What it is not

Three documents get called an AI strategy and are not one.

A capability survey. Twenty pages on what large language models can do, what competitors are reportedly doing, and which categories of tool exist. Genuinely informative, and it commits to nothing. You can read it twice and still not know what happens on Monday.

A tool evaluation. A comparison matrix of vendors with pricing and feature checkmarks. This is procurement, and it is premature: you cannot sensibly evaluate a tool before you have decided which workflow it has to fit into.

A roadmap slide. Four quarters, each with a theme like “expand automation” or “scale AI adoption”. Themes are not decisions. A roadmap becomes a strategy at the point where a quarter contains a named workflow with an owner.

The common thread is that all three avoid the part that carries risk: saying which specific thing you will change first, and therefore which things you will not.

Definition

An AI implementation strategy is a written decision record covering which workflows you will change, in what order, with what oversight, and how you will know whether it worked.

What the document contains

Eight sections. The order matters, and tool selection deliberately sits seventh; a strategy that opens with vendors has skipped the reasoning that should constrain the choice.

#SectionWhat it must answerCommon failure
1Scope and exclusionsWhich functions are in play this year, and which are explicitly outEverything is in scope, so nothing is prioritized
2Workflow inventoryThe candidate workflows, each with its frequency and the people who perform itA list of ideas and aspirations rather than observed processes
3Prioritization and sequenceWhich one is first, second, third, and the scoring that produced that orderA ranked list with no stated criteria, which cannot be argued with or revised
4BaselinesThe current measurement for each prioritized workflow, captured before any changeAbsent, which makes every later claim unfalsifiable
5Governance and risk tiersWhich use cases require what oversight, and who reviews before output takes effectA general commitment to “responsible AI” with no operative rule
6Data and access decisionsWhat data each workflow touches, its classification, and whose identity the system acts underDeferred to implementation, where it becomes the thing that stalls the project
7Tooling and build-versus-buyWhat you will use, and the condition under which you would build insteadPlaced first, and treated as the strategy itself
8Measurement and reviewWhat will be measured, by whom, at what cadence, and what would cause you to stopNo stopping condition, so no project is ever wrong

Eight sections is a working document, not a deliverable to admire, commonly ten to twenty pages. Length is not the signal. Specificity is.

First-hand experience

Where these documents actually break down

Sitting with teams through this, the failure is rarely analytical. It is that section four (baselines) requires someone to go and measure something, and that is unglamorous work with no immediate payoff. It gets deferred to “during implementation”, where it quietly never happens.

The consequence surfaces months later, when leadership asks what the program achieved. Without a baseline the honest answer is a collection of anecdotes, and the project loses funding to something with better storytelling. Teams whose work genuinely helped have lost budget this way.

The other recurring break is section one. Asking a leadership group what is out of scope produces visible discomfort, because exclusion is where a strategy stops being universally agreeable. A document everyone in the room is comfortable with has usually not decided anything.

Reviewing a strategy someone hands you

Whether it came from an internal team or a consultancy, these questions separate a decision record from an expensive survey. Any “no” is worth pursuing before signing anything.

  • Can you name the first workflow?

    Not the first theme or the first phase. The specific repeating process that changes first.

  • Does that workflow have a named owner?

    A person accountable for the process outcome, not for technology generally, and not the person most enthusiastic about AI.

  • Is there a baseline, with a number in it?

    Cycle time, volume, error rate, or cost per unit, measured, not estimated.

  • Does it say what is out of scope?

    If everything is included, nothing has been prioritized.

  • Is there a review point per use case?

    Who inspects output before it takes effect, sized to what happens if it is wrong.

  • Are data classification and access resolved, or deferred?

    Deferred access decisions are the most common cause of a stalled implementation.

  • Is there a stopping condition?

    What result would cause you to retire this rather than extend it. A strategy with no failure condition cannot be evaluated.

  • Does tool selection come after workflow selection?

    If vendors appear before workflows, the reasoning is running backwards.

  • Could a new hire execute from this document?

    The practical test of specificity. If it needs its author present to be actionable, it is a presentation.

Strategy, plan, and roadmap

These get used interchangeably and are different artifacts with different jobs. You need all three eventually; conflating them is how organizations produce documents that read well and change nothing.

The strategy decides

  • Which workflows change, and in what order
  • What is explicitly out of scope
  • What oversight each risk tier requires
  • How success will be measured, and what would stop the work
  • Build versus buy, and under what condition

The plan and roadmap schedule

  • Who does what, in which week
  • Which systems get integrated in which order
  • When training happens, and for whom
  • Dependencies, and what blocks what
  • Budget phasing and resourcing

How long it should take to write

Less time than most organizations spend, and more observation than most expect. The analysis is not the expensive part; watching how work actually happens is.

For a single function, two to four weeks is a reasonable shape: one to two weeks observing and documenting real processes, a few days scoring candidates, and a few days writing. Most of the elapsed time is scheduling access to the people doing the work.

A strategy that took a quarter to produce is usually a sign that the scope was never constrained: section one was skipped, so the document had to survey everything. A strategy produced in a two-day workshop without observation is usually a capability survey with a decision-shaped cover.

On strategies that arrive with a tool already chosen

If a document recommends a specific platform before naming the workflows it will serve, the reasoning ran backwards. If the author resells or is partnered with that platform, that is worth knowing before you read further.

This is not an accusation of bad faith. It is that a tool-first strategy optimizes for what the tool does well, which may or may not be where your time is actually going.

Limitations of this article

This describes strategy documents for organizations of roughly ten to a few thousand people, with existing business systems and no dedicated machine-learning function. Organizations training their own models, or deploying under a named regulatory regime, need sections this outline does not include.

The eight-section structure is a working template rather than a validated instrument. No study supports it, no effect size is claimed, and a competent strategy could be organized differently. What is defensible is the underlying test: a strategy that names no workflow, owner, baseline, or stopping condition has not made a decision.

This article also assumes you want a written strategy at all. For an organization with one obvious workflow and a capable owner, writing a strategy first may be procedural overhead. The better move is to run that one workflow properly and write the strategy from what you learn.

External references

Relevant to the governance and risk sections. Listed as useful reference, not as endorsement of any single framework.

Frequently asked questions

Do we need a written strategy before doing anything with AI?

Not always. If you have one obvious workflow, an owner, and the access to change it, run that one properly and write the strategy from what you learn. Writing first would be overhead.

A written strategy earns its cost when there are competing candidates, more than one function involved, real data-sensitivity questions, or a need to justify sequencing to a board.

Who should write it?

Someone who has watched the work happen. That is the binding constraint, not seniority and not technical depth.

In practice it is usually co-written: an operations or process owner who knows how the work really runs, with technical input on feasibility and access. A strategy written entirely from interviews with leadership tends to describe the documented process rather than the real one.

How is this different from a digital transformation strategy?

Structurally, very little. The sections are close to what any competent systems-change document would contain.

Two things differ in emphasis. Section five matters more, because AI failure modes are fluent and plausible rather than obviously broken, so oversight has to be designed rather than assumed. And section six matters more, because sending data to a third-party model provider raises retention and training-use questions that an internal system does not.

How often should it be revised?

Review at least every six months, and immediately when a prioritized workflow completes or is retired; the sequence should change based on what you learned.

Rewriting it in response to every new model release is a sign the strategy was tool-led. If a vendor announcement invalidates your strategy, the strategy was a procurement document.

Part of the pillar: AI implementation strategy

Author:
Martin Zialcita
Published:
Last reviewed:
Corrections:
Editorial policy