First Steps: Folding-In AI Without Breaking What Works - CFO Secrets Mailbag
Why this is in the vault
Mailbag-format issue that runs the AI-for-CFOs II argument (deterministic-vs-probabilistic decomposition) into a different question - how do you bolt AI onto an existing finance function without destabilizing the systems that already work? The answer is incremental fold-in, not greenfield rebuild. The piece also includes a side-essay rejecting the Ben-Horowitz "wartime CEO / peacetime CEO" dichotomy, arguing effective leaders must operate the full cycle. RDCO files this for two reasons: (1) the incremental-fold-in posture maps directly onto how RDCO is unhobbling its own COO agent (ratchet rules, don't replace the whole stack), and (2) the wartime/peacetime critique is portable Sanity Check material - it's an anti-conventional-wisdom move with a clean owned argument.
The argument
Two arguments stitched together:
On AI adoption. Don't blow up your existing finance systems to fold AI in. Pick the workflows where AI augments today's tooling, install governance at the edge, and let the high-trust systems stay high-trust. The real risk is destabilizing the close, not under-adopting AI. (This is the operational mirror of the AI-for-CFOs II "decompose into atomic units" frame.)
On the wartime/peacetime trope. The framing is overused and lets leaders specialize in only one mode, when the operating reality is that businesses cycle through both phases continuously. Effective CFOs (and CEOs) need to switch modes inside a single quarter, not pick a lane. Specialization in one mode is a liability, not a strength.
Operational specifics
- Fora claim: scaled 3x in one year while keeping finance team flat, cited from the embedded Campfire + Ramp webinar.
- Monte Carlo simulation referenced for covenant stress testing - one of the few places he names a probabilistic technique as the right tool.
- "Austin's case" of owner distributions continuing during flat / loss-making months - illustration of a governance failure that fold-in AI would not catch but a controller would.
- Oracle hired an industrial CFO to manage "AI build out" - flagged in the Noteworthy section as a category signal.
- Physical inventory audits flagged as the durable AI-resistant accounting task - controls require human eyes on physical assets.
Mapping against Ray Data Co
Against the harness-engineering thesis cluster
- Incremental fold-in over greenfield. This is exactly how RDCO is unhobbling the COO agent. We do not greenfield-rebuild the harness; we ratchet rules and add hooks ([[06-reference/2026-04-08-better-harness-evals-hill-climbing.md]]). The CFO advice is the buyer-side mirror: don't burn down your close to fold in AI; bolt AI onto the workflows where it augments the existing close.
- Audit gates remain audit gates. The "physical inventory audits stay human" line is the buyer-side analog of audit-newsletter-outputs.py - certain gates must remain deterministic-or-physical regardless of how much AI you fold in.
- Reject the binary. The wartime/peacetime critique is structurally adjacent to the harness-engineering rejection of "agentic vs deterministic" as a binary. Both arguments say: stop picking a lane, master the full cycle, and route work to the right mode case-by-case.
Against MAC
The "don't destabilize what works" posture is exactly what MAC enables operationally. MAC's Stop / Pause / Go severity tiers are the mechanism by which you fold-in AI without destabilizing the close: AI failures route to severity-tier triage instead of cascading into book-of-record corruption. Worth a Sanity Check candidate angle: "The Secret CFO says fold-in carefully. Here's the data-quality layer that lets you."
Against Sanity Check
The wartime/peacetime takedown is a portable contrarian-but-not-dismissive move - the same balance the SC voice is reaching for. Worth studying as a structural pattern: take a popular framework, name what it gets right (the diagnosis of leadership-mode mismatch), name what it gets wrong (encouraging specialization), end with a sharper alternative (master the cycle). That structure works for SC topics like "the AI agents vs. workflows debate" or "the SaaS vs services debate."
Founder leadership note
The wartime/peacetime critique is also worth a private read for the founder. RDCO is in an early-stage compounding phase that looks peacetime from the outside but is wartime in target-system development - founder needs to operate both modes inside a single week, not pick one.
Sponsorship
Top-of-newsletter banner is a Campfire + Ramp co-webinar promotion (recurring Campfire relationship; Ramp returning as co-sponsor, having previously appeared in AI-for-CFOs III). Mid-content carries an ERP-implementation survey link. Footer carries the SimCFO simulation tool with a leaderboard. Sponsor placement is adjacent rather than embedded; editorial spine reads independent. The Campfire / Ramp pairing is the pattern to watch - they appear to be co-sponsoring the AI-adjacent issues, which is a meaningful signal about which vendors see CFO Secrets readership as their ICP.
Related
- [[06-reference/2026-05-11-cfo-secrets-ai-for-cfos-series-synthesis]] - the AI-for-CFOs synthesis; the deterministic-vs-probabilistic frame this Mailbag operationalizes
- [[06-reference/2026-04-28-cfosecrets-finance-stack-of-the-future-unbundled-erp]] - companion Mailbag on the unbundled-ERP argument; same sponsor pattern
- [[06-reference/concepts/2026-05-10-harness-moat-two-layers-portability]] - harness-moat thesis; incremental fold-in is the buyer-side mirror
- [[06-reference/2026-04-30-mac-bet-architecture-audit]] - MAC bet architecture; severity-tier triage is the data-quality fold-in mechanism
- [[06-reference/2026-04-08-better-harness-evals-hill-climbing.md]] - hill-climbing harness evals; the build-side analog of incremental fold-in