the old work.
An AI strategy is a list of decisions the company is willing to be held to.
Where should an AI strategy start?
It should start with a business outcome and the constraint that prevents it, and it should leave the answer open. An AI-first mandate does the opposite. If every function must submit AI use cases, then the biggest AI mistake companies make is starting with AI: the solution is fixed before discovery begins, and discovery turns into justification. A policy change, a simpler form or the removal of an approval never enters the competition.
My rule is that every serious AI investment faces three credible options: an AI-enabled intervention, a non-AI intervention, and deferral. A sponsor should be able to name the strongest non-AI alternative and explain why AI should beat it. Before committing delivery capacity I would ask for a five-line brief covering the outcome, the constraint, the alternatives, the observation that would make AI preferable after full operating costs, and the owner with the stop rule.
There is a second filter after the first. An AI program can be efficient and still be strategically irrelevant, because AI also makes low-value work cheaper. The question I use is whether improving this workflow would still deserve investment if AI did not exist. If the answer is weak, technical feasibility should not rescue the case.
How do you know an AI use case has a business case?
You check whether there is a believable route from technical performance to captured value. The ten-minute test asks four things: which unit of work changes, who can capture the change, what the counterfactual is, and what the full operating cost includes.
The capture step is where most arithmetic breaks. In my worked example, 12,000 requests a year with three minutes saved each is 600 gross hours. After eligibility, oversight and corrections, 330 net hours remain. At an illustrative $60 an hour that is $19,800 of capacity equivalent, which is not $19,800 of cash savings. The figures are invented for the example. The point holds in real data too. A randomized field experiment across 66 firms and 7,137 knowledge workers found that generative AI access cut email time among users by about two hours a week, yet the study detected no corresponding change in the quantity or composition of tasks from individual access alone from individual access alone.
Saved time is capacity, not cash
If staffing, throughput and outcomes stay the same, the financial line has not moved. Name the person authorized to change staffing, procurement, backlog priorities or service commitments. If the benefit depends on a decision nobody owns, the case has a missing link.
The UK's May 2026 guidance on evaluating AI interventions makes the same distinction between technical performance and impact. It calls for a theory of change, a baseline and a comparison with what would have happened otherwise. It is public-sector guidance and no corporate ROI formula, yet the discipline travels.
Who should own the value of an AI initiative?
A business leader whose own numbers move should own the value claim. The AI team should not own AI value. BCG reports that only 14 percent of CEOs clearly define the P&L impact of all their AI initiatives, and a later BCG release says more than 80 percent of AI spending now sits outside enterprise IT. BCG does not publish when the first survey was fielded, so I read both as directional. The money is moving into the business while the definition of what it should return is not.
I split ownership in two. Capability and risk accountability covers the platform, evaluation, security and standards, and it is best done once, centrally. Value accountability is a promise that a specific workflow will produce a better number, and it is local. The test I propose is to write two names per initiative: the person whose budget or target changes if it works, and the person who absorbs the cost if it fails. If both names belong to the AI team, the initiative is a technology project reporting to itself.
The same logic applies to buying. Most AI strategies are software procurement plans with better typography, because a vendor can deliver a platform but cannot decide which activity loses resources. I would ask the accountable owner to finish four sentences with the product put away: which unit of work will change, which decision right or rule will change and under whose name, what existing activity will be stopped or reallocated if it works, and what harm or cost would reverse the change. A vendor can sign the supplier ledger. Only the business can sign the operating-change ledger.
What should an AI pilot be for, and when should it stop?
A pilot should be a decision instrument, and the decision to stop has to be written before the results arrive. Your AI roadmap is probably missing the only column that matters: what evidence would change our decision. Without it, the same request to continue can be justified after almost any result.
A pilot with no kill criteria is not an experiment. A usable criterion is a small contract with six fields: the business claim, the comparison, the evidence window, the threshold, the decision owner and the next action. I distinguish three reasons to stop. A safety stop can suspend a trial at once, an economic stop means the value case no longer clears its hurdle, and a learning stop means the pilot can no longer answer its own question. Before approving a pilot, I ask the sponsor to complete one sentence: if we observe a result by a date, then a named person has the authority to stop it, and the next funding tranche will not be released unless a stated condition is met.
Gartner's 2027 CIO and Technology Executive Agenda survey of 2,002 technology executives reports that 13 percent say their enterprise has realized significant or transformational AI value. Organizations with guidelines for when pilots should not advance were 2.3 times more likely to be in Gartner's high-performing group, but only 31 percent had such guidelines. That is a vendor-reported association, not proof that a checklist causes results. It does suggest that stopping deserves explicit criteria.
When a pilot ends and the result looks good, a pilot has to deserve to survive. I ask for five exhibits at a review chaired by the future operating owner: the outcome against a real alternative, the result when a second team runs it, the complete operating bill, a named owner who can take it over, and a way out. The decision is not pass or fail. It is fund a bounded production transition, redesign one unproven assumption, or end the pilot and keep the learning. The sponsor should be able to advocate. The operating owner should be able to refuse.
Can your company absorb the change?
An AI strategy needs a plan for absorption, because the model finishes before the organization does. Stop asking what AI can do, ask what your company can absorb. I treat absorption as four boundaries: the system produces an output, someone accepts it against a quality and risk standard, a team integrates it into a changed decision or workflow, and the organization operates the new way of working, including exceptions. Speeding up the first boundary does not open the last two.
The practical test is the second team. A pilot team learns the system's weak points and silently compensates for them. A different competent team, working from the documented process and ordinary permissions, shows what the organization actually owns. I would classify every rescue by the original team as context, technical, judgment, relationship or authority rescue, and watch whether the need for exceptional help falls over time. The measure I would watch is support decay, since documentation completeness says little about who can run the system. This is my proposed diagnostic, not an established benchmark.
Which metrics belong in front of executives?
Only the metrics whose change could rationally change a business decision. Five AI metrics I would remove from an executive dashboard: licenses activated, prompts or weekly active users, hours saved, model benchmark scores and pilots launched. They remain useful for operations, engineering and evaluation. They are weak as the headline because each can rise while the business stands still.
I would replace them with a short decision contract per material investment: the outcome and the unit being changed, the comparison, the net effect after integration, review and vendor costs, and the person authorized to scale, redesign or stop it. The test I apply to any executive metric is simple. If the number improves next month, what would we do differently, and if it deteriorates, what would we stop? If the answer to both is nothing, it may be a good diagnostic, but it is not yet an executive metric. The number I would most like to see is the count of investment decisions resolved with adequate evidence, including the ones that ended.
Does every AI decision need a strategy this heavy?
No, and I would resist applying all of this to everything. For a bounded, low-risk task with stable inputs, clear ownership and measurable value, buying a standard product may be the most rational strategic decision. Requiring a transformation program for a small productivity tool can destroy the economics of the purchase. The amount of strategy should scale with the consequences of being wrong. A small deployment needs a short decision record: the problem, the owner, the costs and risks that matter, and when to stop.
Early exploration is a second limit. Demanding a precise return before anyone understands the opportunity can suppress valuable learning. The answer is to fund exploration as learning and deployment as performance, with different evidence for each, and to require that even an exploratory initiative ends in a decision.
A third objection is that business units lack the technical judgment to own AI value, and that decentralizing everything produces duplicated tools and uneven controls. I agree. Risk, standards and the platform stay central, and only the value claim moves.
Finally, be careful with the evidence under all of this. The McKinsey and BCG figures are self-reported surveys, the Gartner finding is an association, and the frameworks in this essay are my own proposals. None has been tested in a controlled comparison against other ways of running an AI program.
What is the step-by-step sequence for building an AI strategy?
This is the order I would follow for each material AI investment, and then again for the portfolio.
- Name the business outcome and the constraint that blocks it, and write the five-line brief before any model is chosen.
- Compare the AI option with the strongest non-AI option and with deferral, using the same value, risk and cost criteria.
- Run the ten-minute test: the unit that changes, who can capture it, the counterfactual and the full operating cost.
- Write two names, the business leader who owns the value claim and the owner of the capability, and fix the baseline before launch.
- Write the kill criteria and complete the stop sentence before the pilot budget is released.
- Hold a survival review chaired by the future operating owner, and run the second-team test before scaling.
- Report one page of decisions and outcomes, and keep activity metrics on the operational dashboard.
- Review the whole portfolio and ask what you would stop funding to scale the next initiative.
Sources
url: https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/blogs/pulse-check-series-latest-ai-developments/ai-transformation-predictions-2026.html title: Enterprise AI trends 2026: AI transformation strategy publisher: Deloitte date: 2026