How Do We Measure How Blind We Are?
CJ Gustafson uses quotes from finance-tech CEOs (collected for his 2026 agentic finance market-map report) to argue that companies mandate AI adoption but cannot see what it does, and that the missing layer is event-level records and action-level observability.
Why this is in the vault
A compact statement of the "accuracy is not confidence" framing for agents in finance, which pairs directly with how we think about verifying agent output.
Mapping against Ray Data Co
The most concrete connection is the fresh-eyes verification gate ([[feedback_fresh_eyes_subagent_for_own_artifacts]]) and the agent-output-integrity work: CJ's accuracy-by-confidence quadrant (worst box is confident but inaccurate) is the same failure we gate against when an agent reports its own work as done. His claim that records need lineage back to the causing action ("a debit and credit is always the effect, never the cause", quoting Numeric's CEO) supports our deterministic audit layers and ops logging over self-reported status. It also gives phData sales credibility material: "no benchmark exists for internal AI spend ROI; product-cost AI will show ROI before employee/OpEx AI" and his predicted 2027 churn wave are usable framing for AI Services conversations. Weak spots: the essay is quote-driven and unquantified, and every quoted CEO is a report partner with a stake.
The core argument
Every CFO has an AI mandate, but teams see licenses and credit bills, not outcomes. Without event-level record keeping and action-level observability, financial records cannot be traced to their causes. Accuracy and confidence differ; autonomy needs both. Agents may leave you less blind than people, since an agent's reasoning stays with the company while a departing employee's does not. Management structures (reporting lines, budgets, reviews) exist for people and not yet for agents. No ROI benchmarks exist for internal AI spend, and he expects a 2027 churn wave as cheap-to-adopt tools are cheap to rip out and vendor discounts fade. The maps show what a product can do, not how much autonomy to turn on; the right setting depends on the cost of a mistake. AI quality is a gradient, like water that may or may not be clean.
⚠️ Sponsorship
Brex powers the masthead (paid sponsor) and is tied to an upcoming cash-management series. The report being promoted (download via a lead-capture form) is built with partners Abacum, Airwallex, BILL, cfo.ai, NetSuite, Numeric, Pigment, Proxon, Tropic and Vic.ai, and the essay quotes several of those same companies' CEOs (Abacum, Numeric, Pigment, Proxon, Vic.ai), so the sourcing is partner-selected, not independent. Conor McGregor gag and kids-need-shoes line are tongue-in-cheek self-promo. Commercial finance-media operator; treat conclusions as informed opinion.
Related
- [[2026-11-12-mostlymetrics-procurement-ai-spend-market-maps]]
- [[2026-11-05-mostlymetrics-agentic-finance-stack-2026-preview]]
- [[2026-11-02-data-engineering-weekly-294-agent-governance-controls]]
- [[2026-07-19-every-model-easy-part-measurement]]
- [[feedback_fresh_eyes_subagent_for_own_artifacts]]