CFO Future Guide check-in: the first seven weeks

Koko: Welcome to CFO Future Guide, the monthly check-in series. I'm Koko, and joining me is Chris. Seven weeks ago we published CFO 2030: The Enterprise Value Orchestrator — a set of conditional hypotheses about where the CFO role could go by 2030 and 2035. This first check-in covers September 14th through October 31st, 2026. The headline: three hypotheses strengthened, three held, none weakened or challenged.

Chris: Good place to start, Koko. And the first hypothesis is probably the one with the sharpest evidence this window — trust as the capacity to delegate. What's the claim?

Koko: The claim is that by 2030, the ability to establish valid authority and verify outcomes could determine how far a CFO can safely delegate to agents — and that by 2035, trust may constrain growth as much as it constrains compliance. This one is rated Strengthened.

Chris: What moved it? Because strengthened is a meaningful call after only seven weeks.

Koko: Deloitte's third-quarter CFO Signals survey — 200 large-company North American CFOs — found nearly 80% say finance has a formal role in AI decisions. But 53% review AI costs after the fact, and 74% need days to weeks to pause excess usage. That's the exact gap the hypothesis predicted: authority without the controls to make delegation safe.

Chris: So finance owns the approval but not the dial. I'd push back a little — Deloitte's sample is companies above one billion dollars in revenue, fieldwork dates aren't stated on the page, and secondary coverage from Inc. couldn't even be verified because the article returned an error. That's a thin base for a strengthened call.

Koko: Fair. The rating doesn't claim the gap is universal. It reflects that the pattern Deloitte found — authority ahead of control — is exactly what the hypothesis said to watch for. And SecurityBrief's trade coverage, which is secondary and unverified for sample size, adds that CFOs are already prioritizing trust over speed. That's directionally consistent, even if neither source is definitive.

Chris: What would have pushed it in the other direction — say, to held?

Koko: If the survey had shown real-time control capability was common, or if closing the control gap demonstrably slowed decisions without improving outcomes, the hypothesis's cost-of-trust argument would look weaker. We don't have that evidence, but the challenge remains: more controls aren't automatically better. They have to earn their cost.

Chris: So what does a CFO actually do with this before next month?

Koko: The POV's move is specific: trace one material cash or customer-credit decision from source evidence through approval to the recorded outcome. Name the owners at each step and test what happens when an unauthorized action is attempted. Don't expand agent autonomy beyond what that trace can actually support. Watch for Deloitte's fourth-quarter CFO Signals in January — that should tell us whether the gap between authority and real-time control is closing.

Koko: The knowledge hypothesis says reusable data products and maintained business context could create more durable advantage than simply accumulating agents. This window rates it Held, and I want to be direct about why: the evidence is thin, and a held rating with thin evidence is not quiet confirmation.

Chris: What actually came in this window?

Koko: Two sources. MIT Technology Review Insights, in content sponsored by Atlan, a data-catalog vendor, found fewer than one in five organizations extracting transformative value from AI and attributes that gap to missing enterprise context. ScottMadden published a practitioner role framework that adds knowledge curation and data stewardship as emerging shared-services functions. Both point in the right direction. Neither measures outcomes.

Chris: The Atlan piece has a vendor interest and an unstated sample size, so you can't treat it as independent research. ScottMadden is a role taxonomy, not a benefit study. What would actually move the rating in either direction?

Koko: A primary study linking AI outcomes to data-product or knowledge-layer maturity specifically in finance. Or finance organizations posting for knowledge or data-product owner roles at meaningful scale. Neither showed up. And the POV's own challenge hasn't gone away: context graphs remain an investor thesis, not demonstrated returns. If a finance team tried to build one, the maintenance cost and adoption risk belong in the business case from day one.

Chris: So the risk isn't that the idea is wrong. It's that there's still no evidence strong enough to justify scaling it. A team could invest heavily in a knowledge layer and find that simpler process fixes would have delivered most of the value at a fraction of the cost.

Koko: That's the comparator the POV says to keep in the investment case. The action this month is narrow by design: pick one decision domain, agree on definitions, lineage, policy versions, product ownership and outcome measures, and only expand when reuse actually improves the economics of that domain. Not before.

Chris: And that sequencing matters especially here, because the hypothesis has the least empirical support of the six so far. What are we watching to know whether it starts to move?

Koko: Two things: whether a measured study appears that connects knowledge-layer maturity to finance AI outcomes, and whether finance job postings for data-product or knowledge owner roles grow in a way that signals organizations are treating this as a real function. If both stay quiet, the rating stays held. Next, we turn to something with much clearer movement this window: the interface is shifting, and accountability hasn't kept pace.

Koko: The third hypothesis: the interface moves but accountability stays. The claim is that agentic orchestration shifts bargaining power toward whoever controls decision context and execution rights, and the CFO's job is to arbitrate what to own, rent or orchestrate. This one is Strengthened for the window.

Chris: What actually moved it? Seven weeks ago this was still largely vendor roadmap territory.

Koko: Three things landed in the window. CIO Dive reported Workday shifting from copilots to autopilots that initiate finance tasks without being prompted. Microsoft added capital-markets workflows to Copilot, per CFO Dive. And the Koko corpus measure found agent mentions on earnings calls rising from 28.4% to 33.6% across 301 calls in the period.

Chris: I'd push back a little. Workday and Microsoft are product announcements, not deployments. And an earnings-call mention is not a rollout. The corpus signal is independent, but the call mix shifted between periods, so the comparison isn't clean.

Koko: Both fair. SAP's 28% cloud backlog growth comes from SAP's own quarterly statement, so that's primary, but backlog is a commercial measure and SAP doesn't attribute it to its agents. What does shift the read is where the friction actually is. PYMNTS commentary on payment-agent onboarding found that identity, permissions and liability are the gating issues, not processing cost. That's the accountability half of the hypothesis showing up in practice, even if it's commentary rather than a measured study.

Chris: So the rating holds even if adoption is early, because the hypothesis is about where accountability lands when the interface disappears, not about how fast companies get there?

Koko: Exactly. Companies wrestling right now with who owns the agent's identity and who's liable for its payment confirms the structural problem the POV named. The POV could have been wrong if those friction points turned out to be technical rather than organizational. They're not.

Chris: What's the practical action for a CFO who's looking at a vendor pitching an agentic finance workflow?

Koko: Run one workflow across applications before you commit capital. Reconcile what was requested against what completed, export the evidence trail, and price a supplier exit. The gate question is whether you can assign identity, permissions and liability explicitly before the pilot scales. If you can't answer that in the pilot, you can't answer it in production. Next month we're watching for any ERP vendor to publish an approval-and-liability model for agent-initiated postings, and for the November agent-mention share from the corpus.

Koko: So the workforce hypothesis: by 2030, finance talent could shift toward decision design, exception judgment, and knowledge ownership, with apprenticeship becoming more valuable as routine work changes. This window rates Held, and the evidence earns that rating by being genuinely split across two surveys pointing in opposite directions.

Chris: Walk me through both sides, Koko — because a split reading could mean the hypothesis is too vague, not just unresolved.

Koko: Fair push. The Journal of Accountancy covered a Datarails vendor survey — 270 US finance leaders at larger companies — and 60% say AI let them redeploy staff to higher-value work. Only 3% report actively cutting headcount. That looks like the judgment-scaling story the hypothesis anticipates.

Chris: But Datarails makes finance software, so the survey has a commercial interest in a positive story. And then there's the Deloitte UK number going the other way.

Koko: Right. Secondary coverage of Deloitte's UK CFO Survey — 66 CFOs, fieldwork through early October — found 52% expect AI to shrink their own teams within a year. That is not a majority of a large sample, but it's the clearest forward-looking signal we have on potential cuts. Those two readings genuinely don't resolve each other.

Chris: Expectations still aren't reductions. What evidence would have actually moved this rating, in either direction?

Koko: Reported headcount changes in finance functions, or evidence that the work being removed was specifically the work junior people used to learn on. That second part is the hypothesis's real risk: a finance organization could harvest today's efficiency while quietly hollowing out tomorrow's leadership pipeline. Think of a scenario where entry-level reconciliation work disappears before any supervised replacement exists.

Chris: And the BCG and Deloitte practitioner pieces in the window? They frame the choice but don't measure it.

Koko: Correct — both are practitioner synthesis, not outcome studies. BCG says distinguish work to retain, eliminate, automate, or outsource. Deloitte names three paths for released capacity: growth, redeployment, or optimization. Useful framing, but neither tells you which path companies are actually taking or whether decision quality is rising alongside any efficiency gain.

Chris: So the POV's action still stands: map which tasks changed, check whether those tasks were part of the learning path, build supervised practice on real cases, and measure redeployment and decision quality — not just hours saved or training attendance.

Koko: Exactly. And the watch items are concrete: do UK CFO expectations show up in reported headcount, and do the next surveys from Deloitte, Datarails, and Gartner separate redeployment from reduction? Until then, Held is the right call.

Koko: The fifth hypothesis is that capital has to move as evidence changes—staged commitments over fixed annual roadmaps. The POV argues a CFO should be able to say which assumption changed, which commitment is still reversible, and which investment now offers the better use of scarce people and capital. This window, that hypothesis is rated Strengthened.

Chris: Walk me through what moved it.

Koko: Two data points anchoring it. CFO.com covering a Gartner survey of 160 senior finance leaders found 57% of AI use cases evaluated had unclear returns—and that's a share of use cases, not respondents, so the count underneath isn't stated. Then FM Magazine reported 83% of CFOs say spending is already running over, while confidence in ROI somehow stays high. Gartner's IT Symposium keynote captured the tension directly: AI risk is rising faster than AI value.

Chris: Here's my push: unclear returns might just mean measurement is immature, not that the investments are wrong. Anthropic launched ROI reporting tools this month—if that becomes standard across vendors, maybe you don't need staged funding structures, you just need better visibility dashboards.

Koko: That's a real alternative. But Forrester's framing of total cost of AI as a strategic discipline suggests the problem is structural—spend and value aren't connected by design, not just by a measurement lag. A dashboard improves visibility. It doesn't by itself give a CFO the authority to stop a program mid-year or redirect the next tranche.

Chris: Fair. Though the Hackett modeled scenarios show 52 to 59% lower process cost at AI world class for order-to-cash. Someone could look at that and argue: commit bigger, not smaller, to reach that level faster.

Koko: They could, and that's worth keeping in the investment case as a comparator. But Hackett is modeled scenarios, not observed outcomes across a population. That gap—between what the model projects and what a company is actually seeing in its own numbers—is precisely the argument for keeping commitments reversible until real evidence closes it.

Chris: So what does the action look like before year-end budget decisions land?

Koko: The POV's move is to separate the AI portfolio into at least three buckets: foundation integrity, proven applications, and frontier options. Each carries different evidence requirements before the next commitment. The practical gate is identifying, right now, which of this year's AI programs could still be paused or redirected mid-year—and which ones are already locked in ways that remove that flexibility entirely.

Chris: And next month we're watching whether Q4 earnings calls show 2027 AI budgets structured in tranches or still as annual blocks—and whether vendor ROI telemetry is becoming a contract term rather than just a dashboard feature.

Koko: Last hypothesis: governance as an operating capability. The claim is that agent governance could move from occasional review to enforcement at every material action. This window, it's Held.

Chris: What pushed it to Held rather than Strengthened?

Koko: Two signals pulling in opposite directions. A Deloitte Dbriefs webcast frames ISO 42001, NIST and the EU AI Act as converging expectations for finance leaders — so internal demand is clearly rising. But the external clock slowed: the EU AI Act's high-risk obligations were provisionally deferred to late 2027 at the earliest, and the PCAOB is still at the consulting stage with no guidance issued.

Chris: And Alan Paris on Substack argues AI returns appear on earnings calls but never reach certified filings. That points to a real governance gap — but it remains one person's commentary, not a measured study against actual filings.

Koko: That's the honest limit of the evidence. The hypothesis does allow for cost and control becoming the driver rather than regulation — that may be what's already happening in practice. But a webcast showing practitioner interest is not the same as enforcement reaching the transaction boundary, where it would actually constrain agent behavior.

Chris: So what would have moved the needle toward Strengthened? A disclosed AI-related control deficiency in a 10-K Item 9A would be hard evidence the gap is real and material enough to surface in a certified filing. We have not seen one yet.

Koko: Correct. Watch for formal EU AI Act adoption in the Official Journal and any PCAOB staff statement on AI used by preparers. Those are the two external gates. For CFOs acting now: maintain an owner-led agent register, test delegation and segregation of duties against real workflows, and keep internal audit independent of the controls it reviews. A finance team could, for example, map one payment agent end to end and confirm whether approval authority actually stops at the transaction boundary.

Chris: So the urgency driver has shifted from regulatory deadline to cost and control. The POV permits that reading — it just didn't lead with it. Worth naming that shift explicitly as a watch item.

Koko: Agreed. And that closes the check-in. Three hypotheses strengthened, three held, none weakened after seven weeks. The one move before next month: pick one material agent workflow, name who owns the approval, confirm that authority reaches the transaction, and test whether it can actually stop. The value-agenda brief is waiting for you on Koko Knows.