The project closed on the same morning that production received its first material data change. The delivery budget expired, the specialists moved away and monitoring had not yet entered an operations plan. Finance saw a completed asset; users saw a living service whose next decision had no funded owner.

Annual project funding assumes bounded scope, a temporary team and a completion event. A production AI capability keeps accumulating obligations: service operation, representative evaluation, data and policy movement, provider change, workflow adoption, security, incidents, control evidence, benefit capture and eventual removal.

Calling those obligations overhead creates an artificial project cliff. Funding visible features while evaluation and recovery wait produces a backlog that looks productive and a service that is not supportable. Yet replacing the cliff with permanent entitlement is equally weak. A team can persist despite low adoption, absent captured value and a cheaper substitute.

The answer is persistent but conditional funding. Stable capacity covers obligations that exist for as long as the decision service operates. Incremental capital is released against evidence about user outcome, service and risk health, unit economics, adoption, reuse and captured value. Stop rights and retirement liability remain inside the contract.

Project discipline still has a place. A one-off migration, regulatory remediation deadline or bounded asset build can have scope, budget, dates and a real finish. The recurring decision service supported by that work remains a product because its outcome and operating obligations continue.

This article’s AI product funding covenant separates five envelopes and joins them through a continuation receipt. It is a management funding mechanism, not an accounting policy. Financial-reporting recognition depends on applicable standards, facts and authoritative accounting judgement.

Original synthesisSeparate five capital envelopes, then join their evidence

The discovery option buys information, not a miniature production entitlement. It names the uncertainty, experiment population, evidence event, maximum exposure, expiry and next decision. Real-options research explains why irreversibility and uncertainty can justify waiting, staging or abandoning rather than committing at once.[1][2]

Capability build funds assets that can reach a definable release state: a workflow configuration, data contract, evaluation system, control, interface or migration. Its evidence covers the integrated product configuration rather than treating model completion as the finish line.

The run and assurance base does not compete with optional feature proposals. It buys the minimum capacity needed to keep the existing promise supportable. This includes people as well as technology: operations, evaluation, domain review, incident command, independent challenge and evidence production.

The outcome and adoption tranche prevents platform funding from floating free of workflow responsibility. A named owner can change work, adoption, service or capacity use. The tranche states how use, decision quality, downstream outcome and captured benefit will influence the next funding choice.

The reserve recognises that switching and removal cost money. A vendor change, policy revision or failed outcome may require migration, rollback, user transition, evidence preservation and dependency removal. Treating retirement as free biases continuation decisions towards sunk cost.

The envelopes do not dictate separate cost centres or financial-reporting categories. They preserve management visibility over distinct obligations and options. Applicable accounting treatment requires separate analysis; the covenant cannot decide capitalisation.[15]

Funding cadenceKeep the base stable and make increments answer to evidence

The continuation receipt is quarterly or event-triggered, not an artificial monthly reauthorisation of a stable production service. It records decision population, observed outcome, causal confidence, adoption, service and risk health, unit cost, captured benefit, remaining uncertainty, the next option, stop condition and retirement liability.

The run base continues while the product remains authorised because abrupt withdrawal would itself create operational risk. Continuation does not protect its current size or roadmap. Evidence can expand the population, hold investment, redirect an outcome tranche, narrow scope, repair the service or finance retirement.

Incremental tranches are decisions, not rewards for activity. Teams do not earn capital by producing more features, prompts or experiments. The evidence condition states what uncertainty will change, which owner can act on the result and what alternative becomes available. Output metrics cannot stand in for an outcome or operating receipt.

Product-flow disciplines help finance smaller batches and economically sequenced work rather than one annual bundle.[6] This does not require fragmenting a stable service into tiny funding events. It separates continuing capacity from optional increments whose value and urgency differ.

Sunk cost is absent from the forward case. Build expense may explain the current configuration but cannot justify new capital. The receipt compares expected outcome, risk, cost and retirement liability from today against narrowing, substitution and exit. A lower-cost alternative reopens the product boundary even after a successful build.

Shared reuse appears in both service economics and allocation. A common component can lower marginal cost across products, but reuse does not prove outcome. Each consuming workflow retains an adoption and capture receipt, while the shared service publishes its cost, reliability, change and concentration evidence.

Uncertainty disciplineBuy information before buying irreversible exposure

A discovery option is appropriate when one answer can materially change the investment path. It begins with the decision that follows evidence. If no plausible result would alter build, population, control or abandonment, the activity is research without an option boundary.

Technology-positioning research shows how options can stage uncertain investment.[3] The label must not excuse indefinite experimentation. Work on real-option boundaries warns that open-ended learning without expiry, exclusivity or abandonment discipline is not necessarily an option.[4]

The tree names the uncertainty, test, maximum spend, exposure ceiling, expiry and branches. A positive result may justify a build tranche. A mixed result can narrow the population or buy one more bounded test. A negative result closes the option and funds removal of experimental dependencies.

Reversibility affects price and sequence. Cheap, reversible probes can precede detailed design. A consequential live test requires stronger safeguards and a credible fallback. Waiting also has a cost, including delayed benefit or decaying access, so the receipt includes the value and deadline of information.

The project counter-case clarifies the boundary. A data-centre migration with one destination, cutover and acceptance event is project-funded. The AI decision service running on it keeps its product covenant because monitoring, adoption, change and outcome continue after infrastructure acceptance.

Negative controlsPrevent features and entitlement from consuming the covenant

The feature raid begins before a visible demonstration. A new interface consumes engineers reserved for reliability, evaluators are reassigned to content, and the retirement reserve fills a provider bill. The feature ships; monitoring gaps and recovery debt appear later in another budget.

Ring-fencing protects a minimum obligation, not a department. The covenant names run, evaluation, incident, challenge and retirement capacity that cannot be diverted without a decision to restrict the product. If leaders choose the feature, they also sign the reduced service or population boundary created by that choice.

Production ML research documents hidden debt across data dependencies, feedback, configuration and system interaction.[9] NIST likewise frames ongoing monitoring, response, change and decommissioning as lifecycle work.[7][8] The base gives that work an explicit funding home.

Permanent entitlement is the opposite capture failure. A team reports uptime and output while eligible adoption is low, no incremental outcome is visible, cost rises and substitutes improve. Persistent ownership keeps evidence and exit responsibility together; it does not exempt the product from narrowing or termination.

False capitalisation is a boundary error. Management may call an envelope build investment or run base to make trade-offs visible. Those labels do not determine whether expenditure is recognised as an asset or expense. Finance must apply relevant standards and local requirements to actual facts, with professional advice where needed.

Continuation evidenceMake the next capital decision from a joined receipt

The receipt begins with population and outcome, then records causal confidence. A before-and-after movement may be encouraging without proving attribution. Finance sees whether the comparison, mechanism and confounders support expansion or only further observation.

Adoption is eligible use, bypass, correction and sustained workflow change, not account activation. Captured benefit has an operating action and finance confirmation. Potential time becomes realised value only when demand, service, capacity, staffing or cost changes.

Service and risk health sit beside outcome because value earned through unstable or non-compliant operation is not supportable. Reliability objectives should drive explicit trade-offs under empowered ownership.[10] Incidents carry remediation and future liability rather than disappearing after recovery.

Unit cost includes provider, platform, data, human review, exception, support, control and change effort. Marginal economics show what expansion costs at the next population boundary. Reuse credits are visible but do not double-count a common benefit across products.

The next option states the actual choice: expand one segment, hold and observe, repair service, test an alternative, narrow, migrate or retire. The stop condition is forward-looking. Retirement liability estimates transition, evidence preservation, contract exit, dependency removal and residual support.

Worked successNarrow the credit-file product when one segment lacks value

The credit-file product begins with a discovery option covering a bounded population. The uncertainty is whether source assembly and provenance reduce analyst rework without weakening evidence. The option funds representative observation, analyst corrections and a manual fallback, then expires at a release decision.

A positive result opens a build tranche for workflow integration, evaluation, controls and service recovery. Release then activates a ring-fenced run and assurance base. The product owner cannot trade that base for additional document features without changing population or service obligations.

The outcome tranche funds adoption support and operating redesign. Its receipt joins eligible use, file completeness, provenance, correction, cycle time, unit cost and captured capacity. The lending operations owner, not the product team alone, confirms how capacity is used.

Evidence later differs by segment. Standard files show incremental outcome and acceptable cost. A specialist segment has high exception effort and no credible improvement over manual preparation. The covenant expands the standard population, narrows funding away from the specialist segment and preserves the manual route.

Narrowing creates retirement work even though the whole product continues. The reserve removes specialist routing, preserves required evidence, informs users, adjusts vendor volume and verifies that automated processing has stopped for that segment. It also funds a later alternative test if new evidence justifies one.

Adoption research stresses complementary organisational capabilities rather than technology purchase alone.[11] Dynamic-capability research frames sensing, seizing and transforming as continuing managerial work.[5] The covenant pays for those moves while requiring evidence that they improve the decision promise.

Finance operating rhythmDecide at evidence events without destabilising the service

A finance cadence has three clocks. The base is reviewed against operating obligation and authorised population. Optional increments are reviewed when their named evidence arrives. Portfolio allocation compares products on outcome, exposure, economics and alternatives at a regular executive interval. These clocks can meet without forcing every service through one annual gate.

Base review asks whether the current service remains authorised and what capacity the promise requires. A fall in demand can shrink support and provider commitments. New regulation or concentration may increase assurance and reserve needs even when feature scope is unchanged. The base is evidence-responsive, not fixed forever.

Tranche review begins from the prior decision. It tests whether promised evidence exists, whether uncertainty changed and whether the next option still matters. A missed receipt does not automatically cancel a live service, but it prevents optional expansion and triggers repair, narrowing or an explicit exception.

Portfolio review removes local sunk-cost bias. It compares incremental capital with other products, manual configurations and exit. Shared dependencies and concentration are visible across product receipts. A common provider discount may improve unit economics while making the retirement reserve and switching option more valuable.

Finance and the product owner make different contributions. Finance tests economic consistency, allocation, evidence and capture. The product owner integrates service, outcome, change and retirement consequences. The workflow owner controls adoption and operating capture. Risk and assurance challenge the evidence within their mandates.

Artificial precision is avoided. Causal confidence can be high, bounded or weak with reasons. Unit cost can show ranges where exception demand is volatile. Retirement liability can state known commitments and unresolved dependencies. Conditional funding needs decision-useful uncertainty, not invented certainty.

Government evidence also shows that funding growth can outpace procurement readiness and impact measurement.[12] The covenant responds by tying incremental capital to adoption, evidence and an executable next decision rather than assuming that spending or deployment demonstrates impact.

For applicable high-risk systems, lifecycle, monitoring and deployer obligations can create continuing work that cannot be treated as an optional feature.[13] Operational-resilience principles similarly keep dependency, continuity, response and learning visible.[14] The envelope design turns those obligations into funded capacity without giving them an unlimited claim on capital.

Budget ownership follows the product promise rather than the organisational home of each specialist. A platform team may employ the evaluator and operations may employ incident staff, yet their reserved capacity remains visible in the product covenant. This prevents distributed cost from becoming invisible responsibility and lets portfolio leaders see the full forward commitment.

Exceptions are funded and time-bounded. If a control, data source or provider cannot meet the normal covenant, the receipt states the temporary population, compensating work, cost, owner and expiry. Repeated renewal becomes evidence that the base or product boundary is incorrectly designed, not a routine way to defer the investment decision.

Decision quality also matters after a funding route is chosen. Finance records what evidence was expected, which alternative was rejected and why. At the next event, reviewers can distinguish a reasonable decision under uncertainty from a covenant that quietly changed its premise. Learning then improves capital allocation rather than merely explaining variance. The record also prevents later success from making an undisciplined original commitment appear justified.

Original synthesisChoose the funded envelope and its next evidence condition

The AI Product Funding Covenant Designer begins from lifecycle state. An uncertain pre-production proposition can buy discovery. A live authorised service with strong receipts can fund its base and the next outcome tranche. A weak service or evidence boundary holds expansion while repair is funded.

Population matters. When value is credible for one segment and absent for another, narrowing preserves the proven promise without defending the original scope. The unfunded segment receives an explicit manual, substitute or retirement route rather than falling into unsupported operation.

Stop applies to new funding when forward value is not supportable, incident liability is unacceptable, no capture owner can act, or a better alternative dominates after switching and retirement. Stop still finances safe retirement. It never means abandoning users, records or dependencies.

The designer does not issue an accounting conclusion. Tranche amount is management intent, and the returned envelope describes capacity and evidence. Accounting recognition, measurement and disclosure remain outside the instrument.

Read the route as a capital decision, not a grade

The healthy live-service case funds two things for different reasons. Its run and assurance base continues because users and controls depend on the authorised service. Its outcome tranche continues because joined evidence supports a further adoption or population decision. If the second case weakens, finance can hold the tranche without destabilising the first obligation.

The uncertain case buys only enough capability to answer the named question. It does not acquire a production population, permanent team or implied follow-on budget. The option receipt states who will decide at expiry and what happens to data, vendor access and experimental workflow if the proposition is abandoned. Cleanup is part of option cost.

The repair case distinguishes a problem worth fixing from a service worth expanding. Existing users retain a supportable fallback while the ring-fenced repair closes recovery, monitoring, risk or outcome evidence. Leaders cannot compensate for the gap by adding a visible feature. Expansion resumes only when the missing receipt changes the funding decision.

The narrowing case treats heterogeneous value as useful evidence. A product need not defend every original segment to retain a proven one. Costs, outcomes and exception demand are compared at the population boundary. The covenant pays for transition out of the weaker segment and records whether a future discovery option could reopen it.

The stop case ignores the amount already spent. It compares future capital, future outcome, alternatives and liability. A live capability may still require several funded periods of user transition, records preservation, vendor exit and verification. Retirement finance is evidence of disciplined termination, not continued product advocacy.

The one-off migration case can coexist with all five routes. Migration scope, cutover and acceptance remain inside a project budget. Delay or overrun is managed as project performance. After acceptance, the recurring AI product’s run, evidence and outcome obligations continue under the covenant, preventing the infrastructure finish line from becoming a second product cliff.

ConclusionFund continuity, evidence and exit as one conditional promise

Annual projects create accountability for bounded work, but they cannot be the sole funding unit for a recurring AI decision service. The obligations that matter after release are not residual overhead. They are the cost of maintaining evidence, service, authority, change, outcome and safe exit.

Persistent funding is not permanent entitlement. The five-envelope covenant protects a supportable run and assurance base, stages uncertain discovery and build, conditions outcome and adoption capital, and reserves capacity for change and retirement. Each envelope has an evidence boundary and a next decision.

The continuation receipt keeps outcome, causal confidence, adoption, service, risk, unit economics, capture, uncertainty and liability in the same capital conversation. It prevents a good uptime number, attractive feature or sunk build cost from dominating. It also prevents weak causal evidence from cancelling an operational duty that still exists.

Project controls remain useful for migrations, deadlines and assets with a real completion event. The product covenant continues around the decision service those projects support. Finance can therefore compare bounded delivery and persistent operation without forcing one to impersonate the other.

Conditional routes make the model credible. Finance can buy information, fund a healthy base, hold and repair, narrow to a proven population or stop new capital while financing retirement. Expansion is one possible result, not the premise of product funding.

The durable principle is simple: capital should follow the living decision promise and the evidence that can change it. A product earns neither a cliff at release nor protection from termination. It receives the capacity to operate responsibly and the obligation to prove, adapt, narrow or end.

Evidence spineReferences

1. Pindyck, R. S. (1990), ‘Irreversibility, Uncertainty, and Investment’, NBER Working Paper 3307; subsequently published in Journal of Economic Literature (1991), 29(3), 1110–1148. doi:10.3386/w3307.
2. Dixit, A. K. and Pindyck, R. S. (1994), Investment Under Uncertainty, Princeton University Press. Princeton University Press.
3. McGrath, R. G. (1997), ‘A Real Options Logic for Initiating Technology Positioning Investments’, Academy of Management Review, 22(4), 974–996. doi:10.5465/amr.1997.9711022113.
4. Adner, R. and Levinthal, D. A. (2004), ‘What Is Not a Real Option: Considering Boundaries for the Application of Real Options to Business Strategy’, Academy of Management Review, 29(1), 74–85. doi:10.5465/amr.2004.11851715.
5. Teece, D. J. (2007), ‘Explicating Dynamic Capabilities’, Strategic Management Journal, 28(13), 1319–1350. doi:10.1002/smj.640.
6. Reinertsen, D. G. (2009), The Principles of Product Development Flow, Celeritas Publishing. Celeritas Publishing.
7. National Institute of Standards and Technology (2023), ‘AI Risk Management Framework: Core’. NIST AIRC Core.
8. National Institute of Standards and Technology (2023), ‘AI RMF Playbook’. NIST AIRC Playbook.
9. Sculley, D. et al. (2015), ‘Hidden Technical Debt in Machine Learning Systems’, NeurIPS, 2503–2511. Google Research.
10. Thurgood, S., Ferguson, D., Hidalgo, A. and Beyer, B. (2018), ‘Implementing SLOs’, in The Site Reliability Workbook. Google SRE Workbook.
11. Organisation for Economic Co-operation and Development, Boston Consulting Group and INSEAD (2025), The Adoption of Artificial Intelligence in Firms. doi:10.1787/f9ef33c3-en.
12. Organisation for Economic Co-operation and Development (2026), ‘Adopting and governing AI in government’, Digital Government Outlook. OECD.
13. European Union (2024), ‘Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence’. EUR-Lex.
14. Basel Committee on Banking Supervision (2021), Principles for Operational Resilience. Bank for International Settlements.
15. International Accounting Standards Board, ‘IAS 38 Intangible Assets’. IFRS Foundation.