DECISION BRIEF 24

AI Investment Strategy: Fund a Portfolio, Not More Pilots

An AI investment strategy should fund a governed portfolio of business outcomes, not an annual technology budget or an expanding list of pilots. Release capital in stages, compare every use case against a baseline, and preserve the right to scale, redesign, or stop before integration costs make the decision difficult to reverse.

The leadership implication is direct: the CEO should decide the portfolio thesis, capital envelope, and stop rules. Business and technology leaders should earn the next tranche with evidence.

Baca versi Indonesia

Executive deck

The conventional assumption is that strategic conviction should be expressed through a large AI budget and a wide pipeline of experiments. That logic confuses activity with option value. A pilot is useful only when it reduces uncertainty about an outcome, a control, or a capability that matters.

AI changes the investment decision in four ways. Models and vendor capabilities evolve quickly; experimentation is cheap relative to integration; operating costs can move with usage; and value depends on workflow, data, adoption, and governance working together. These conditions reward staged commitments and punish budgets that become entitlements.

Human judgment remains irreducible. A system can forecast cost, simulate scenarios, and monitor performance. It cannot own the company’s risk appetite, decide which customer or workforce consequences are acceptable, or determine which capability must remain strategically distinctive.

Evidence signal

The case for portfolio discipline is stronger than the case for either blanket acceleration or blanket restraint.

Signal	What it means for the decision	Scope and limitation
65% of U.S. CEOs said they expect to allocate at least 10% of capital budgets to AI in the next year.	AI is moving into material capital-allocation territory.	KPMG, 15 October 2026; 400 U.S. CEOs within a global study of more than 1,300 large-company CEOs. Intentions are not realized returns.
Companies in BCG’s survey planned to spend about 1.7% of revenue on AI in 2026, roughly double the prior year.	The cost of weak prioritization rises as the envelope grows.	BCG, 15 January 2026; 2,360 executives in 16 markets and nine industries, including 640 CEOs. Self-reported plans vary by sector.
Only 12% of organizations said they consistently assess AI value against implementation cost across the enterprise.	A larger budget without portfolio instrumentation can scale uncertainty faster than value.	KPMG Global AI Pulse Q3 2026; 2,131 senior leaders in 20 jurisdictions at organizations with at least US$50 million revenue. Self-reported maturity.
IBM reports an average useful AI-model life of about 14 months in its 2026 Tech Leader Study.	Long approval cycles and fixed multi-year assumptions can outlive the underlying asset.	IBM Institute for Business Value, June 2026. The average does not apply equally to every model, workflow, or industry.

These figures are global and U.S.-weighted, not a spending benchmark for Southeast Asian companies. They establish direction: capital is increasing, asset assumptions can age quickly, and measurement discipline is not keeping pace.

The decision question

The decision is not “How much should we spend on AI?” It is “Which uncertainties deserve funded options, which outcomes deserve scale capital, and what evidence will make us stop?”

A useful answer separates three kinds of capital:

Foundation capital keeps shared data, security, identity, evaluation, and integration capabilities reliable.
Proof capital funds bounded tests designed to resolve a named uncertainty.
Scale capital expands a workflow only after value, control, adoption, and unit economics survive a real operating test.

Mixing these categories hides cross-subsidies. A pilot can appear cheap because platform, change-management, and control costs sit elsewhere. A foundation can appear unproductive because its value is distributed across use cases. Portfolio reporting must make both visible.

Point of view

The CEO should approve a portfolio constitution, not a permanent project list. Each investment needs an accountable business owner, a counterfactual baseline, a funding stage, a review date, a control boundary, and a pre-agreed stop condition.

This is not financial engineering around technology. It is a way to preserve strategic option value. The company can move quickly where evidence is strong while limiting loss when a model, vendor, workflow, or market assumption changes.

The portfolio should be small enough for the executive team to understand. If leadership cannot explain why a use case exists, what it displaces, and what would cause it to stop, the portfolio is already too opaque.

Thought process

The recommendation follows five observations.

Spending pressure is rising. CEO surveys show AI moving from experimental budgets into material capital decisions.
Reported value is not the same as measured economic value. Benefits may be real while cost allocation, adoption effort, and risk exposure remain incomplete.
AI assets are perishable. Model capability, pricing, regulation, and vendor architecture can change within a planning cycle.
Integration creates irreversibility. The deeper a workflow depends on one model, data path, or vendor, the more expensive it becomes to change direction.
Therefore, capital should follow evidence in tranches. A portfolio gate makes speed conditional on learning, not on optimism or fear of missing out.

The strongest counterargument is that staged funding can slow a strategically urgent transformation. That risk is real. The answer is not to remove gates; it is to make them fast, explicit, and proportional to reversibility. A low-risk internal assistant should not face the same approval burden as an agent authorized to commit money or communicate with customers.

Four strategic options

No single funding model fits every company. The choice depends on strategic concentration, data readiness, risk, and the cost of waiting.

Option	When it works	Main advantage	Main failure mode
One concentrated platform bet	The capability is strategically differentiating and architecture choices are sufficiently reversible.	Speed, coherence, and bargaining leverage.	Vendor concentration and premature commitment to an unproven operating model.
A broad pilot portfolio	The company is still mapping high-value problems and can keep tests genuinely bounded.	Fast learning across functions.	Pilot theatre: activity rises while ownership and adoption remain weak.
A staged outcome portfolio — recommended default	Several use cases compete for shared capital and infrastructure.	Capital moves toward evidence while weak bets stop early.	Requires disciplined measurement and executive willingness to stop favored projects.
Deliberate wait-and-watch	External capability is improving quickly and the cost of waiting is lower than the cost of lock-in.	Preserves cash and avoids immature commitments.	Capability debt and loss of organizational learning if waiting becomes passive.

The recommended default is a staged outcome portfolio. A concentrated bet may still be right when a capability is central to strategy and switching paths are credible. Waiting may be right when the technology is not the binding constraint.

Five portfolio gates

Every AI investment should answer five questions before receiving more capital.

1. Outcome gate

The use case must improve a decision, revenue flow, cost position, risk exposure, or customer outcome against a defined baseline. Adoption and model accuracy are supporting measures, not the final business outcome.

2. Full-cost gate

The business case must include model usage, data work, integration, security, evaluation, human review, change management, vendor management, and exit costs. A cheap token is not a cheap workflow.

3. Control gate

Decision rights, permissions, evidence logs, human handoffs, incident response, and rollback must match the harm the system can cause. Greater agency requires stronger controls before scale.

4. Adoption gate

The workflow must be used by the people who own the outcome, not merely demonstrated by the project team. Incentives, job design, and management routines must reinforce the new behavior.

5. Strategic-option gate

Leadership must know what becomes difficult to change after the next tranche: vendor dependency, data architecture, workforce structure, customer promise, or regulatory exposure. Commit only when the evidence justifies that loss of flexibility.

Decision rights

AI investment strategy fails when every executive can sponsor a pilot but no one can stop one.

CEO action	What it means
Decide	Portfolio thesis, total capital envelope, strategic outcomes, risk appetite, and enterprise stop rules.
Delegate	Technical evaluation, experiment design, procurement analysis, and day-to-day delivery within explicit limits.
Instrument	Full cost, outcome baseline, adoption, control performance, concentration exposure, and decision logs.
Escalate	High-impact use, material customer or workforce effects, irreversible architecture, weak evidence, or repeated control failure.
Stop	Orphan pilots, unowned benefits, missing baselines, expired assumptions, and projects that cannot explain the next learning objective.

The CFO should challenge capital logic and cost completeness. The CIO or CTO should own architecture adaptability and technical controls. Business executives should own outcomes and adoption. Risk, legal, security, and HR should set constraints and escalation triggers. None of these roles removes the CEO’s accountability for portfolio direction.

The 60-day smallest credible test

Do not begin by redesigning the entire investment process. Apply one portfolio review to three to five active AI use cases that share a business domain or capital pool.

Days 1–10: establish the portfolio record

Name one accountable business owner for each use case.
Record the baseline, intended outcome, full cost estimate, decision rights, dependencies, and next irreversible commitment.
Classify current funding as foundation, proof, or scale capital.

Days 11–40: run and instrument

Measure the business outcome and adoption in the live workflow.
Capture human-review effort, exceptions, control failures, and variable cost.
Maintain a comparable non-AI baseline or a credible before-and-after reference.

Days 41–60: hold the capital gate

Scale when the outcome crosses the pre-agreed threshold, control performance is acceptable, adoption is real, and unit economics remain within the approved range.
Redesign when the outcome is promising but workflow, data, adoption, or cost prevents reliable performance.
Stop when ownership is absent, the baseline cannot be verified, a critical control fails, data rights are unclear, or the next tranche creates concentration without justified advantage.

Success is not “all projects scale.” Success is that capital moves toward the strongest evidence, weak assumptions become visible, and at least one stop or redesign decision can be made without political damage.

Decision gate

Approve the next tranche only when five statements are true:

The business outcome and counterfactual are explicit.
Full cost and variable-cost behavior are visible.
The accountable owner can explain adoption and operating change.
Controls are proportionate to system agency and potential harm.
The organization understands which strategic options narrow after commitment.

If any statement is false, fund learning rather than scale. If the missing evidence cannot be produced within a bounded test, stop.

Implications for decision-makers

For CEOs: make stopping a normal portfolio action, not an admission of strategic failure. The ability to reallocate capital is part of AI fluency.

For boards: ask how many investments changed stage, not only how many use cases exist. Review concentration, reversibility, and outcome evidence alongside risk.

For CFOs: require full-workflow economics. Separate shared foundations from use-case costs without pretending either category is free.

For business leaders: accept ownership of adoption and results. AI cannot rescue a workflow that the operating team will not redesign.

For technology and risk leaders: shorten review cycles by predefining control tiers. Speed improves when teams know which conditions trigger escalation.

Testable hypotheses

A staged portfolio will reallocate capital faster than an annual project budget because review dates and stop conditions are explicit.
Use cases with accountable business owners and counterfactual baselines will reach clearer scale-or-stop decisions than technology-led pilots.
Tracking full workflow cost will reduce the number of apparently successful use cases, but improve confidence in the projects that remain.

These hypotheses should be tested against the company’s own portfolio. Survey evidence cannot substitute for operating data.

Continue reading

Source notes

Evidence reviewed 19 October 2026. Survey results describe respondent intentions and self-reported practices; they do not establish a universal spending benchmark or causal return. The IMF evidence concerns financial-system exposure and market expectations, not the ROI of any individual company’s AI portfolio.

About the author

Antovany Reza writes from an operator’s perspective across general management, growth, partnerships, P&L, go-to-market, AI strategy, and digital transformation. The CEO Decision Lab uses public evidence to help senior leaders make consequential choices without outsourcing accountability.

Discuss the decision

If your AI budget is growing faster than your ability to compare outcomes, bring the portfolio, current evidence, and the next irreversible commitment. The useful conversation is not whether to “do AI,” but which bets deserve the next unit of capital.

LinkedIn · X · Facebook · WhatsApp

Discuss a decision with Antovany Reza