Agents are on the table in almost every finance leadership conversation right now. An agent that runs the close, chases receivables, or drafts the board narrative overnight. The interest is well placed: agentic AI really can orchestrate the multistep, repeatable work that fills a finance team’s month.
The question worth asking isn’t whether agents deliver value, but what it needs underneath it to be something you’d trust with a number that reaches the board.
A statistic you may have seen, read correctly, explains why.
What the 95% figure really tells finance leaders
MIT’s State of AI in Business 2025 found that roughly 95% of organizations see no measurable return on their generative AI investments. The number has been traveling through boardrooms as proof that AI isn’t ready.
But Bain & Company reads it more usefully. The problem the data exposes is the approach, not the models. The organizations seeing returns are embedding AI into real workflows, grounding it in good context, and measuring business outcomes. The ones seeing nothing pour money into visible, top-line pilots while the back-office work with the fastest payback goes starved. Bain’s message to a CFO is not to spend less on AI but to spend differently, aiming the next dollar at the use cases that move cash, cost, and risk.
That reframing carries real weight in finance. Finance processes are repeatable, data-rich, and policy-bound, which Bain identifies as the very conditions where AI delivers. Far from being the laggard in the AI story, finance is one of the places AI already works, when it’s pointed at the right problem and built on the right foundation.
The agent is the consequence, not the starting point
Here the agent conversation and the 95% conversation turn out to be the same one.
An agent is AI given room to take multistep action: read an invoice, match it to a contract, apply the tolerance, post the entry, escalate only the exception. That is valuable work, and it’s also the kind that fails quietly when the foundation isn’t there. An agent acting on data nobody validated, applying rules it inferred rather than rules finance defined, producing an output no one can trace, doesn’t make finance faster. It reproduces the 95% problem with more autonomy and higher stakes.
Bain locates the real constraint precisely. What holds finance AI back is not the price of the models or the sophistication of the prompts. It’s whether the context is ready: shared definitions for the chart of accounts and master data, ERP and EPM systems connected under a policy-aware retrieval layer, every action traceable to the data and logic behind it. An agent is only ever as trustworthy as the governed context it runs on.
Market enthusiasm for the visible layer of the agent is understandable. Whether it works depends on something underneath: whether the business logic feeding it was defined by finance or inferred by a model. Deterministic logic is defined by finance. An agent applies that logic. It should never be the thing deciding what your revenue recognition policy means.
Where the logic has to live
A finance leader needs a way to judge any AI initiative, agentic or not, before it touches a number that matters. The real test is whether you could stand behind the output: see what the workflow did, follow the logic it applied, and trace the result back through its data if an auditor asked. An agent that meets that bar is one you can scale. There is more to say about how a leader applies that test in practice, and a piece next month lays it out in full.
Meeting the bar comes down to an architecture question finance leaders increasingly own: where does the logic live? The answer that makes AI trustworthy is a governed business logic layer that sits between your raw enterprise data and the models and agents that consume it, preparing and validating the data and enforcing the rules finance owns.
That is where Alteryx sits. It is not another AI model competing for the forecast, and it does not replace your ERP or EPM systems. It does the work AI cannot do reliably on its own, so the agent built on top has something solid beneath it. Build that governed foundation first, and the agent becomes a natural next step once it’s in place.
Where this leaves the agent question
The interest in agents is the right instinct, and the teams acting on it now will have a head start. The only adjustment is the order of operations. An agent is what you earn once the data is governed, the logic is defined, and the outputs are traceable enough to defend in front of your auditor and your board.
Get that right, and the agent stops being a leap of faith. It becomes the next obvious step, and the question shifts from whether you can trust it to where you put it to work next.