Agentic AI in finance, and the 17% that matters
You have been shown the adoption chart by now. Some version of it: a bar that was small in 2023 and is large in 2026, under a headline about how finance is embracing AI agents.
The number underneath is more interesting. According to CFO Connect’s State of AI in Finance 2026, published in March, 56% of finance leaders use AI, roughly double 2023. But 45% of teams are in limited pilot mode and only 17% have AI in core workflows.
So five in six finance teams that report using AI are not running it on anything that matters yet. That gap is the whole story, and it is the one worth planning against.
This is the strategic read. If you want the tool-by-tool version, AI agents for finance is the hub and it covers what to install. Checked 29 August 2026.
What agentic means, briefly
An agent takes a goal rather than an instruction. You describe the outcome, it decides the steps, calls the tools it needs, and comes back with the work done. A script follows the steps you already wrote. A chatbot answers the question you already asked.
The distinction matters for finance because an agent makes decisions you did not explicitly approve. That is exactly the property that makes it useful, and exactly the property your controls were not designed for.
What the 2026 numbers say, and how much to trust them
Two surveys get quoted constantly. They say different things, because they asked different people, and that difference is worth understanding before you cite either in a board paper.
CFO Connect, State of AI in Finance 2026
Published 11 March 2026. Headline findings: 56% of finance leaders use AI, double 2023. 45% remain in limited pilot mode against 17% in core workflows, a split it attributes to a General Atlantic poll. And 68% of CFOs say they have been slow to adopt because they do not know where to start.
It also makes a point most coverage skips: finance still ranks last among all business functions on AI deployment. Not second-slowest. Last.
The caveat: the report does not disclose a sample size. That does not make it wrong, but a percentage without an n is a direction, not a measurement, and it should be quoted that way.
Consero, 2026 CFO Survey
Run with Cascade Insights in Q1 2026, and this one does publish its method: 102 finance leaders, 71% CFOs and 29% VPs of Finance, all at private equity or venture-backed companies. Fully embedded AI in finance rose from 22% to 42% year on year, and 97% now use or test AI in some form.
The caveat: PE and VC-backed finance teams are not a cross-section of finance. They are smaller, faster-moving, under investor pressure to show operating leverage, and generally less constrained by legacy systems. A 97% figure from that population is real and it is not your company unless it is.
What both agree on
| Finding | Source |
|---|---|
| Adoption is rising fast and genuinely | Both |
| Depth is far behind breadth | CFO Connect, 17% in core workflows |
| Not knowing where to start is the top blocker | CFO Connect, 68% of CFOs |
| AI is adding headcount, not cutting it | Consero, 87% hiring more finance staff |
That last row is the one I would put in front of a nervous team. In the Consero population, 87% are hiring more finance people, not fewer.
For the 68% who do not know where to start
The finance AI library
The prompts, templates and the fill-in AI policy I use on real finance work, so the first step is a starting point rather than a blank page. Free.
Where agents are genuinely in production
The pattern in what has shipped is consistent, and it is not the glamorous end of finance.
Invoice processing is the most common entry point. HPE has put an agent called Alfred into production on the transactional side of its finance function, covering credit and collections through accounts payable and receivable. Not forecasting. Not strategy. The high-volume, rule-heavy, deeply boring work.
That is the right place to start and the reason is structural. Transactional finance has the three things an agent needs: enough volume that automation pays back, clear enough rules that the agent can be judged right or wrong, and a reviewer who would catch an error before it reached a statement.
Forecasting has none of those. The volume is low, the rules are judgement, and no one can tell you a forecast was wrong until the quarter closes. It is the last place to point an agent, and it is the first place most people want to.
Why finance is last, and why that is defensible
Finance ranking last on deployment gets framed as a failure of nerve. I think that reading is wrong, and the reason sits in what this month’s tooling research turned up.
Across the four major AI surfaces a finance team would reach for, the audit story is inconsistent. Some log everything, some log nothing, and the differences are not advertised. If your job includes signing something, adopting a tool whose activity your compliance team cannot see is not caution, it is the correct answer.
The slowness is also self-inflicted in one specific way. The most cited blocker is that 68% of CFOs do not know where to start, ahead of budget, risk and skills. That is a problem you solve with a decision rather than a project.
What I would do about it
Four things, in this order, and none of them require a strategy document.
- Pick the boring process. High volume, clear rules, a reviewer who would catch a mistake. That is where the production deployments are, and it is where you can prove the thing works before anyone important is watching.
- Settle the audit question first, not last. Ask what your compliance team can see before you pick the tool, because the answer varies more than you would expect and it eliminates options quickly.
- Write the policy before the pilot. Approval modes, what data the tool may touch, who reviews the output. Two pages, not twenty. The template in the library covers it.
- Keep the human review step and mean it. Every production example I can find has one. The agent removes the typing, not the accountability.
The 17% figure is the opportunity, not the warning. Most finance teams saying they use AI are running it on the edges. Getting one real process onto it puts you in a small group quickly.
Where to go next
AI agents for finance is the practical hub: what counts as an agent, which tools are worth learning, and where to start on Monday. If you want to see one built end to end, building AI agents in n8n walks through three working finance agents.
For the tools themselves, Claude Cowork and Copilot agents are the two most finance teams will meet first, and both write-ups cover what an admin can and cannot see. If the underlying problem is a month that repeats itself, start at finance automation instead.
The distinction between 56% using AI and 45% still limited to pilots is far more useful than the headline adoption chart—scaling agentic workflows requires measurable controls, not just experimentation.