CFO 2035: four futures, different capital choices

Koko: Welcome to Koko Knows. I'm Koko, and today we're opening CFO 2035: four futures, different capital choices. With me are Sam and Max. Sam stress-tests the logic; Max runs the economics. Let's start with trust — specifically, whether it becomes the real constraint on how far a CFO can delegate.

Sam: I'm Sam. Before we get there — trust is already a control requirement. What's the actual claim here, Koko?

Koko: The conditional hypothesis is that by 2030, how far you can safely delegate to agents depends on your ability to establish financial meaning, verify authority and confirm outcomes. By 2035, that capacity could limit growth as much as capital does. PwC frames it as CFO ownership of reliable data and credible decisions — though that's normative guidance, not a forecast.

Max: Max here. Operationally that resonates. The Hackett Group ties segregation of duties and ongoing model monitoring directly to sustained value. What concerns me is cost — more governance infrastructure can slow the decisions you're trying to accelerate.

Sam: Exactly. Autonomy backed by thick controls might cost more than the efficiency gain. So what would actually change this hypothesis?

Koko: If tighter controls consistently erode net returns without catching meaningful errors, the threshold for delegation should shift. NIST's risk framework is voluntary — it doesn't define the right financial control boundary. The enterprise has to set that based on evidence, not framework coverage.

Max: So the practical test is: can you trace a material cash or credit decision from evidence through approval to the recorded outcome, name every owner, and test what an unauthorized action would expose? That's the action — finance controller owns it, and agent authority expands only where that trace is clean.

Sam: That's a real decision gate. Management owns the controls; audit assesses independently. If the trace has gaps, you're not ready to expand scope — and that tension connects directly to what happens when knowledge itself becomes productive capital.

Koko: So here is the second hypothesis on Koko Knows. Agents might be less valuable than the knowledge they share. If every agent queries different definitions of the same customer, you get faster contradictions, not better decisions. The conditional claim is that reusable, maintained data products could create more durable advantage than accumulating more agents.

Max: That rings true operationally. A finance team without agreed policy versions spends more time resolving which number is right than acting on it. The question is whether treating that knowledge as a funded, owned product actually changes the economics, or just adds a new maintenance burden.

Sam: Foundation Capital frames context graphs as a major software opportunity, but that is an investor thesis. SAP describes connecting data with business context for AI systems, but that is a vendor product page. Neither tells us what the return looks like for a specific enterprise.

Koko: Agreed. McKinsey envisions agents helping FP&A anticipate gaps, but those are illustrative cases without a general causal finding. And W3C PROV-O gives a solid vocabulary for provenance, but it does not validate that a decision was correct. The evidence supports the direction, not the size of the payoff.

Max: The risk Sam is pointing at is real. Recorded history can lock in bad exceptions as policy. A credit concession approved during a supply disruption should not become the default. The knowledge layer needs owners who can retire content, not just accumulate it.

Sam: What would change this hypothesis? If a second team reuses a defined data product and exception rates drop measurably, that is a signal. If maintenance costs keep climbing without adoption, you are in stranded intelligence, not compounding enterprise.

Koko: So the action now: CFO and CDO co-fund one decision domain with agreed definitions, lineage, policy versions and a named product owner. The decision gate is explicit. Before funding a second domain, confirm the first improved a pre-agreed economic measure.

Max: That also sets up the 2030 question correctly. By then, finance should know which decisions depend on that layer and whether reuse is actually reducing cost. That is a testable condition, not a forecast. And it connects directly to where we are heading next.

Koko: This is chapter three of CFO 2035: four futures, different capital choices. I'm Koko. The interface is moving — headless SaaS, agents calling other agents — and my hypothesis is that bargaining power shifts toward whoever controls decision context and execution rights. The CFO has to decide what to own, what to rent and what to orchestrate.

Max: I'm Max. That's the right question. SAP's published architecture separates capabilities, a context layer and orchestration. Salesforce documents headless agent invocation through APIs. So the technical direction is visible. But what does that actually change for a finance team running month-end?

Koko: It changes who owns the meaning. An agent can reach five systems and still misread a commercial exception as approved policy. Connectivity solves access. Knowledge engineering addresses meaning. Governance determines what action is allowed. Trusted Finance has to survive the loss of the familiar screen.

Sam: I'm Sam. And that's where I push back. Open protocols — MCP, A2A — don't make application semantics interchangeable. The A2A spec defines communication between agents; it doesn't define what a payment approval means in your jurisdiction. The authoritative ledger and the deterministic calculation still have to exist somewhere.

Koko: Agreed, and that's the concentration risk. If one supplier owns the context, the orchestration and the evidence, switching becomes expensive even when the underlying model is easy to replace. Oliver Wyman flags exactly this — negotiate portability and audit rights before the dependence becomes invisible.

Max: McKinsey's agent-economics interview makes a related point: measure cost per accepted business outcome, including review and rework. Not model consumption — accepted outcome. That's the number that tells you whether the architecture is actually working.

Sam: And the MCP security documentation is worth reading carefully. Authorization risks appear at every tool connection. A working connector is not the same as a valid financial delegation. The CFO needs business-level authorization and evidence at the action boundary, not just a successful API call.

Koko: So the action is concrete: run one workflow across the proposed architecture, reconcile requested against completed actions, export its evidence, test a prohibited action and price a supplier exit. That exercise is more informative than any diagram. Own the context that makes an outcome accountable — which connects directly to our next topic.

Koko: Here's a tension I want to test. Gartner's survey of 204 finance leaders found more AI investment leans toward productivity than decision quality. If that split holds, we may be automating the routine work junior staff learn from without replacing the learning path.

Sam: That's the risk. Automating tasks isn't the same as building judgment. If the exceptions disappear before people learn to handle them, the expertise pipeline quietly empties. Training attendance won't show that.

Max: Operationally, the gap shows up when something breaks. Someone needs to explain why a flagged item is material. If no one junior has seen real exceptions before, escalation goes nowhere fast.

Koko: IBM's execution research, which is survey-based rather than causal, connects AI benefits to operating maturity—not just tooling. That points toward supervised practice on actual workflows, not classroom modules.

Sam: What changes the hypothesis? If teams measure redeployment and decision quality instead of hours saved, and those numbers hold up, the learning-path redesign becomes credible. Without that evidence, the efficiency gain may be borrowing from future capability.

Max: The concrete move is mapping which tasks changed, who owns which decision rights now, and giving early-career staff reviewed cases with real stakes. Six months in, measure adoption and decision accuracy, not seat hours.

Koko: By 2035, in any of the four futures, apprenticeship and accountability may be scarcer than the technology. The compounding scenario assumes a workforce that can judge exceptions at scale. Stranded intelligence is partly what happens when that assumption fails.

Sam: So the decision gate is clear: if redeployment and quality don't move in six months, stop booking the capacity benefit and fund the learning-path redesign first.

Koko: Capital must move as evidence changes—that's the hypothesis here. Physical AI, quantum and compute are advancing, but the CFO 2035 reading asks us to treat that acceleration as a condition to test, not a certainty. Fixed technology roadmaps may be the wrong instrument entirely.

Sam: The challenge is obvious though. DARPA's Quantum Benchmarking Initiative is still evaluating whether quantum systems can deliver computational value exceeding cost—by 2033. That's an evaluation horizon, not a deployment guarantee.

Max: And operationally, energy is a real constraint. The IEA notes that lower energy per AI task can coexist with rising total consumption. So an operator can't just underwrite the device; they have to underwrite the infrastructure stack.

Koko: Gartner's Physical AI Hype Cycle abstract—admittedly only the abstract was accessible—cautions that hype and risk currently outweigh business outcomes. That supports staging capital rather than committing to a roadmap.

Sam: So what actually changes the hypothesis? If you see reproducible economics against a classical comparator, or workforce absorption beyond a single pilot site, then expanding makes sense. Repeatedly missed milestones without useful learning should close the option.

Max: The action I'd prioritise: separate foundation funding from frontier options with explicit gates. And post-quantum cryptography is distinct from quantum computing—NIST has standardised algorithms available now, independent of whether quantum advantage ever materialises commercially.

Koko: Right. CFO and CISO own that cryptographic inventory today. For physical AI, run one bounded pilot, underwrite productive hours and full operating cost, and require a workforce plan before booking any efficiency benefit.

Sam: The decision gate: does the pilot exceed its classical comparator on the economics you actually underwrote? If not, pause before scaling. That connects a 2030 learning position to a 2035 capital posture.

Koko: Last chapter of CFO 2035 on Koko Knows. My hypothesis: governance stops being a periodic review and becomes sustained enforcement at every material agent action. Across all four futures, an agent register without enforcement authority is just a list.

Max: Operationally that matters. If an agent can initiate a payment or change a vendor record, the control has to sit where the action happens, not in a separate dashboard reviewed monthly.

Sam: Fair, but Workday's agent registry was announced as in development, and NIST's agent standards initiative is exactly that, an initiative. Voluntary guidance isn't enforcement, and we can't infer legal requirements from vendor roadmaps.

Koko: Agreed. The Hackett Group's framing helps here: traditional controls, AI-specific controls and monitoring aren't three phases, they're concurrent. If that changes the hypothesis, the signal is an actual regulatory mandate or audit finding tied to agent logs.

Max: The operating question is ownership. Vendors like Microsoft and ServiceNow each offer a control plane, but if obligations differ by jurisdiction, one dashboard may not cover the enterprise.

Sam: And 2030 to 2035, standards could converge or fragment. Either way, the control design has to be portable enough to adapt without rebuilding from scratch.

Koko: So the action now: risk and control owners build an owner-led register, test delegation boundaries, segregation of duties and rollback, and set a gate to revisit when obligations or evidence change.

Max: That closes the episode well. The CFO's mandate across all four futures stays constant: protect, create and orchestrate value. Governance is the operating capability that keeps that mandate credible. Listeners, the full reading is CFO 2035: four futures, different capital choices, on Koko Knows.