AI can reduce the effort required to explore, specify and build a product idea. It cannot determine whether that idea deserves scarce capital, customer attention or organisational change. As execution becomes cheaper, investment judgement becomes more important because the organisation can pursue more plausible options than it can responsibly fund.

More capacity does not create a priority

When the cost of producing a prototype falls, more ideas become executable. That is valuable, but it changes the constraint. The portfolio can fill with convincing demonstrations, local automations and partially funded products before leaders have agreed which customer or operating outcomes matter most.

DORA's 2025 research describes AI as an amplifier of the surrounding organisational system. The finding is bounded to software delivery, but the investment implication travels. Faster execution amplifies clear priorities and strong feedback. It also amplifies weak selection, fragmented ownership and work that survives because starting is easier than stopping.

A buildable idea therefore needs two separate judgements. Can the organisation create and operate it? Should the organisation allocate attention, change capacity and continuing cost to it? AI can improve evidence for both questions. It cannot reconcile competing ambitions or accept accountability for the trade-off.

This distinction matters most when local teams can begin without waiting for central capacity. A small experiment may be cheap, yet the aggregate portfolio still consumes customer attention, data access, review time, integration capacity and future support. Leaders need visibility before inexpensive starts become expensive commitments.

SourcesDORA

Begin with the decision the investment must improve

Product proposals often begin with a feature, agent or platform. Investment should begin with a decision or outcome that is currently expensive, slow, unreliable or strategically limiting. This makes value observable before a preferred implementation turns into a commitment.

NIST's AI Risk Management Framework asks organisations to document business value, intended purpose, scope and risk tolerance. It is a risk framework rather than a portfolio method, but it reinforces a sound sequence: establish purpose and context before deciding whether development or deployment should proceed.

Name who experiences the problem, what changes if it is solved and which measure would alter the next funding decision. A claim such as 'use AI in customer operations' is too broad. A claim such as 'reduce avoidable rework in renewal preparation while preserving accountable approval' can be tested against current performance and alternatives.

The baseline is part of the investment case. Without current elapsed time, error, failure demand, cost or customer behaviour, a team may prove that a new system works while remaining unable to show that it improves the operation. Evidence should describe the present constraint before it celebrates the proposed capability.

SourcesNIST

Compare the AI option with credible alternatives

An AI-enabled product is rarely the only route to an outcome. A policy change, data repair, design-system improvement, deterministic workflow or removal of unnecessary work may create more value with less operating complexity. The comparison must include doing nothing and improving the current system.

This is where rapid prototypes are useful. They can test whether a capability changes behaviour, whether context is available and whether users trust the proposed role. A prototype is evidence about assumptions. It is not evidence that the complete investment is desirable or operationally ready.

Require every option to expose its continuing obligations: human oversight, data stewardship, evaluation, support, model and infrastructure cost, security, change management and exit. The most impressive demonstration can become the least attractive investment once these obligations are visible.

Comparison also protects the organisation from solution momentum. Once a named technology has a sponsor, discovery can become a search for confirming evidence. A credible review gives the simpler route equal design attention, applies the same outcome measures and records why added adaptability is worth added uncertainty.

Connect customer evidence to operating economics

Customer value and economic value must meet. DORA's user-centricity guidance recommends using feedback and product-performance measures to keep AI-assisted work directed towards user needs. Provider-associated research should not be treated as a universal effect size, but the operating principle is credible: capacity needs a clear value signal.

The FinOps Foundation defines unit economics as relating technology use and cost to a measurable unit of a product or service. Its framework is concerned with technology economics, not complete product value. Used carefully, it helps leaders connect adoption, service outcome and marginal operating cost rather than approving a build budget without understanding the cost of success.

Choose a small set of linked measures. Include the outcome the customer or operation receives, the behaviour that indicates usefulness, the fully loaded cost of producing that outcome and the risk or quality constraint that must not deteriorate. Avoid compressing the system into one return number before uncertainty has been reduced.

Economics should be observed at the useful unit. Cost per model call may help engineering, but it rarely explains the investment. Cost per completed case, accepted recommendation or retained customer connects technical consumption to an outcome. It also reveals when adoption grows faster than value or when human correction dominates the apparent automation gain.

Fund the next uncertainty, not the complete promise

Early investment should purchase evidence. The next tranche might test demand, workflow fit, context readiness, authority, technical feasibility or unit economics. Each stage should have a named decision date and an explicit choice to continue, scale, redirect, pause or stop.

GOV.UK's agile governance guidance argues for evidence-based decisions at the right level, with teams clear about their authority and escalation boundaries. The public-service context is specific, but the governance pattern is useful. Frequent decisions require explicit authority, not a succession of steering meetings that can discuss evidence without changing the investment.

Stopping is a productive outcome when an important assumption fails. The portfolio gains capacity, avoids continuing cost and preserves what was learned. Leaders should make the evidence reusable so a rejected route does not have to be rediscovered by another team six months later.

This requires portfolio hygiene. Every active investment should have a current outcome, accountable owner, next uncertainty, evidence date and decision state. Work without these elements is not necessarily wrong, but it is no longer governable as a deliberate investment. It should be clarified, contained or stopped before more capacity is added.

SourcesGOV.UK

Use a five-part investment test

Apply the following test before approving a major build and at each evidence-based funding point. Its purpose is to expose the quality of the investment decision, not to manufacture certainty.

End each review with one explicit portfolio response: explore a bounded uncertainty, continue the current commitment, scale when evidence and operating capacity support it, redirect while preserving useful learning, or stop and return capacity to the portfolio. The response should name its owner, evidence threshold and effect on funding and team authority.

AI expands what a team can attempt. Leadership still has to decide which change matters, what evidence earns further commitment and what the organisation will stop doing to create room. The fastest builder is not automatically the strongest portfolio. The stronger portfolio converts greater capacity into fewer, better and more reversible investments.

  • Outcome: Which customer, operational or strategic decision will improve, and for whom?
  • Evidence: Which critical assumption is known, inferred or still untested?
  • Alternative: What simpler intervention could produce the same outcome?
  • Economics: What are the continuing cost, ownership and risk obligations if usage grows?
  • Authority: Who can authorise exploration, then continue, scale, redirect or stop the investment when evidence changes?

Sources and further reading

  1. DORAState of AI-assisted Software Development 2025
  2. DORAUser-centric focus
  3. NISTAI Risk Management Framework Core
  4. FinOps FoundationCapability: Unit Economics
  5. GOV.UKGovernance principles for agile service delivery