01 Start with a value decision
The CFO’s capital mandate is to protect, create and orchestrate enterprise value. Trusted Finance is central to that mandate. The capital question is not how much to spend on AI. It is which enterprise capability deserves the next commitment, what decision it will improve and how management will know whether the investment worked.
A coherent investment case joins five things that are often funded separately: the business workflow; its data products; its knowledge and policy; the people who will use and govern it; and the technology that enables action. Every case should also identify what existing expense or activity can be retired.
This agenda is Koko’s proposed management practice, informed by the research. It is not a prescribed allocation percentage or an accounting policy. IBM emphasizes finance execution maturity, Kearney connects capital to strategic objectives, and Forvis Mazars warns that isolated spending can fail to support the integration needed for scale. IBM Institute for Business Value ↗ Kearney ↗ Forvis Mazars ↗
02 Four funding purposes, four evidence standards
| Funding purpose | What belongs here | Evidence required for the next commitment | How to avoid waste |
|---|---|---|---|
| Trusted Finance: integrity and resilience | Financial reporting and reconciliation, liquidity and asset protection, critical data controls, identity, security, continuity and required remediation. | Material exposure, accountable owner, tested remediation and an operating plan. | Prioritize by consequence and dependency; do not invent speculative revenue to justify essential work. |
| Reusable capability | Data products, definitions, policy knowledge, evaluation assets and workforce proficiency. | A specific first consumer, a credible second use and maintained quality and cost. | Stop infrastructure without consumers; avoid forcing every use into one oversized platform. |
| Proven applications | Workflows with measured benefits and feasible operating support. | Repeatable net outcomes, acceptable exceptions, adoption and capacity to expand. | Count integration, reviews and support; retire displaced tools or processes. |
| Strategic options | Bounded exploration of quantum, physical AI, new interfaces or emerging ecosystems. | A material uncertainty to resolve, limited commitment and a dated exercise, redesign or stop decision. | Buy useful learning or access; do not continue merely because the technology is interesting. |
A company’s mix depends on its exposures, maturity and strategy. Under the stranded intelligence scenario, the next dollar may move from additional agents to reliable knowledge, integration and workforce adoption. Under compounding enterprise, expansion of proven value streams can earn more capital. Under slower progress, preserve relevant options and improve current operations. The four scenarios are designed to change the portfolio, not decorate the strategy deck.
03 What to have on the table by day 90
| Workstream and accountable partnership | Days 1–30: establish | Days 31–60: build and test | Days 61–90: make the decision |
|---|---|---|---|
| Trusted Finance — controller with treasury, risk and technology owners | Map material reporting, cash and execution risks; identify control and evidence gaps. | Test reconciliations, financial meaning, delegated approvals, exception handling and recovery for the selected workflow. | Confirm the operating boundary and remediation budget; retain independent challenge and recurring control evidence. |
| Enterprise value portfolio — CFO and business leaders | Map two material decisions, baselines, constraints and existing spend. | Define scenarios, value hypotheses and full-cost comparisons. | Approve scale, redesign, defer or stop; transfer resources accordingly. |
| Data foundations — CDO/CIO with named domain owners | Identify critical entities, conflicting definitions, lineage and access gaps for those decisions. | Deliver a governed data product with a consumer, owner and service expectations. | Accept it only if it meets decision needs; fund a second use where justified. |
| Knowledge engineering — policy owners with data and operations teams | Map rules, contracts, exceptions, effective dates and expert dependencies. | Create a small maintained knowledge package and a reviewed exception set. | Evaluate decision improvement and maintenance cost; retire unreliable or stale content. |
| Agent governance — risk/control owners with technology and business owners | Inventory agents, tools, identities, permissions, money movement and external dependencies. | Test boundaries, evaluation cases, logs, escalation and rollback. | Approve a defined operating envelope, a control owner and recurring support budget. |
| Workforce — CHRO, CFO and operational leaders | Map work that disappears, changes or grows; identify judgment and skill gaps. | Run supervised work-based learning and redesign roles and incentives. | Verify proficiency, employee adoption and the destination of released capacity. |
| Architecture and procurement — CIO/CISO, procurement and finance | Map systems of record, APIs, orchestration, licenses and exit exposure. | Compare native, external and owned orchestration on one workflow. | Contract for permissions, evidence, service levels, consumption visibility and portability. |
| Frontier options — business sponsor with R&D/COO/treasury | Identify a relevant exposure beyond current finance AI. | Run a bounded comparison or partner assessment. | Record what was learned and whether the option deserves another commitment. |
These workstreams share an outcome and a decision date. Their budgets should reveal dependencies. A pilot that requires a new data product and permanent review team is not a low-cost pilot simply because its model bill is small.
04 Make trust part of the investment decision
Every material proposal should explain the value it seeks to create and the value it could expose. Identify the financial records and assumptions it depends on, the actions it may take, the accountable business and control owners, the failure consequences and the conditions that require escalation or suspension.
Define a proportionate trust threshold before scaling. For one workflow, it may require a reconciled data source, approved definitions, tested permission boundaries, recoverable execution and evidence that exceptions reach an accountable person. Another may safely operate with estimates and human review because it only prepares an internal scenario. Apply consequence and materiality rather than one universal threshold.
Fund protection on its own evidence. A necessary reporting correction, liquidity control or resilience measure need not invent growth revenue to earn priority. For discretionary risk reduction, show exposure ranges and assumptions, then test whether the control reduces that exposure. Keep modeled avoided loss separate from realized cash savings. Do not count the same outcome again as growth, productivity and protection.
The scale decision joins three judgments: the economic case is credible; the Trusted Finance requirements fit the intended use; and the enterprise can sustain the people, knowledge and controls behind the workflow. If any fails, narrow the scope, repair the gap or stop. Hackett’s control guidance and NIST’s lifecycle approach inform this discipline; the thresholds remain company-specific management judgments. The Hackett Group ↗ NIST ↗
05 Turn data foundations into products people use
Start with a decision whose information is unreliable or expensive to assemble. A cash-conversion product might combine receivables, customer terms, disputes, payment history, inventory and supplier commitments. A customer-economics product might connect revenue with service costs, returns, credit exposure and capacity.
Every product needs a named owner and consumer; business definitions; source lineage; permitted uses; freshness and quality expectations; issue resolution; a cost to serve; and a retirement condition. Scale by adding meaningful consumers and maintaining dependable service, not by increasing the count of tables in a catalog.
The first product should not wait for the entire enterprise to be perfect. Conversely, urgent use cases should not keep recreating incompatible customer, supplier or profit definitions. Fund the common pieces that remove repeated friction. Data-product principles are established; their economic value has to be demonstrated in the particular enterprise. Zhamak Dehghani ↗
Acceptance test: another team can use the product for a real decision without a bespoke reconciliation project, and the original owner can maintain it at an acceptable cost.
06 Underwrite the economics honestly
Use a cash-flow case for economic evaluation and a separate capability scorecard. Include implementation, integration, data preparation, knowledge maintenance, model and platform consumption, controls, human review, training, support, transition and exit costs. Compare against the best realistic alternative, which may be conventional automation or a simpler process change.
Count capacity only once. Hours released become cash savings only when an actual spending decision changes. If people move to growth or risk work, evaluate the resulting outcome separately. Avoid claiming the same benefit as both headcount savings and productive redeployment. Agent costs can grow with the amount and complexity of work, even when per-unit model prices fall. McKinsey ↗
Illustration—assumptions only: suppose 100,000 cases a year use 20 minutes of manual handling at a loaded labor rate of $60 an hour. The baseline effort is 33,333 hours, or $2.0 million of labor capacity. A proposed workflow reduces handling to 8 minutes per case, releasing 20,000 hours, with a capacity equivalent of $1.2 million. If $500,000 of annual platform, data, control and support cost plus $600,000 of implementation are incremental, the first-year capacity-equivalent benefit after those costs is $100,000. That is not $100,000 of cash savings unless the capacity is actually converted. If additional review increases handling to 12 minutes, the first-year capacity-equivalent result falls to negative $300,000. Any incremental error, transition or redundancy costs would further change these results.
Use discounted cash flows, the company’s hurdle rates and scenario assumptions for a real investment decision. This simple example demonstrates why review effort and benefit conversion can dominate a cheap inference bill; it is not a valuation recommendation.
07 Sustain agentic governance after launch
Treat governance as funded operations. Record an accountable owner, purpose, deployment context, permitted data and actions, underlying identity, model and tool versions, evaluation results, cost limits, and escalation and retirement rules for each material agentic workflow.
Define the operating envelope explicitly. An agent may gather evidence and prepare a payment proposal while a separate authorized person or controlled process approves it. Transaction limits, counterparty changes and exceptional conditions may alter the required approval. Enforce the policy at the system that can perform the action, not solely in instructions to a model.
Maintain segregation of duties across the combined human and machine workflow. Test adversarial inputs and excessive permissions, evaluate material changes, monitor drift and incident severity, and rehearse rollback. Preserve a record of what was requested, approved and executed. Structured business rationale and evidence are useful; a model’s asserted internal reasoning is not an auditable substitute.
The Hackett Group discusses traditional, AI-specific and monitoring controls. NIST supplies lifecycle risk guidance, and MCP documentation identifies connector-level security risks. These inform the design but do not replace company-specific control judgment. The Hackett Group ↗ NIST ↗ Model Context Protocol ↗
Acceptance test: the owner can show why an action was allowed, demonstrate that a prohibited action is blocked and recover when a dependency fails.
08 Upskill toward a different contribution
Give learning a work assignment. An FP&A professional might build and defend a scenario, assess model error and recommend an intervention. An accountant might evaluate policy exceptions and improve the evidence record. A business partner might trace a customer decision through revenue, margin, cash and service consequences.
Fund protected learning time, expert review and rotations into the business. Keep experienced practitioners involved in the knowledge being encoded. Pair early-career employees with reviewed exceptions and live operating decisions so automation does not remove the route to judgment.
Measure demonstrated proficiency, adoption and outcomes. Do not treat a course certificate as permission for consequential autonomous work. Deloitte’s integrated transformation approach and Accenture’s CFO research reinforce the need to connect people, technology and operating change. Deloitte ↗ Accenture ↗
09 Run a portfolio, not a permanent pilot program
Monthly, review material outcomes and incidents with operating owners. Quarterly, make capital decisions: expand a proven pattern; repair a prerequisite; replace an expensive dependency; retire unused products; or stop a thesis that has failed its test. Reopen strategic scenarios when the technology or commercial environment changes materially.
The board should see value realized, value still hypothesized, capital committed, the next irreversible decision, residual exposure and the capability being built. Do not blend those states. An owned ontology or reusable workflow may be economically valuable without qualifying for balance-sheet recognition; accounting treatment requires a separate policy assessment.
The discipline is simple to state and difficult to sustain: fund Trusted Finance and the capabilities that help the enterprise protect value, make better choices, demonstrate what followed, and make the next commitment on the evidence.