Stop Treating AI as the Strategy

A chief executive asks every division to produce an artificial-intelligence plan. Within weeks, the company has a catalogue of promising ideas: automated customer service, synthetic marketing content, faster financial reporting, intelligent search and an internal assistant capable of answering employees’ questions.

The list creates momentum, but not necessarily strategy.

Each project may demonstrate that the technology works, yet none answers the harder commercial questions. Which customer problem has become more urgent? Where is the company losing margin, speed or market share? Which workflow should be redesigned rather than merely accelerated? What advantage will remain once competitors can buy access to comparable models?

This is where many corporate AI programmes begin to lose direction. The organisation treats adoption as evidence of progress, counts pilots and user licences, and describes AI as a strategic priority. What it has not established is the business strategy that AI is expected to serve.

That distinction is becoming more important as experimentation spreads. McKinsey’s 2025 global survey found that almost nine in ten respondents reported regular AI use somewhere in their organisations, yet most companies were still struggling to translate experimentation into scaled enterprise impact. Only around 39 percent reported any measurable effect on earnings before interest and tax, and most of those attributed less than five percent of EBIT to AI.

The gap does not suggest that AI lacks value. It shows that access to the technology is becoming easier than the organisational work required to capture it.

AI can support strategy, alter the economics of a process and occasionally enable a new business model. It is not a substitute for deciding where the company will compete, which customers it will serve or why it should win.

Begin with the business constraint

A useful AI programme starts with a problem whose commercial consequences are already understood.

A manufacturer may be losing production time because maintenance teams cannot identify equipment deterioration early enough. An insurer may employ skilled underwriters who spend too much of the day extracting information from documents. A professional-services company may know that its experts have produced valuable knowledge but cannot retrieve it quickly enough to reuse it. A retailer may have excessive inventory because forecasts respond too slowly to local demand.

These are strategic or operational constraints. AI becomes relevant only after the organisation has identified which decision, task or information gap contributes to the problem.

The difference can be seen in customer service. “Introduce an AI chatbot” is a technology initiative. “Reduce the time required to resolve routine account queries while preserving access to a qualified employee for complex or vulnerable customers” is a business objective.

The second formulation establishes the outcome, the customer boundary and one of the principal risks. It allows the company to compare AI with other interventions, including clearer invoices, better staff training, improved website navigation or the removal of a process that generates unnecessary calls.

Without this step, companies frequently automate symptoms. The chatbot answers questions created by confusing policies; the summarisation tool processes reports that nobody needs; and the internal assistant searches a fragmented knowledge base whose underlying information is obsolete.

AI can make a poor process move faster. That does not make the process more valuable.

Distinguish strategy from use case and platform

Leadership teams often collapse three separate decisions into one.

The strategy defines the commercial objective: improve retention among profitable customers, reduce unplanned production outages or shorten the time required to bring a product to market.

The use case describes the work AI may perform: predict equipment failure, summarise technical records or generate an initial product design.

The platform supplies the models, data infrastructure and tools through which that use case is delivered.

The order matters. When companies begin by purchasing a broad AI platform, they can feel pressure to find enough applications to justify the commitment. The technology portfolio then drives business priorities instead of supporting them.

The opposite mistake is to permit hundreds of uncoordinated tools. Employees solve local problems quickly, but the company accumulates duplicated contracts, inconsistent security controls and data moving through systems that central management cannot see.

A disciplined organisation maintains a relatively small technology foundation while allowing business units to propose use cases against common criteria. The foundation may include approved models, identity controls, secure access to enterprise data, monitoring and procurement standards. It should enable experimentation without predetermining where the value lies.

Technology architecture is important, but it becomes strategic only when connected to the organisation’s choices about customers, economics and capability.

Concentrate investment where the economics are visible

The enthusiasm surrounding generative AI has encouraged companies to spread modest investments across many functions. This produces impressive pilot numbers and weak concentration.

BCG’s 2024 research classified 49 percent of surveyed companies as remaining largely in proof-of-concept mode, while only four percent had developed what it described as mature AI value engines. The firms reporting stronger results tended to concentrate their efforts on a limited number of core business processes rather than distributing resources evenly across numerous small initiatives.

A useful portfolio should contain several types of application, but it needs a clear hierarchy.

Some uses improve personal productivity: drafting routine text, summarising meetings or searching documents. These can save time across a large workforce, although the value is often dispersed and difficult to realise unless the organisation changes how that time is used.

Other applications improve a defined workflow. They may accelerate claims processing, reduce false fraud alerts or help engineers diagnose faults. Their value can usually be measured through time, cost, accuracy, throughput or customer outcomes.

The most strategically significant applications change the economics of a product or business model. AI might allow a company to serve a previously uneconomic customer segment, personalise a service at scale or create a data advantage that improves with use.

The company does not need every initiative to be transformational. It does need to know which category each initiative belongs to and fund it accordingly.

An employee writing assistant should not receive the same executive attention as a system that could change underwriting decisions across an insurance portfolio.

Redesign the workflow rather than adding another screen

Many organisations insert AI into an existing process without altering the process itself.

An employee receives an AI-generated summary but must still copy it manually into another system. A model recommends a decision, yet the same number of approvals remain in place. A customer-service assistant proposes an answer, but the agent must search several databases to verify it.

The technology creates an additional step rather than removing one.

McKinsey’s research on organisations capturing value from generative AI identifies workflow redesign as one of the practices most strongly associated with impact. Other important practices include embedding the technology into daily systems, establishing clear adoption road maps, training people by role and tracking well-defined performance indicators.

Workflow redesign begins by mapping how the work is currently completed. Which activities require judgement? Which exist because information is difficult to retrieve? Where does the process wait for approval, and which errors create rework later?

The organisation can then decide which steps AI should automate, which it should support and which should remain under human control.

A mortgage lender, for example, may use AI to extract information from documents and flag inconsistencies, while preserving human responsibility for credit decisions and unusual cases. The benefit does not come solely from faster document reading. It comes from restructuring the workflow so underwriters spend more time evaluating risk and less time assembling files.

The operating model must also specify what happens when the model is uncertain. An AI system that handles routine cases efficiently can still fail commercially if its exceptions are sent into an understaffed queue with no clear owner.

Give every initiative a business owner

AI programmes are often led by technology teams because those teams understand the infrastructure and suppliers. Technology ownership is necessary, but it is not enough.

The executive responsible for the commercial process should own the result.

If AI is being introduced into procurement, the procurement leader should be accountable for whether supplier decisions improve. If it supports marketing, the marketing leader should answer for conversion, brand quality and customer acquisition costs. The technology function remains responsible for architecture, security and reliability, but it should not be asked to manufacture a business case on behalf of another department.

Each initiative needs one named owner with authority over the workflow, budget and adoption decisions. That owner should establish the baseline before implementation and agree how benefits will be measured.

Without a baseline, almost any pilot can be described as successful. Employees may say it feels faster, the demonstration may work well and senior leaders may enjoy using it. None of those observations establishes financial value.

The relevant metric depends on the application. It might be the cost per resolved query, production downtime, claims leakage, average sales-cycle length or the proportion of engineering time devoted to repeat work.

The organisation must also decide how value will be realised. Saving an employee ten minutes does not automatically reduce cost or increase revenue. The time must be aggregated, redirected towards higher-value work or reflected in greater throughput.

A credible business case states not only how much time AI might save but what the company will do with it.

Data readiness is a strategic choice

Companies frequently describe data quality as a technical obstacle discovered during implementation. In reality, it reflects years of business decisions.

Customer records may be duplicated because regional units were allowed to use different definitions. Product data may be incomplete because no executive was accountable for maintaining it. Valuable knowledge may remain in personal files because incentives favoured individual ownership rather than organisational reuse.

AI exposes these weaknesses because models require accessible, relevant and sufficiently reliable information. A sophisticated system connected to poor data may produce answers quickly and confidently, but not correctly.

The solution is not to clean every dataset before beginning. That can become an expensive programme with no clear end. The company should improve the information required for its priority use cases and assign permanent ownership to it.

For a sales application, that may mean defining which customer interactions must be captured, who can amend a record and how quickly outdated information should be corrected. For an industrial model, it may involve consistent sensor calibration, maintenance labels and equipment identifiers across sites.

Data governance should follow business value. The information most important to pricing, risk, customers and operations deserves the strongest ownership and quality controls.

The same principle applies to proprietary data as a competitive advantage. Many companies assume that possessing large volumes of information will differentiate them. It will do so only when the data are legally usable, relevant to the decision and organised well enough to improve the system.

Do not confuse model access with competitive advantage

Most companies can purchase access to similar foundation models. As those models improve and prices decline, the model itself becomes a weaker source of lasting differentiation.

Advantage is more likely to come from how the company combines technology with proprietary information, domain knowledge, distribution and redesigned processes.

A bank and a retailer may use the same language model, yet the bank’s value depends on integrating it with risk policies, customer records and regulated approval processes. The retailer’s advantage may depend on demand data, supplier relationships and the ability to alter inventory decisions quickly.

The difficult work lies around the model: data preparation, system integration, evaluation, workflow design and employee adoption.

This also means that switching models may become necessary. Performance, price, regulation and data-location requirements will change. An architecture designed around one supplier can become expensive to alter later, particularly when prompts, evaluation systems and applications are closely tied to proprietary features.

Companies do not need to avoid strategic technology partners, but they should understand where dependence is being created. Contracts should address data use, model training, audit rights, service continuity and the practical ability to migrate.

The objective is not theoretical vendor neutrality. It is sufficient leverage to prevent the AI platform from becoming the company’s unexamined strategy.

Governance should follow the decision at risk

Some companies respond to AI risk by creating a central committee that reviews every experiment. The result is often delay, informal evasion and employees using unapproved tools outside the official process.

Others issue broad principles and leave individual teams to interpret them, producing inconsistent controls.

A proportionate model begins with the consequence of error.

An internal tool that helps an employee rewrite a presentation carries different risks from a system that recommends medical treatment, determines access to credit or communicates directly with vulnerable customers. Governance should become more demanding as the system gains autonomy, handles more sensitive information or affects more consequential decisions.

The NIST AI Risk Management Framework organises this work around four continuing functions: govern, map, measure and manage. Its generative-AI profile also addresses risks such as confabulation, privacy, information security, harmful bias, intellectual property and human overreliance.

In practice, the business should document the intended use, data involved, decision owner, performance tests and escalation procedure. It should know whether output is advisory or binding, which people can override the system and how incidents will be recorded.

Models must be monitored after deployment because user behaviour, data and operating conditions change. A system that performed well during testing can deteriorate once customers learn to interact with it or employees begin using it for tasks outside its original scope.

Governance is therefore not the approval meeting held before launch. It is the operating discipline that continues afterwards.

Adoption is a management responsibility

An AI tool can be technically sound and commercially irrelevant because employees do not use it or use it badly.

Adoption fails when workers receive a generic demonstration without understanding how their role will change. It also fails when management describes AI only as an efficiency programme and employees reasonably interpret efficiency as a threat to their jobs.

Leaders should explain the intended operating model with precision. Which tasks are expected to disappear? Which decisions remain human? How will performance be assessed when output increases? What support will be available when the tool fails?

Role-based training is more effective than general enthusiasm. A lawyer needs to understand confidentiality, source verification and professional responsibility. A sales representative needs to know when a suggested message becomes misleading. A manager needs to recognise automation bias and challenge a recommendation rather than treating the model as an authority.

Senior executives must also use the systems they endorse. Leadership involvement matters not because every chief executive should become an expert prompt writer, but because visible use exposes practical limitations and signals that adoption is part of operating change rather than an isolated IT exercise.

The company should provide a simple route for employees to report inaccurate outputs, security concerns and workflow problems. Feedback has to reach the product or implementation team quickly enough for the system to improve.

AI adoption is not completed when licences are distributed. It is completed when the new behaviour becomes the normal way valuable work is done.

Measure value at three levels

The first level is technical performance. Does the system produce sufficiently accurate, reliable and timely output under realistic conditions?

The second is workflow performance. Does it reduce handling time, increase throughput, improve consistency or allow employees to spend more time on higher-value activity?

The third is business performance. Does the change improve revenue, margin, customer retention, risk outcomes or another strategic measure?

A system can succeed at the first level and fail at the other two. A highly accurate summary tool creates little value when employees do not trust it or when summaries were never the constraint. A faster claims process may damage the business if it increases inappropriate payouts or customer complaints.

Benefits and costs must be measured together. Model fees may be modest while integration, data engineering, human review and change management are substantial. Usage costs can also grow unpredictably as adoption expands.

The financial model should include implementation, operation, monitoring, security and the cost of correcting errors. It should distinguish temporary productivity gains from structural improvements in the economics of the process.

Management should close initiatives that do not produce sufficient value. AI portfolios need the same capital discipline applied to other investments. A failed pilot is inexpensive when it provides an early answer; it becomes costly when the company keeps it alive to preserve the appearance of progress.

A practical 90-day reset

During the first month, leaders should stop asking departments for AI ideas and instead identify the small number of business outcomes that matter most. Each priority should be translated into a measurable constraint: time, cost, error, capacity, conversion or customer experience.

The next month should be used to map the relevant workflows and data. Teams need to understand how work actually moves, where judgement is required and which information is missing or unreliable. Potential AI interventions can then be compared with simpler process, policy or software changes.

During the final month, the company should select a small portfolio of initiatives with named business owners, baselines and explicit stopping rules. At least one should be capable of producing measurable value within a short period, while another may test a more strategic capability whose payoff requires deeper integration.

The organisation should also establish common technology and governance foundations: approved models, data-access controls, evaluation procedures, incident reporting and procurement standards. These foundations should be strong enough to reduce risk without becoming an excuse to centralise every decision.

At the end of 90 days, the board should not ask how many AI projects are running. It should ask which business constraints are being removed, what evidence supports the claimed benefit and what must change in the operating model for the result to scale.

AI deserves serious strategic attention because it can alter cost structures, decision-making and the design of products. That does not make it a strategy in its own right.

A strategy explains where the company will create value and why it is positioned to do so better than its competitors. AI can strengthen that answer, but it cannot supply one.