A few weeks ago, a CFO at a €60M Finnish industrial group described his finance operation in one sentence: "Our accountants do everything by hand in Business Central. Nothing is automated." He wasn't asking for an AI strategy or an agentic workflow. He wanted his team to stop manually keying journals.
That conversation was one of more than 100 I had this year with heads of finance at European companies in the 50–500 employee range. As part of it, I ran a cold-outreach campaign to 109 of them and saw a 12.8% engagement rate. The conversations that landed had little to do with the language the AI-finance category uses. Nobody raised "autonomous accounting." Nobody asked about replacing their team with agents. They asked about bank reconciliations, e-invoice approval routing, and why their tax codes still need manual posting.
This is a field report on that gap — what vendors market, what analysts forecast, and what the buyers in front of me actually asked for.
Disclosure: I'm building Artifi, an AI tooling layer for finance teams, so I have a stake in this market. The interviews below are field notes, presented as I found them rather than as a product argument.
What the category markets
Most AI-finance vendors describe a similar promise. Vic.ai positions "VicAgents" as "task-specific, intelligent AI agents trained to autonomously execute finance workflows" (Vic.ai). Trullion frames the trajectory as the "evolution of AI in accounting" toward autonomous agents (Trullion). AppZen emphasizes reviewing 100% of transactions instead of 5% sampling. HighRadius, Numeric, ChatFin, and Tipalti converge on a comparable arc: the autonomous finance function is arriving, and the cost of waiting is competitive disadvantage.
The analyst class broadly echoes the direction. Gartner projects that CFOs who implement "strategic AI deployment" will add 10 margin points of growth by 2029 (Gartner). McKinsey reports 44% of CFOs using generative AI across five or more use cases, up from 7% a year earlier (McKinsey). One survey found 79% of CFOs saying at least 25% of their accounting workload is being handled by agentic AI (Journal of Accountancy).
Taken together, these sources describe AI agents already running meaningful parts of the finance function.
What buyers asked for
The conversations I had pointed somewhere narrower. Five workflow patterns recurred. None are dramatic; all are specific.
1. Multi-jurisdiction tax-code posting on high-volume transactions. A fintech processing more than 100,000 Stripe transactions a month across the EU, Estonia, the US, and rest-of-world has to group each transaction into a tax-code bucket before it posts to the GL correctly. Today an accountant does this by hand every month. What drew the buyer's interest was a single concrete description of that job — grouping the Stripe transactions by tax code and posting them — rather than a platform-level pitch.
2. E-invoice approval routing. A mid-market Estonian transport-services group described its accounts-payable bottleneck as a routing problem, not a data-entry one. When an e-invoice arrives, someone identifies the right approver, escalates when that person is away, chases when it stalls, and reconciles when it is re-routed. Their AP system handles the document; it does not handle the human routing graph.
3. Deduplication in multi-party sales records. The same group resells transport services through a chain — the same shipment appears as a sale in their system, again in their customer's, and again in the retailer's record. Reconciling and transforming those duplicates consumes real hours each week before anything can be invoiced cleanly. It is a problem few vendors describe publicly, and a recurring cost for the company.
4. Internal reporting apps built by finance, not engineering. A Finnish software vendor with 300–500 staff located its pain above the close: limited internal analytics on Business Central, and no quick way to build a one-off board report without filing an engineering ticket. The ask was for the finance team to build small internal tools itself — "self-serve analytics," a fifteen-year-old category they still did not have.
5. Bank reconciliation. Several CFOs raised this independently. A Gartner survey cited by reconciliation vendors found 18% of accountants make financial errors daily and 59% monthly, much of it a byproduct of repetitive manual work (Kolleno). BlackLine and Trintech have sold reconciliation automation for over a decade, yet adoption among 50–200-person European companies remains uneven, and much of the work is still done in spreadsheets.
The common thread: the ask was a workflow, not a category. Buyers were not shopping for an AI platform; they had a specific, recurring job that — once described plainly — was enough to start a conversation. One CFO at a Finnish industrial group put it directly: "We won't switch ERP, but we'd buy the layer."
What didn't come up
What was absent is as informative as what recurred.
No one asked to replace their team. In 100+ conversations, it did not come up once. The CFO at a 60-person company has a three-person finance team and is not trying to get to two; she is trying to stop the three from working evenings.
No one asked for a fully autonomous close. "AI closes your books for you" did not resonate; "clear the prepaid and accrual cut-offs on the bills export" did — the same end-state at a different level of abstraction.
No one asked for agents acting on their own authority. Even AI-positive CFOs wanted approval gates. Gartner's 2025 survey shows finance AI adoption at 59%, up a single point from 2024 after a sharp jump from 37% in 2023 (CPA Practice Advisor) — consistent with a first wave that has adopted and a second wave that is watching.
And below roughly €1M in revenue, the demand was largely absent. An accounting-firm owner with about 200 SMB clients on Merit was blunt that her sub-€1M clients do not want an AI interface for their books; the interest appeared above €1M, where finance tends to become a dedicated function.
Why the gap exists
The vendor pitch is not irrational; it is calibrated to a different audience. Venture-scale fundraising rewards a platform story sized to the global finance function — "autonomous AI agents for the office of the CFO" tends to raise more readily than "we automate one tax-code posting workflow for fintechs." The homepage and the pitch deck are often the same document, and its primary readers are investors and analysts.
Buyers appear to read it that way too. Two-thirds of mid-market CFOs call human oversight of agentic AI "extremely or very critical," and 86% of finance teams report encountering hallucinated data while using AI for finance tasks (Journal of Accountancy). The MIT finding that circulated in late 2025 — 95% of generative-AI pilots delivering no measurable P&L impact — reached most CFO inboxes (Fortune). In financial services specifically, one analysis estimates 80% of AI projects do not reach production, and of those that do, 70% do not deliver measurable value (Financier Worldwide). When Klarna replaced 700 customer-service staff with AI, the reversal — rehiring humans in 2025 after satisfaction fell — drew far less coverage than the original move (Entrepreneur).
The result is two sides calibrated differently: vendors producing platform-level messaging because their funding model favors it, and buyers discounting that messaging because their risk model favors caution. In the conversations I had, the exchanges that progressed were the ones where a single workflow was described in plain language.
What this means for each group
For vendors: in these interviews, narrow, workflow-specific messaging correlated with more meetings, while broad "office of the CFO" positioning correlated with more skepticism. The MIT analysis also found vendor-bought AI reached success roughly 67% of the time, versus about a third as often for internally built tools (AI Magazine) — relevant to where buyers currently place their trust.
For finance teams: the conversations that went somewhere were the ones where the buyer could name a workflow in quantified terms ("four days a month on this," "two of five people do nothing else in week one of the close"). Specific pain tended to attract specific, testable solutions; vague pain attracted vague pitches.
For the category: the distance between platform-level marketing and workflow-level demand is wide enough that it is likely to close in one of two directions — vendors describing more precisely what their product does inside a real finance team on a given Tuesday, or buyers solving these workflows themselves with general-purpose tools.
A starting point, not a ceiling
One thing the interviews don't settle is whether narrow demand stays narrow. The workflow-led pattern describes how buyers enter, not necessarily where they stop. In adjacent finance categories the entry point has tended to widen: Ramp and Brex began with a single workflow — corporate spend — and expanded into reconciliation, bill pay, accounting sync, and treasury; Rippling started in one HR workflow and grew into a platform. If that pattern holds here, a tax-code posting job or a reconciliation agent reads less as a destination than as an on-ramp — the first workflow earning the access and trust that later make a broader system viable. Stated neutrally: the data shows where adoption begins, not where it ends.
The takeaway
Across 100+ conversations, the most consistent signal was that mid-market demand concentrated on specific, nameable tasks rather than on "AI" as a category. The buyers who engaged could finish the sentence "we spend [N] days a month on [one task]" — and that, more than any platform narrative, was what determined whether a conversation went anywhere. Whether those entry points stay narrow or widen into broader systems is the open question the next few years will answer.