AI strategy development is the process of deciding where AI belongs in your business, what it may do, and who will remain accountable for its operation. Start with a measurable business result, define the process and permissions, then specify the evidence required before release. The strategy should give leadership a clear basis for funding, narrowing, expanding, or stopping each proposed system.
Key takeaways
- Organize the strategy around business results, operating ownership, system boundaries, and release evidence.
- Assign an owner for the process result separately from the owner responsible for technical operation.
- Require separate approval for accessing real data and for taking recurring production actions.
- Evaluate the complete process, including review and exception handling, before expanding the system.
What should the four pillars of your AI strategy cover?
Use four pillars as a planning structure: business value, operating ownership, system design, and evidence. Treat these as the framework proposed here, rather than a universal taxonomy. For each pillar, require a written artifact that a funding or release decision can reference.
Business value: define the result and the alternative
Write a result statement that names the process, the intended improvement, and the measure you will use. For example, propose reducing unresolved purchase-request exceptions while preserving required approvals. Compare the proposed system with a native software feature, a deterministic automation, and a process change before authorizing custom development.
Keep individual assistance separate from enterprise performance in the business case. Ask what must change in the process before an individual improvement becomes an operating result.
Operating ownership: assign decisions, not just sponsors
Name the executive sponsor who can fund or stop the work, the process owner who accepts its business behavior, and the technical owner responsible for running it. Give risk, security, and privacy reviewers explicit approval responsibilities within their remit. Also name the person who handles exceptions during ordinary operation.
Avoid assigning shared accountability without specifying decision rights. Require one named decision-maker for release, one for suspension, and one for approving a material change. The same person may hold several responsibilities, but record each responsibility separately.
System design: define the operational perimeter
Define the operational perimeter as the exact process, users, data, tools, and actions included in the system. State exclusions with equal precision, including records the system must not access and decisions it must not make. Use a system map to connect each permitted action to its source, receiving system, and responsible owner.
Compsia designs, deploys, and operates custom AI production systems, defining system-specific data, sources, actions, owners, controls, acceptance tests, and operating limits. Our explanation of what an AI business layer is provides context for connecting these decisions.
Evidence: specify what would justify release
Write acceptance scenarios before selecting the final technical approach. Include ordinary work, ambiguous inputs, missing authorization, unavailable tools, and failed actions. Require the process owner to accept the business behavior and the technical owner to accept the operating procedure.
For a generative AI strategy, specify which generated outputs require source checking, which require approval, and which must never become executable instructions. Keep the test set aligned with the permitted use. Do not use a good drafting result as evidence that the system should receive write access.
Build the strategy through five approval steps
As you develop the strategy, record the decisions that need approval and the documents supporting each one. Keep each document short enough for its owner to maintain. Increase detail where an action has material consequences or where reviewers need more evidence.
Step 1: document the current process
Map the trigger, inputs, decisions, handoffs, receiving records, and completion condition. Ask operators to identify exceptions and show how they resolve them today. Record the baseline using the same process boundary you intend to evaluate after launch.
Choose measures that reflect completed work, such as unresolved exceptions, corrected records, or accepted outputs. Record review effort and rework alongside the primary measure. Do not approve a business case that counts generated output while excluding the work required to make it usable.
Step 2: compare approaches before ranking projects
For each candidate, compare keeping the process unchanged, improving the process, configuring existing software, adding deterministic automation, and building a custom AI system. Explain why the selected approach is proportionate to the result. Put the rejected alternatives in the decision record so the team can revisit them if assumptions change.
Apply a feasibility rule before scoring potential value: defer a candidate without an accountable owner, authorized data access, or a workable acceptance test. Then rank the remaining candidates by business importance, implementation burden, operating burden, and reversibility. Do not let a high projected benefit cancel a missing release requirement.
Step 3: separate data access from action authority
Create one approval record for allowing real data into the proposed processing path. Include data categories, authorized users, providers, receiving systems, retention, deletion, and incident ownership. Require reviewers to approve the actual path rather than a general product description.
Create a second approval record before allowing recurring actions that change records or trigger downstream work. Specify allowed actions, authenticated approvals, duplicate handling, partial-failure behavior, and rollback. Compsia's internal release standard separates these data-access and action-capable production decisions, and requires system-specific evidence for each.
Step 4: test the complete operating process
Use the strategy to define how development will proceed, including explicit acceptance, failure, and change tests. Test what happens after an output reaches the reviewer, including rejection, correction, and escalation. Require a usable manual procedure before authorizing recurring operation.
Record the system version, test evidence, unresolved limitations, and approving owners in a release record. Keep this record distinct from a demonstration or presentation. For the handoff into delivery, use the enterprise AI implementation page to frame the implementation questions that still need answers.
Step 5: approve adoption and operation together
Build the adoption strategy around the work people will perform after release. Specify who starts a task, checks an output, handles an exception, and reports a problem. Train those people on the accepted process and its limits before expanding access.
Approve the operating budget alongside the build decision. Include monitoring, support, evaluation, exception handling, provider usage, and controlled changes. Require a per-completed-case view where feasible, rather than reviewing only the total technology bill.
The enterprise AI research distinguishes between tracking total spending on AI and tracking the per-unit cost of inference (enterprise AI findings, S4). Extend your own evaluation to the full operating process, with review and rework recorded separately from inference spending.
Apply the framework to purchase-request exceptions
Use purchase-request exception handling as a generic planning pattern. Define the proposed task as checking a request against approved purchasing rules and preparing a recommendation for a purchasing reviewer. For the first release, exclude payment approval, supplier-master changes, and changes to purchasing policy.
Write the result statement as a proposal to improve exception resolution without bypassing purchasing authority. Assign the purchasing process owner responsibility for the business result and acceptance scenarios. Assign the technical owner responsibility for access, operating alerts, and recovery, with a named purchasing reviewer responsible for each recommendation.
Design the initial system to read only approved request records and the relevant policy versions. Require its recommendation to identify the missing information, the applicable rule, and any ambiguity. When the evidence is incomplete, direct the case to a reviewer rather than authorizing the system to infer permission.
Test conflicting policy versions, incomplete requests, duplicate submissions, restricted records, and unavailable receiving systems. Require a failed update to remain visible in the exception queue. Check that a retry cannot create a second approval request for the same action.
Evaluate the proposed release using completed exception cases, reviewer corrections, unresolved cases, and operating cost. Approve a later write capability only through a separate change decision with its own tests. Keep the original read-and-recommend release available as a fallback if the broader release fails acceptance.
Decide how much review the process can sustain
Set the review requirement before committing to the business case. For every approval step, specify what evidence the reviewer receives, what they must check, and how they can reject or stop the action. Test whether the assigned reviewers can perform that work within the proposed operating model.
Do not treat human review as a substitute for access control or data authorization. Compsia's internal standard requires privacy, access, retention, incident, and confidentiality requirements to remain in force even when a person reviews the output. Use approval to control the specified decision, with separate evidence for the surrounding data path.
Make review capacity a release condition. If reviewers cannot check the required evidence reliably, narrow the task, improve the evidence presentation, or reduce the permitted action level. Do not remove review solely to preserve a projected benefit.
Also define the stopping condition for the whole system. Suspend the affected path if required evidence disappears, the operating burden exceeds the accepted limit, or an unresolved failure compromises an acceptance condition. Assign the suspension decision in advance rather than leaving it to an informal discussion during an incident.
Choose whether to buy, build, narrow, or stop
Use a decision memo that compares the accepted operating result with all the work and costs required to deliver and maintain it. Separate confirmed facts from assumptions that still need testing. Ask the sponsor to approve the assumptions explicitly rather than presenting them as established benefits.
If you use AI to draft the memo, require the accountable owners to verify the evidence and approve the recommendation. Keep the supplied Deloitte enterprise AI source, S4 available alongside the spending claim when reviewing that evidence.
Use the following decision rules:
- Buy or configure when an existing product meets the accepted process and control requirements.
- Build when a defined result requires a system design that the evaluated alternatives do not adequately support.
- Narrow when a smaller data scope or action boundary makes the proposal acceptable.
- Defer when ownership, authorization, adoption capacity, or operating support remains unresolved.
- Stop when the accepted result does not justify all the work and costs required for delivery and operation.
Set expansion criteria before the initial release. Require evidence from the accepted process, a named owner for the additional scope, and approval for any new data category or action. Use managed AI operations as a prompt for the monitoring, support, and change responsibilities to settle before expansion.
Frequently asked questions
What are the four pillars of an AI strategy?
For this framework, use business value, operating ownership, system design, and evidence. Produce a result statement, an owner map, a system boundary, and acceptance criteria for those pillars. Treat the categories as a practical planning choice rather than a mandatory industry standard.
What is the 30% rule in AI?
Do not use a 30% rule as a release threshold without defining what it measures and who authorized it. Set task-specific acceptance and review requirements instead of adopting an unexplained percentage.
How to use AI to make $10,000 a month?
Treat that amount as a target requiring a business model, not an expected consequence of using AI. For an enterprise proposal, identify the income mechanism or operating result, then subtract implementation and recurring operating costs. Approve a bounded test only when its assumptions can be measured, without promising that income level.
Which 5 jobs will survive AI?
Do not base workforce planning on a list of five protected job titles. Assess the tasks in each role and specify which responsibilities people must retain under your proposed operating model. Include exception handling, consequential approvals, and system supervision in that assessment, then plan training against the redesigned responsibilities.
Source availability
Only S4 has a supplied external HTTPS URL in the material available for this revision. Both external links above point to that same source; they do not represent different sources. The URLs and excerpts for S1 and S3 were not included. Meeting the requirement for three different provided S sources linked inline remains blocked until two additional S sources and their URLs are supplied; no replacement sources or support have been invented.
Bring one process to an operations review
Choose one process with an accountable owner and a result worth evaluating. Bring its current steps, decision points, source records, exceptions, and approval requirements to a review of how Compsia designs, launches and operates a production system. Ask for a recommendation to buy, build, narrow, or stop before committing to broader scope.
We look at how your team runs its core processes and show you what we would make more efficient.
Primary references
Continue reading: AI and GDPR Compliance: A System-by-System Framework.