To understand what they want instead, we interviewed and surveyed 50+ finance leaders we highly rate (42 CFOs and other senior finance leaders, such as Heads of Finance and Financial Controllers). We asked not only about AI that answers questions or drafts reports, but also about AI that does the work: reconciling accounts, resolving discrepancies, chasing payments and, eventually, running entire finance workflows. For simplicity, we refer to all respondents as CFOs throughout.

The picture that emerged is far more interesting than a simple story of cautious CFOs resisting automation. Below, we explore individual finance workflows (accounts receivable, accounts payable, FP&A and so on) to uncover everything from why accounts receivable lags payables despite similar comfort with AI taking action, to why close and reporting automation tops CFOs' wish lists.

This is the first of a three-part series. Having explored what CFOs want, we turn to how to put AI to work in Part 2, a practical deployment playbook for finance leaders, before mapping who's building the solutions in Part 3.

55% of respondents already have AI in production (i.e. AI embedded in finance workflows, not merely individual chatbot usage), with another 25% piloting. And general-purpose AI tools like Claude and ChatGPT are almost ubiquitous: 98% of respondents use them (largely driven by enterprise-wide rollouts) compared with:

• 34% using AI capabilities in existing ERP systems (e.g. SAP, Oracle, NetSuite, Workday, Quickbooks, Xero)
• 23% using AI in existing finance software
• And just 6% using standalone AI tools from startups

When we asked an open-ended question about their favourite AI tools, 83% mentioned Claude. This likely reflects which tools companies have rolled out enterprise-wide as much as any head-to-head preference, and our sample may skew towards Claude users.

We aren't disheartened by the small share of CFOs buying standalone AI tools from startups today. We think the market could evolve in three stages:

We asked CFOs two separate questions across 12 finance workflows. First, are they currently using AI, planning to deploy it within the next 12 months, or have no plans? Second, how much autonomy would they be comfortable giving AI in that same workflow, from "not comfortable with AI" through "AI recommends, human decides" and "AI acts with human approval" to "AI acts autonomously within rules"?

That lets us gauge two things: the adoption frontier, where finance teams actually use AI today, and the autonomy frontier, where CFOs believe AI could safely be allowed to act.

The most fascinating findings arose where the two sets of answers meet. "No plans" doesn't automatically mean "no trust". Across the full sample, 202 workflow responses were marked "no plans" for the next 12 months. Yet 84% of those are still comfortable with AI playing some role, and 51% would let it actually act, autonomously or with approval.

This suggests an enormous unrealised delegation pool: willingness is substantially ahead of deployment. Where CFOs had no plans for a workflow, our interviews suggest it usually reflected business priorities rather than reluctance towards agentic AI. Existing systems work well enough, budget is committed elsewhere, the team lacks capacity to implement, or no compelling product exists yet. None of those explanations means they distrust AI.

Being comfortable in principle doesn't establish purchasing intent. Some respondents may picture low-risk actions such as flagging suspicious transactions rather than consequential ones such as freezing payments. Nevertheless, deployment priorities and appetite for autonomy appear to be two different dimensions.

To see where this latent demand sits, we looked only at respondents who aren't yet using AI in a given workflow. A majority would still let AI take action in eight of the 12 workflows, led by accounts receivable (79%), expenses and procurement (77%) and audit and controls (74%).

Appetite for autonomy varies across workflows. Some involve repeatable transactions, clear rules and verifiable outcomes (AP, AR, expense management, fraud detection). Others involve complex judgement, uncertain forecasts or decisions with no single correct answer (FP&A). It's not only an unwillingness to delegate high-stakes work such as tax, compliance and treasury, but also a reluctance to delegate ambiguous work.

When we asked respondents to rank the challenges they face deploying AI, data quality and data silos came first overall, followed by accuracy and trustworthiness, then integration. There's nothing terribly shocking there.

A more interesting finding emerges when we separate teams already using AI in production from those that haven't deployed anything live yet. Of the 24 respondents not yet in production, nine (38%) said lack of AI skills was their biggest challenge. Among the 29 already in production, only one (3%) did. Those further along worry more about data quality, accuracy and integration.

These are different teams surveyed at a single point in time, so we can't say for certain that concerns shift once AI goes live, and a skills gap may itself be part of why some teams haven't reached production. But the contrast is suggestive.

The divide between Tech/SaaS and other industries points to a similar pattern. 69% of Tech/SaaS companies (n = 26) already have AI in production, compared with 41% of other companies (n = 27), and deploy it across 4.3 workflows on average versus 2.9. Yet non-Tech/SaaS companies (financial services, retail and consumer, professional services and others) are planning deployments across almost twice as many additional workflows, partly because they have more left to automate.

This looks more like a gap in execution than in ambition. Tech-native businesses often have the expertise, engineering capacity and infrastructure to experiment with AI and work around product limitations. Other organisations face the harder task of retrofitting AI into fragmented systems, legacy processes and imperfect data.

And there's a catch-22. One finance leader at a small consumer business described a team so stretched that every hour spent learning and implementing AI is an hour taken from today's work. The teams that stand to benefit most from automation may be the least able to dedicate resources to deploying it. Another observed that the bottleneck wasn't using AI, but getting information into an AI-ready format in the first place.

CFOs are, understandably, not especially fond of financial inaccuracies. 58% ranked accuracy and reliability as their number-one criterion when selecting AI products, with ease of integration second by average rank. But what persuades them to give AI more autonomy? Explainability and complete audit logs ranked first and second, ahead of improvements in accuracy and proven case studies.

Accuracy still received the most individual first-place votes, but the message is clear. CFOs don't just want AI to get things right. They want to understand and verify how it got there.

That's why founders need to focus on graduated autonomy: start with recommendations, progress to human-approved actions, and eventually allow autonomy within defined permissions, policies and thresholds. Every action should be traceable, verifiable and, where possible, reversible.

As underlying AI capabilities become commoditised, we think differentiation will increasingly come from three things.

Together these create a compounding flywheel. Each workflow becomes the foundation for the next, giving a narrow point solution a credible path to becoming a broad finance platform. It also points towards consolidation: CFOs have little reason to manage a sprawl of disconnected agents when one or a few providers can reuse context, integrations and controls across them all.

With this deeper understanding of the customer, startups can drive the kind of ROI that CFOs value most: time savings and freed-up strategic capacity.

CFOs' build/buy behaviour is overwhelmingly case-by-case, or "buy, but build where nothing good enough exists". Overall, only 8% express a primary preference for building.

A more interesting question is what CFOs wouldn't want to build in-house. Our interviews point to payments, complex procurement, regulated workflows (payroll, tax and audit) and foundational accounting/ERP. Three factors govern this: the complexity of the workflow and the high likelihood of getting something wrong, the consequences of getting it wrong, and data security and privacy (payroll is very sensitive).

The gap between preference and practice becomes even more striking by deployment maturity:

Custom orchestration and integration is strongly associated with actually getting AI live. That said, none of the respondents already using an in-house solution say they prefer a build-first strategy. Finance teams are building because the market doesn't yet solve their needs, not because they want to. Conversely, the handful who do prefer to build aren't (yet) using in-house solutions. And given the pace of change in AI, the CFOs we interviewed would rather not take on the burden of maintaining tools themselves.

There's also a sector divide. Tech/SaaS companies are more than twice as likely to have built their own AI tools (46% versus 19%), yet almost none say they prefer building; most weigh it up case by case. They build because they can, not because they want to.

We asked CFOs for the top three skills they most want to add or develop in their finance team as AI becomes more prevalent. Process design and automation architecture is the most-selected future skill at 62%, narrowly ahead of data literacy at 60%. That's a one-respondent margin, so the two are effectively joint top.

The first wave of AI adoption has largely been about making finance professionals more productive. The next will be about redesigning the workflows themselves.

We don't think the winners will necessarily be those with the cleverest models, but those that best understand a company's financial context, navigate its messy exceptions and earn the right to act with progressively less supervision. Choosing the right ICP matters too: traditional businesses are more attractive buyers, as they prefer buying over building and typically lack the engineering capacity to build and maintain in-house.

That gap is where most finance AI projects stall. In Part 1, we surveyed 50+ senior finance leaders and found that appetite is running well ahead of deployment: even for workflows where CFOs have no AI plans for the next 12 months, more than half of responses said they'd let AI take action. What holds teams back isn't willingness. It's execution: messy data, fragmented systems, scarce skills and teams too stretched to experiment.

This playbook draws on what we heard from CFOs and finance leaders who have already put AI live, including teams in our portfolio. Read it end to end, or jump straight to the part you need.

The best first project is generally a bounded process that the team dislikes, understands deeply and can test against a trusted result. Not every workflow needs AI: standard rules, integrations and existing product features often solve simpler problems more cheaply.

Score each candidate process from 1 to 5 on the six dimensions below, and start with the highest total. Treat a low score on Verifiability or Reversibility as a red flag, however painful the process.

In the worked example, the payment run has the highest total, but its Reversibility score rules it out as a first project: an error means money has already gone. The reconciliation scores almost as well, every output can be checked, and nothing it does is irreversible, so it's the place to start. ARR shows that a low score can be fixable: agree one definition across systems (see .03) and its Verifiability rises, which is exactly where the CFO in our introduction came unstuck.

When measuring the return, capture a baseline before you start: hours spent, error rates, cycle times and cost. CFOs told us the ROI they value most is time savings and freed-up strategic capacity, so report both the hours returned and what the team now does with them, whether that's a faster close, live cash visibility or analysis that simply wasn't possible before.

Saving time is only half the story. Don't automatically remove the work people find intellectually useful (focus on manual, repetitive, boring tasks), and protect the judgement, challenge and explanation that make finance strategically valuable.

Once you've chosen a process, resist the urge to open a tool. Map it in detail first: teams we spoke with found that the edge cases they uncovered determined what an agent could safely do. For each step in the process, document the following.

We suggest a one-page swimlane diagram (a flowchart showing each step in sequence and who owns it), plus a decision log, sample inputs and outputs, an exception register and a human approval matrix. Start with common cases, but explicitly test for edge cases and manual overrides. Continuing the worked example, here's a swimlane for supplier statement reconciliation: the agent prepares and stages the work; people resolve exceptions, approve and post.

Many CFOs described the same metric appearing differently across systems such as their data warehouse, ERP and FP&A tools. That matters because automating over inconsistent data can scale error faster than it scales productivity. Here's what CFOs are doing about it.

Some are building permissioned semantic layers on exported NetSuite tables before exposing them to agents, though for many businesses this would mean a large data-warehouse project. Vendors are getting better at handling imperfect data, but clean data pays off whether you build or buy, and it's essential if you build in-house.

CFOs want a more consolidated agentic finance stack, with fewer providers reusing the same context, integrations and controls across workflows. Those already on AI-native ERPs expect them to absorb many of today's point solutions, and CFOs more broadly are drawn to end-to-end offerings. Yet they're wary of single-vendor lock-in, and remain open to specialist tools that integrate seamlessly and deliver a meaningful improvement. The right route depends on the workflow.

As a rule of thumb, the CFOs we interviewed for Part 1 were most reluctant to build payments, complex procurement, regulated workflows (payroll, tax, audit) and core accounting/ERP: areas where errors are costly and data is sensitive. See the build vs buy data in Part 1 →

Finance is a low-tolerance environment for plausible but incorrect outputs. A faster calculation isn't a production-ready source of truth. Yet the same CFO from our ARR story also shows where AI does deliver.

The difference is verifiability. The lesson from both: earn autonomy gradually, proving accuracy at each step before granting more. Use an explicit autonomy ladder.

Build these controls in from day one:

Bring internal and external auditors in early, too. Agreeing upfront how AI-supported controls will be evidenced saves difficult conversations at year end.

None of these controls replaces judgement. They tell you where to look, but someone with enough experience still has to question what they find and push back when a number looks wrong.

One finance team protected every Thursday for automation work during a focused two-month period. Another ran a two-hour hackathon, asking everyone to automate an irritating spreadsheet or small task. Both recognise that the people closest to a workflow usually understand its edge cases best. The most AI-forward finance teams we've seen do the following.

Drawn from our conversations with CFOs and finance leaders, this catalogue shows what peers are actually automating across the finance function, and how far along they are. Every entry is live, in pilot or committed to.

Run through the first list before you start building, and the second before any pilot goes into production. Your ticks are kept in this browser only.

None of this requires an AI-native team on day one. Start with one workflow, test it against a result you already trust, and use what you learn to tackle the next. In Part 3, we'll map the startups building agentic finance products, so you can see who might solve the workflows you'd rather not build yourself.