Industry Insight
Finance adopts AI personally. Not institutionally.
Individual use of AI in finance is climbing steeply while functional deployment stays flat. This memo sets out why we think that gap exists, what evidence supports the argument, and (deliberately) which parts of it are not yet proven.
Finance holds the most structured data in the enterprise and runs the least artificial intelligence on it. The usual explanations — risk aversion, regulation, legacy estate — are consistent with that observation but do not account for its shape.
Individual adoption is climbing steeply. Functional adoption is flat. Something is preventing the work itself from being rebuilt around tools that practitioners have already accepted, and it does not appear to be appetite.
The findings at a glance
What we think is happening
These are readings of the evidence rather than findings. Section 6 separates what is established from what is ours.
- The constraint may be epistemic rather than technical. The knowledge needed to act on financial data appears disproportionately tacit — it sits in practitioners, not systems.
- Five generations of automation solved the movement of data. None solved the movement of judgement. Each extended reach; none extended memory.
- Access is commoditising and is not the differentiator. Retrieval protocols standardise access to systems; they carry no semantics about what the data means.
- If that is right, it is a design gap rather than a capability ceiling — which would be the more encouraging of the two diagnoses.
Scope and method
Prepared as background for Finaisse's positioning and published in full rather than summarised. Every figure is attributed and every source linked. Where a source is vendor-published rather than independent, it is marked. Section 6 sets out what this argument does not establish — read it before quoting anything here.
1 · Why is finance slow to adopt AI?
The adoption data does not describe a function that is reluctant. It describes a function whose individuals have adopted and whose institutions have not.
Adoption, 2026
| Finding | Source |
|---|---|
| 56% of finance leaders use AI daily — from 31% (2024) and 17% (2023) | CFO Connect |
| 59% of finance functions use AI — up one point year on year | CFO Connect |
| Finance ranks last among business functions for AI deployment | CFO Connect |
| 68% of CFOs cite not knowing where to start | CFO Dive |
| Data readiness is the leading blocker to AI return on investment | CFO Dive |
| Fully embedded AI up 91% YoY among PE/VC-backed CFOs (n=102) | Consero |
Were the obstacle regulatory or cultural, individual usage would be suppressed alongside institutional usage. It is not. Practitioners have adopted these tools readily; their organisations have not restructured around them. Something prevents the work itself from being rebuilt, and it does not appear to be appetite.
2 · What five generations of finance automation left behind
Practitioner literature converges on roughly five phases. We treat the sequence at greater length elsewhere; what matters here is the residue.
Deterministic bookkeeping
Solved repeatable, well-formed transactions. Left behind anything unstructured or exceptional.
Pattern recognition
Solved classification, matching, prediction. Left behind why a given pattern matters here.
Task execution across systems
Solved swivel-chair labour. Brittle to change; encoded no meaning.
Language over documents
Solved extraction, drafting, summarisation. Left behind grounding — confident error.
Goal-directed action
Solved orchestration and follow-through. Left behind governance, traceability, accountability.
The observation
Each generation extended reach — more formats, more systems, more steps. None extended memory. No generation introduced a durable representation of the judgements practitioners make. Agentic systems therefore reach further while knowing no more, which is the configuration that produces confident, plausible, incorrect action.
3 · Is this a tacit-knowledge problem?
Finance literature offers no strong theory of why this layer resists digitisation. Knowledge-management research does.
Nonaka and Takeuchi's SECI model distinguishes tacit knowledge — personal, context-dependent, experiential — from explicit, codified knowledge, and describes four conversion modes between them: socialisation, externalisation, combination, internalisation. Three findings transfer directly.
- Tacit knowledge is the hardest to store or share, and transfers most effectively through lived experience rather than abstraction (Starmind).
- Organisations typically discover its value only after it leaves — through churn, retirement or a change of role.
- Advantage arises from converting tacit knowledge into explicit, transferable artefacts. SECI names this step externalisation.
Applied: a controller who knows a given vendor short-pays freight holds tacit knowledge. Current practice supports only socialisation — she tells a colleague — which is why the knowledge leaves when she does.
The missing mechanism
As far as we can establish, no finance system in general use performs externalisation — though we would be glad to be shown one. We can find no field on the item where “we established this is a timing difference” becomes durable, machine-readable state. The judgement is formed, acted upon, and discarded.
This reframing is load-bearing. If AI in finance were constrained by model capability, the rational response would be to wait. If it is constrained by the absence of an externalisation mechanism, the response is to build one.
4 · Why AI in finance fails without grounded context
Evidence on ungrounded retrieval is consistent in direction, if not always in magnitude.
Grounding
| Finding | Source and caveat |
|---|---|
| Hallucination rates fall by approximately 87% with well-structured rather than unstructured retrieval sources | Branch8 — single source, treat as directional pending corroboration |
| Enterprise reporting requires grounding in governed business semantics: approved metrics, dimensions, joins, row-level security | GROUND, arXiv 2608.26157 |
| Retrieval grounds knowledge but does not cover action; no single technique is sufficient | Glean |
| Finance-specific retrieval and hallucination benchmarks now exist | FinDER, arXiv 2504.15800 |
The 87% figure is the most consequential number here, and we flag its single-source status deliberately. If it survives replication, the dominant variable in AI reliability is not model selection but the structure of the context supplied before inference. An organisation that structures its interpretive layer would obtain an order-of-magnitude reliability improvement without changing model at all.
Finance is the least forgiving setting for this. Figures must be exact, point-in-time and entity-correct; a near-miss is simply a wrong answer.
5 · Does MCP solve the connection problem?
The obvious objection is that retrieval protocols already address this. Model Context Protocol and comparable standards allow models to reach the ERP, banking, billing and CRM systems. If everything is reachable, is connection not solved?
MCP, 2026
| Finding | Source |
|---|---|
| SDK downloads ~97M/month by March 2026, from ~100k at launch in November 2024 | Cybertize |
| Only 41% of software technical leaders report even limited production use | Cybertize |
| Only 8.5% of servers implement the mandatory OAuth 2.1 standard | Cybertize |
| The “visibility gap” — agents deployed without visibility into their actions — is the most cited enterprise concern of early 2026 | Atolio |
MCP is genuine and important infrastructure. It resolves the adapter problem — the N×M cost of connecting models to systems — and will be table stakes within a year. It is, explicitly, a transport and access standard. It carries no semantics regarding what retrieved data means: no representation of normality, of materiality, or of decisions already taken.
MCP standardises access to systems. It does not standardise meaning across work.
The gap between 97 million monthly downloads and 41% limited production use suggests the field is arriving at the same conclusion. Connection proved to be the tractable half.
6 · What this does not establish
Separating evidence from inference matters more here than the argument does. The split is as follows.
Evidence vs inference
Well evidenced
Finance lags other functions; functional adoption has plateaued
Data readiness is the leading reported blocker
Structured context materially reduces hallucination (single source)
Auditors want reasoning traces; human-in-the-loop is near-universal
Our inference — not proven
That “data readiness” is principally an interpretive problem rather than an integration one. No survey isolates this.
That tacit-knowledge externalisation is the specific missing mechanism. Theoretically grounded in SECI; not empirically tested in finance.
That carrying context improves close outcomes. Plausible, and unmeasured.
We hold the left column as fact and the right as argument. Anyone citing this memo should preserve that distinction.
A further caution: several widely circulated benchmarks in this space are vendor-published marketing assets. They are directionally useful and should not be cited as independent evidence. We have excluded them from the tables above.
7 · How often do finance teams re-derive a decision?
No published study measures how often finance teams re-derive a judgement they have already made. The re-work cost of lost interpretive context has not been quantified.
That is a conspicuous gap, because it would be inexpensive to close. A short instrument administered to a few hundred controllers would establish a baseline and convert this memo's central assertion into a finding.
We would rather see it measured than assumed — including at the risk of the number returning smaller than we expect.
8 · What follows, if the argument holds
Each of these depends on the inference in section 6 being correct. They are stated as consequences, not as established positions.
- A hypothesis worth testing: is the constraint externalisation rather than intelligence? If it is, improved models will not close a gap produced by absent state.
- Governance and context converge. Auditable AI requires stored judgement; solving one substantially solves the other.
- Trust is the adoption variable. Perceived reliability governs acceptance more than raw capability, which makes explainability an adoption mechanism rather than a compliance cost.
- Access is commoditising. Whatever differentiates after that necessarily sits above the transport layer.
- The measurement gap is open. We would expect whoever quantifies context loss in finance to shape the vocabulary around it — but that is a prediction, not a finding.
None of this requires a breakthrough. It requires treating the judgement layer as something a system is accountable for holding — captured at the moment a decision is made, rather than reconstructed under close-week pressure.
Sources
- CFO Connect — State of AI in Finance 2026
- CFO Dive — Top AI adoption challenges facing CFOs in 2026
- Consero — 2026 CFO Survey
- SECI model of knowledge dimensions
- Starmind — the most valuable tacit knowledge
- Branch8 — AI hallucination risk mitigation
- GROUND — grounding enterprise reporting (arXiv 2608.26157)
- FinDER — finance retrieval benchmark (arXiv 2504.15800)
- Glean — when LLMs hallucinate in enterprise contexts
- Cybertize — MCP adoption report 2026
- Atolio — enterprise MCP limitations at scale