Connectors Not Models: How Integration Wins CFO Buy‑In

CFOs and ops teams buy decision‑ready artifacts: board packs, variance bridges, scenario models. These artifacts must link every material number to a workbook tab, dashboard, or named owner. OpenAI’s finance playbook enforces source citations and conservative editing rules: don’t invent metrics, flag unstated assumptions, and cite the workbook tab behind each claim. That is product specification, not a model capability. (openai.com) A raw LLM can produce a plausible narrative that does not match the source files. Finance treats that as noise. A system that surfaces the exact Excel tab and attaches a QA memo converts the same language generation into an actionable artifact. Building that system requires file connectors, a mapping from enterprise schemas to prompts, deterministic citation logic, and a UI that separates sourced facts from interpretation. The model writes; the rest makes the writing trustworthy.

Three properties recur in every successful operational deployment: permissions, provenance, and templates. Permissions limit reads to what users already access. Provenance ties each sentence and number to a concrete source. Templates encode CFO‑grade structure and tone. OpenAI’s how‑to pages treat these as non‑optional: use Google Drive, SharePoint, Box and require citations for every material number. Those are guardrails, not optional features. (openai.com) Implementing this stack requires engineering and policy work: deploy connectors that respect org auth, instrument audit logs for every extraction, generate deterministic citation metadata, and map domain objects (P&L lines, headcount plans) to prose blocks. Each piece raises the cost to replicate and turns a one‑off productivity trick into part of the finance operating rhythm: draft creation, owner verification, auditor trail, leader sign‑off.

Adoption follows a two‑step internal funnel. First, an analyst or ops lead integrates the assistant into a workflow using templates, connectors, and prompts. Second, the finance leader adopts the output because it reduces review time and improves auditability. The buyer is the accountable person. For CFOs that means clarity on sources, change‑logs, and owner follow‑ups. OpenAI’s business and data science guides assume teams will provide trackers, dashboards, meeting notes and owner context, and that Codex’s role is to assemble a reviewable first pass. (openai.com) This distribution pattern favors companies that package domain‑specific templates and onboarding playbooks. The engineering emphasis shifts from training custom models to building connectors, prewired templates, role‑based permissioning, and a governance UI. The seller becomes the template author and the internal champion who trains owners to verify drafts — not the ML team that swapped checkpoints.

Tying AI drafts to a firm’s canonical files, owners, and decision flows creates real switching costs. Microsoft’s Copilot illustrates this: embedding the assistant in OneDrive, SharePoint, and Teams makes content and permissioning native to the assistant. Enterprises value that continuity because it reduces exposure, simplifies compliance, and eases audits. (learn.microsoft.com) Model quality sets baseline competence. Durable enterprise advantages arise from product engineering that solves connectors, governance, and UX. If you build for finance and ops, prioritize auditable connectors to systems of record, deterministic provenance and citation rules, and a library of vetted templates and prompts for CFO workflows. Do this well and you stop competing on model logits and start trading on process, trust, and operational lock‑in — the assets that survive commoditization of base models.