Is This AI Quality or Just Theme Exposure? | AI Strategy

Inside the AI conversation sits a capital-allocation problem. Executives now approve AI budgets with a kind of defensive confidence: competitors are spending, boards keep asking, and employees are experimenting on their own anyway. The easy response is to spread money across a broad portfolio and hope winners surface later, and that approach no longer fits the size of the commitment.

The buildout has moved past experiment budgets into infrastructure commitments, hiring plans, vendor contracts and a large share of management attention. Once a company has made commitments of that kind, it is making a capital-allocation decision and should judge it as one.

The board question should be: what quality of business are we creating with this AI spend?

What investors are watching in the AI stack

The investment world offers a useful habit: follow the cash, then ask who holds a lasting advantage once the cash has moved through the stack. An Excess Returns discussion with Tom Hancock of GMO (Grantham, Mayo, Van Otterloo) described AI in layers: applications, compute, models and infrastructure. Money paid by one layer becomes revenue for the next. End users pay for applications, applications buy compute, and compute providers fund models, chips, power and data centers. For an operator, the useful lesson is to treat each layer of the stack as a separate opportunity instead of lumping them together.

Goldman Sachs Research estimates that consensus 2026 hyperscaler capital expenditure has risen to $527 billion. It also notes that investors are getting more selective, rewarding capital spending that has a visible link to revenue and marking down spending that squeezes operating earnings or leans heavily on debt, which is a reasonable rule for a management team to apply to its own budget.

The adoption surveys make the same case in operating terms. Boston Consulting Group's (BCG) July 2026 work found that only a small share of companies qualify as AI leaders, and what sets them apart is real business performance, talent development and the habit of reinvesting productivity gains into scale and new opportunities. A separate BCG CEO survey found that many companies get targeted value from AI but struggle to turn it into enterprise impact, because the links to P&L, workflows, accountability and funding are weak. McKinsey's 2025 State of AI survey likewise placed workflow redesign and CEO oversight of governance among the factors most associated with bottom-line impact. Deloitte's 2025 enterprise survey found that most experiments were still scaling slowly even as AI budgets kept rising. Read together, these sources suggest that AI spending pays off when it strengthens a business system that will last; spending that buys optionality without changing the company's economics or operating model mostly produces the appearance of progress.

Five tests before an AI investment gets serious capital

An investment should clear all five of the tests below before it receives serious money.

Test	What It Proves	Failure mode
Advantage	The initiative uses proprietary data, distribution, workflow depth, trust, speed, or domain knowledge that competitors cannot quickly copy.	The company buys a tool that every rival can buy, then calls access a moat.
Value Path	The team can name the P&L line, baseline, adoption assumption, cost curve, and time horizon before scale funding begins.	Leaders count users, prompts, demos, or model upgrades instead of financial impact.
Workflow	Work, roles, incentives, controls, and decision rights change because the AI system is now part of the operating model.	AI is bolted onto a broken process and then blamed for not transforming it.
Resilience	The plan can survive model churn, vendor repricing, capacity limits, governance failures, cyber incidents, and uneven adoption.	The business becomes dependent on a fragile AI layer before it understands the failure modes.
Balance Sheet	Commitments are staged so learning precedes lock-in and fixed costs do not outrun proven demand.	Management turns forecasts into contracts before the use case has earned them.

The first test gets the least attention. AI advantage rarely comes from the model alone. In most businesses it comes from what the model is combined with: trusted customer relationships, proprietary data, regulated processes, switching costs and managerial judgment. Many software businesses turn out to be harder to disrupt than their code base suggests, because a competitor can recreate the product's surface while the business stays protected by years of data history, compliance work, implementation depth and procurement trust. An AI application can look spectacular in a demo and still fail this test if it owns none of the complements that would let it keep the value it creates.

Why a venture-style portfolio of pilots fails

Companies should not run AI as a venture portfolio of internal pilots. The comparison sounds rigorous because it accepts failure, but in practice it often excuses under-managed activity. A venture investor can hold a basket of uncertain bets because one winner can pay for many failures. An operating company usually cannot, since every pilot consumes scarce attention, leaves integration residue behind, teaches employees to expect constant tooling churn, and may bring data, legal or customer-experience risk with it. A pilot that never scales can also teach the organization that AI is a side project.

Capital budgeting under uncertainty is the better model. Fund learning and preserve options, but put more capital in only when the evidence improves, and kill or merge initiatives that do not build a real complement. The roadmap should be a set of business changes, with experiments feeding into them rather than piling up alongside.

Where to spend

None of this means spending less by default. It means spending where AI changes the structure of work or the economics of growth enough to make the company more valuable. Three areas stand out.

Fund bottleneck removal. Where a workflow is held up by judgment latency, compliance queue time or case triage, AI may turn scarce expert attention into higher throughput. The aim is more useful work from the same organization, and headcount cuts should not be the default story.

Fund proprietary learning loops. The strongest AI systems improve as the company uses them, through better-labeled exceptions, richer customer context, cleaner feedback and stronger control logs. A system that gets no more useful with use is probably just a feature.

Fund trust-critical augmentation. In regulated or high-stakes workflows, the near-term value may lie in better preparation, review and monitoring rather than full autonomy. That can still be strategic: a company that makes experts faster and more consistent in a trust-sensitive process may build a stronger moat than one chasing autonomy where failure is unacceptable.

What boards should ask management for

Boards do not need to approve every model choice, but they should hold large AI spending to a higher standard. For each major initiative, ask management for a one-page quality memo that answers five questions:

Which durable complement does this initiative strengthen?
Which P&L line will move, and what baseline will finance use?
What change in workflow, role, incentive or decision rights does it require?
Which vendor, model, data, security, regulatory or capacity risk could break the value case?
What evidence would trigger a decision to stop, shrink, merge or double down?

The fifth question does the most to protect capital. AI programs fail when leaders count persistence as success, and naming the stop evidence in advance is what prevents that.

Sorting the most visible initiatives this week

This week, take the five AI initiatives with the most executive visibility and put each into one of three buckets.

Quality bets have a clear complement, a named P&L path, accountable workflow owners and staged funding. These deserve stronger management support.

Learning options are uncertain but useful because they answer a specific question. Give them short cycles, narrow scope and explicit kill criteria.

Theme exposure exists mainly because the company wants to be seen doing AI. Stop these, merge them, or recast them as a narrower learning option.

The exercise will be uncomfortable because it separates ambition from evidence, and the number of pilots a company runs says little about whether its spending is making the underlying business more productive or harder to copy. Initiatives that pass belong in the strategy with staged funding; those in the third bucket should be cut back even when someone defends them as innovation.