The finance AI conversation is conducted almost entirely in generalities. "AI will transform the finance function" is not a claim that can be acted on, evaluated, or disagreed with. It is a mood.
This brief replaces the mood with specifics, using the only evidence base we can speak about with authority: our own. We report what the finance software market describes its AI as doing, measured across the vendors we track, with the method and its limitations stated plainly — because a research brief that overstates what its data supports is worse than no brief.
Method, and what this can legitimately claim
Population. 71 finance software vendors in the dilynx research base, spanning the close, reconciliation, AP, expense, FP&A, treasury and procurement markets. Of these, 50 (70%) carry curated AI capability analysis derived from cited sources.
What was measured. Two things. First, the distribution of publishable capability claims held by AI-describing vendors across the nine finance domains — where AI-positioned vendors actually operate. Second, the language of the curated AI descriptions themselves — what those vendors are described as doing.
Every claim in the underlying base carries at least one independent citation, and claims are publishable only when they pass the provenance gate. Nothing counted here is unsourced.
What this cannot claim, stated up front:
- It measures positioning, not deployment. A vendor describing anomaly detection is not
evidence that customers use it or benefit from it.
- It measures our tracked population, which is weighted toward the mid-market and toward
the domains we cover most deeply. It is not a census of the market.
- Language frequency is a proxy for emphasis, not for capability quality. A vendor may
do excellent work it describes poorly.
- It is a point-in-time reading of a fast-moving market, as of the review date.
Those limitations are real and they narrow the findings considerably. What survives is still worth reporting, because it contradicts the prevailing conversation in a specific and checkable way.
Finding 1 — Claims concentrate in two domains
| Finance domain | Claims | Share |
|---|---|---|
| Procure-to-Pay | 16 | 27% |
| Record-to-Report | 16 | 27% |
| Treasury & Cash Management | 10 | 17% |
| Plan-to-Perform (FP&A) | 8 | 14% |
| Finance Data & Systems | 4 | 7% |
| Governance, Risk & Controls | 4 | 7% |
| Order-to-Cash · Tax & Compliance · Finance Operating Model | 0 | 0% |
More than half of all capability claims held by AI-positioning vendors sit in just two domains: procure-to-pay and record-to-report. Both are high-volume, transactional and rule-dense — precisely the conditions under which machine learning has been productive for years, and precisely the conditions under which the work is least dependent on judgement.
The top individual capabilities follow the same logic: accounts payable automation, budgeting and forecasting, account reconciliation, ERP and data integration, SOX controls and audit readiness.
The implication for a buyer is unglamorous and useful. If your AI ambition sits in order-to-cash, tax, or the operating model itself, the market we track offers materially less than the discourse implies. That is not a reason to abandon the ambition; it is a reason to expect to build or partner rather than buy, and to price accordingly.
Finding 2 — Predictive language dominates; operational language does not
| Capability language | Vendors | Share of the 50 |
|---|---|---|
| Prediction | 16 | 32% |
| Forecasting | 10 | 20% |
| Anomaly detection | 6 | 12% |
| Coding / classification | 6 | 12% |
| Narrative generation | 4 | 8% |
| Copilot / assistant framing | 4 | 8% |
| Recommendation | 3 | 6% |
| Matching | 2 | 4% |
| Agentic | 1 | 2% |
Prediction and forecasting together dominate — appearing in roughly half the descriptions — while matching appears in 4% and coding in 12%.
This is the inverse of where the operational value in transactional finance actually sits. Matching and coding are high-volume, tolerant of a review step, and directly reduce close and invoice-cycle effort. Prediction is harder to validate, harder to control, and its value is contingent on a decision changing as a result.
Two readings, and honesty requires holding both:
- The charitable reading: matching and coding have become so standard that vendors no
longer treat them as differentiators worth describing, so the language understates them. This is plausible and probably explains part of it.
- The less charitable reading: prediction sells. It maps onto executive aspiration, and
it is harder to disprove in a demo than a match rate is.
Both readings point a buyer the same way: evaluate matching and coding performance explicitly, because the market is not competing on it in its language even where it is competing on it in its product. Ask for match rates against your own exception population, not the capability list.
Finding 3 — "Agentic" is a discourse phenomenon, not yet a market one
One vendor in fifty — 2% — carries agentic language in its curated description.
Set against how thoroughly the term dominates conference agendas, vendor marketing and finance media, this is the most striking number in the analysis.
The honest interpretations, in order of likelihood:
- Lag. Our curated analysis reflects sourced, corroborated capability rather than the
current marketing cycle, and the gap between a positioning shift and citable evidence of capability is real. Some of this is timing.
- Substitution. Vendors may describe agentic behaviour in other words — "automated
workflow", "touchless processing" — that our language analysis does not capture as agentic. This is a genuine limitation of the method.
- The discourse is ahead of the product. The simplest reading, and consistent with the
pattern in the previous two findings: the market is discussing a capability class it has not yet largely shipped into finance.
We cannot distinguish between these with the evidence available, and it would be dishonest to pretend otherwise. What we can say is narrower and still useful: as of this review, an executive planning around widely available agentic finance capability is planning around something our evidence base does not yet show.
That is not a prediction that it will not arrive. It is a statement about now, and it should be read against our benchmark data: close automation coverage sits at a median of 55% and a top quartile of 75% across our peer set. A market where a quarter of routine close steps remain manual in the top quartile is not a market that has broadly deployed agents. The two datasets are consistent with each other.
What this means for a buyer
Four practical consequences.
- Ask what the AI does, in process terms, and check it against this distribution. If a
vendor's AI story is predictive but your problem is throughput, the story is not addressed to your problem.
- Test matching and coding on your own exception population. It is under-described in
market language and is where transactional value concentrates.
- Discount agentic claims until the evidence is specific. Ask what action the system
takes without a human, under what boundary, and what the audit file contains. Most current claims dissolve at that question — which is itself informative.
- Expect less coverage outside P2P and R2R. If your ambition is in O2C, tax or the
operating model, budget for building or partnering.
What we will do next
This analysis will be re-run and republished on each substantive expansion of the research base. The agentic figure in particular is the one to watch: if interpretation (1) is right, it should rise materially within a year. If it does not, interpretation (3) was correct and we will say so.
- AI in finance concentrates where the work is transactional. More than half of tracked
capability claims sit in procure-to-pay and record-to-report; three of nine domains have none at all.
- **The market's language over-indexes on prediction and under-indexes on matching and
coding** — the inverse of where operational value sits. Evaluate the latter explicitly.
- "Agentic" appears in 2% of curated descriptions. Whatever the explanation, an
executive planning around broadly available agentic finance capability today is planning around something the evidence does not yet show.
Where this leaves you. Take the AI capability you are currently being sold and locate it on Exhibit 2. If it sits in the top two rows, ask what decision will change as a result. If it sits in the bottom rows, ask for performance against your own data — that is where the question gets answered.