Six to nine months ago, CFOs at roundtables had to be coaxed into talking about AI. By the CFO Alliance roundtables in Texas in early November 2026, it was the subject everyone wanted. The questions had also moved on. They were no longer about what the tools could do but about return on the investment, risk, and why so little of the early promise had reached the way finance teams actually work.
A simple model came up in every room to describe where teams are with AI adoption in finance: curious, connected and compounding.
The three stages
Curious. People are using AI tools on their own. Drafting emails, summarising contracts, writing a formula or a first-pass variance comment. There are real time savings for individuals, and no change to how the finance function runs.
Connected. At least one process has been changed. AI is connected to the systems it works on, through an integration or a protocol such as MCP, rather than fed by copy and paste. Typical first candidates are invoice coding, reconciliation matching, close checklists and the first draft of monthly commentary.
Compounding. A function has been redesigned around AI. The close, accounts payable or reporting runs differently end to end, and each improvement makes the next one easier.
At a recent finance leadership event of around 200 people, a show of hands put most of the room at curious, a few at connected and very few at compounding. That matched what the Texas rooms described.
Why AI adoption in finance stalls at curious
Building has become the easy part. Several CFOs made the same point: the tools work, and a capable analyst can prototype a useful workflow in a day. What fails is adoption. ERP implementations have failed this way for years, and AI is following the same pattern. The system is built, people keep working the old way, and the savings never reach the process.
Individual time savings also have a habit of disappearing. One CFO described a task that went from ten hours to three. The team did not reinvest the seven hours; it absorbed them. Another firm tied part of its incentive pay to how AI time savings were redeployed, so that the hours went into new analysis rather than quietly back into the week.
The order that works
Start with people. Decide what the team needs, what will change for each role, and what they are worried about. Process comes before tooling: agree how the work should run, get buy-in, then implement.
Find the champions. Most teams have one or two people already working at the connected or compounding stage on their own. Give them a process to own and let others see the results.
Set guardrails before you spread. Write down which data AI tools can see, which outputs need human review, and who signs off. An AI use policy does not need to be long, but it needs to exist before adoption widens.
Then spread. Move to the next process once the first is working and trusted, not before.
Control access before agents touch the ledger
One CFO at the Austin roundtable described using an AI assistant to help close the books, drafting journal entries with direct query access to the accounting data. The productivity was real. So was the point raised immediately afterwards: before any AI agent gets direct database access, the data layer needs proper authentication, permissions scoped to what the agent needs, and a log of what it did.
Read-only access for analysis is a different risk from write access that posts entries. Treat them differently, and keep a person accountable for anything that reaches the ledger.
Buy versus build has shifted
Platforms that were an obvious purchase six months ago may now be worth building, because building is cheaper and faster than it was. The reverse also holds: a quick internal build can turn into software nobody maintains. Decide on the basis of who will own it in two years, not how fast it can be made.
What AI adoption will not fix
It will not fix a broken process. Automating a close that has no calendar, no reconciliation standard and no review step produces faster errors.
It will not remove the need for judgment. AI output still needs someone with the domain knowledge to say when it is wrong, which we take up in AI in finance has a confidence problem.
Frequently asked questions
What are the stages of AI adoption in finance teams?
A common model has three: curious, where people use AI tools individually; connected, where at least one finance process has been changed with AI integrated into the systems; and compounding, where a whole function has been redesigned around AI.
Where are most finance teams with AI today?
Most are at the curious stage, using tools individually without changing processes. In rooms of finance leaders in late 2026, few described themselves as connected and very few as compounding.
Why do AI projects in finance fail?
Mostly at adoption rather than build. The tool works, but people keep the old process, time savings are absorbed rather than reinvested, and nobody owns the change. Starting with people and process before tooling avoids most of this.
Should AI agents have direct access to accounting data?
Only with controls in place first: authentication at the data layer, permissions limited to what the agent needs, a log of its actions, and human review of anything posted to the ledger. Read-only analysis and write access are different risks.
How should a CFO start an AI rollout in finance?
Pick one process, agree how it should work, find the people already getting results, write a short AI use policy, and move to the next process only when the first is working and trusted.
Where CFOLogic fits
We run finance operations for growing businesses and use AI inside the close and reconciliation work, with a finance professional reviewing and signing off what it produces. See Finance Team as a Service.