This is part 2 of Finance AI Field Notes, a series grounded in a July 2026 scan of ~105 live job postings plus verified market research. Part 1: The finance systems role is quietly becoming an engineering job.
If you want to know how finance leaders plan to adopt AI, you can ask them — or you can read what they're hiring for. The two answers don't quite match, and the gap is the most useful thing I've learned this year.
What the surveys say
Start with the self-reported picture. Bain's 2026 CFO survey found roughly 60% of finance organizations still have AI in pilot or limited production; only 15–25% have scaled anything into full production across finance. At the same time, budgets are moving: 56% of CFOs are raising AI spend by 15% or more this year. Money in, production out — the classic shape of a function that believes but hasn't yet operationalized.
On the how, L.E.K.'s Office-of-the-CFO survey is the cleanest data point I've found: about 56% of CFOs prefer platform solutions with embedded AI from vendors they already use, versus roughly 31% who prefer best-in-class stand-alone tools. The stated reasons are ease of adoption, data consistency, and vendor rationalization. In other words: "I'd rather my ERP grew agents than buy an agent product."
Gartner, meanwhile, predicts finance organizations using cloud ERP with embedded AI will close 30% faster by 2028 — but the same press release names why most CFOs are stuck in early adoption: "data quality, integration complexity, skills gaps, and inconsistent multientity support in the market." Keep that list. It's the honest one.
What the vendors' numbers say
If 56% of CFOs want embedded AI from their existing vendor, the existing vendors should be winning. Are they? The German-speaking SAP user group (DSAG) surveyed its members: about 3% of SAP customers run SAP Business AI in production, and 77% of the SAP enterprises active in AI rely on non-SAP tools to do it. SAP's own chief product officer conceded that adoption of Joule Studio — the build-your-own-agents tool — "has been minimal."
So the stated preference is embedded; the deployed reality is pilots plus workarounds. Announcement velocity is not adoption. That gap between what the incumbent demos and what the incumbent's customers actually run is where every other approach lives.
What the hiring pages say
Here's the signal the surveys don't capture. In July 2026 I collected about 105 live job postings where companies hire people to build or own AI inside the finance function. Not data-science roles — finance roles: Heads of Finance Systems asked to "replace manual finance workflows with AI/LLM-powered agents" (Scale AI), a Director of Accounting whose JD states "software agents, automated reconciliations, and compressed close cycles are the mandate, not an aspiration" (CSC Generation), a crypto exchange hiring an architect for "an agentic finance layer that runs alongside the existing stack today" (Kraken), a 94,000-employee insurer creating a "Finance Agents Engineering" unit that "owns the agent layer of the close" (Aon).
Roughly 70% of the corporate postings use build language, not buy language. That's a third adoption path the surveys under-count, because "we hired someone to build it" doesn't fit neatly into platform-vs-point-solution questionnaires.
The three real paths (and who takes them)
Reading surveys and postings together, finance leaders are sorting into three lanes:
1. Wait for embedded. Trust the ERP vendor's roadmap. Rational for the risk-averse; the DSAG numbers say you'll wait longer than the keynote implied. This is the majority position by preference and the slowest by delivery.
2. Buy point solutions. Close tools, AP automation, reconciliation SaaS. Fast for the specific workflow; adds another silo to the stack Gartner already flagged for "integration complexity."
3. Hire a builder. One or two people, inside finance, wiring LLM workflows to the ERP. This is the fastest-growing lane in the hiring data and the least examined. Its strength: agents shaped exactly to your process. Its weakness: everything — governance, audit evidence, upgrade path, key-person risk — depends on what that one hire builds under time pressure.
What I'd actually advise
Whichever lane you're in, three tests keep you honest. First, the auditor test: for any AI-touched workflow, can you show who approved what, when, with what evidence? If the answer depends on screenshots, you're not in production, you're in a demo. Second, the multientity test — Gartner named it for a reason: a pilot that works for one entity and dies at consolidation isn't a pilot, it's a dead end. Third, the bus-factor test: if your builder resigns, does the automation survive the quarter?
The leaders getting this right aren't choosing a lane on principle. They're choosing per-workflow — embedded where the vendor is genuinely there, bought where the point tool is mature, built where the process is theirs alone — and holding all three to the same control standard.
This series draws on a July 2026 scan of ~105 live finance-AI job postings plus the verified survey data cited above. I build Artifi, a governed platform for exactly the third lane. Next: the goals CFOs are actually setting for AI — taken from the success metrics written into these job descriptions.