AI in Finance: Why It’s Promising but Not Ready Yet

Only 15% of executives said AI improved profit margins over the last year. Another survey found just 5% saw widespread value. Those numbers should land hard if you’re a bank or fintech leader being asked to “roll out genAI” in 2026.

What’s happening isn’t a failure of ambition. It’s a mismatch between how AI is sold (“easy button”) and how AI behaves in production (“jagged frontier”). And finance feels that mismatch more than most industries because mistakes aren’t just embarrassing—they can be regulatory breaches, customer harm, fraud losses, or bad credit decisions.

This post is part of our AI in Finance and FinTech series. The goal here is simple: translate the broader business reality—AI’s future is clear, its present is messy—into practical steps Australian finance teams can take to get value now, without betting the farm.

The AI reality check: value is real, but uneven

Answer first: AI value exists today, but it’s concentrated in narrowly-scoped workflows with clean data, tight guardrails, and clear success metrics.

The RSS story captured what many execs are now saying privately: AI still doesn’t “work right now” in the way leaders expected after the ChatGPT shockwave. The issue isn’t that models can’t produce fluent text. The issue is operational reliability—repeatability, correctness, auditability, and safe failure modes.

A wine app had to spend six weeks coaxing a chatbot to be less polite and more honest. That’s funny—until you map it onto finance:

• A “helpful” model that won’t say “no” can become a policy-bending assistant.
• A model that invents details (hallucination) can create false compliance narratives.
• A model that’s inconsistent across runs can break controls and audit trails.

This is why Forrester’s prediction that companies will delay a meaningful chunk of planned AI spending is believable. Not because AI is fading—because organisations are getting stricter about what counts as “working.”

Why finance adoption feels harder than other industries

Answer first: Finance is data-rich but format-fragmented, heavily regulated, and dependent on deterministic processes—exactly the conditions where ungoverned generative AI causes friction.

The “jagged frontier” shows up fast in banking

Researchers call it the “jagged frontier”: models can crush complex tasks (math, code) and still fail at basics (calendar entries, consistent summarisation, interpreting time ranges like “last week”).

In finance, that frontier appears as:

• Great first drafts of customer emails, but shaky interpretation of policy exceptions
• Fluent summaries of a credit memo, but errors in numeric details and covenants
• Strong pattern recognition in fraud signals, but confusion when fields change names or formats

And when your environment includes core banking systems, multiple CRMs, broker feeds, legacy data warehouses, and third-party risk platforms, the model is constantly being asked to “read patterns that don’t exist.”

Data formatting is the unsexy blocker that decides ROI

A line from the source that matters for finance leaders: many firms rely on data compiled from broad sources, formatted differently, which can mislead AI tools.

Here’s what that looks like in real financial operations:

• Transaction descriptions differ by merchant, acquirer, country, and channel
• Customer identity attributes vary across onboarding, KYC refresh, and servicing systems
• Credit bureau data arrives with differing schemas over time
• Incident notes and call centre transcripts are unstructured and inconsistent

If you don’t standardise and label what matters, you’ll end up spending more time arguing with model outputs than using them.

Where AI works right now in finance (and where it doesn’t)

Answer first: AI delivers reliable value in finance when it augments decisions, automates low-risk tasks, and operates inside strong controls—especially in fraud, servicing triage, and document-heavy ops.

Below is a practical “do this / not that” map I’ve found useful when advising teams.

High-confidence use cases (good places to start)

These are the areas where AI in finance typically delivers measurable outcomes within one or two quarters.

1. Fraud detection augmentation

   • Use AI to prioritise alerts, cluster patterns, and summarise investigator notes.
   • Keep final actions (declines, blocks, reporting) behind rules and human approval.

2. Customer service triage (not full replacement)

   • AI classifies intent, retrieves account context, drafts replies, and routes complexity.
   • This matches what we’re seeing broadly: humans are still essential for empathy and edge cases.

3. Credit operations and document processing

   • Extract key fields from statements, payslips, bank statements, and covenant packs.
   • Pair it with validation rules and confidence thresholds.

1. Compliance and policy assistance (retrieval-first)

   • Use retrieval-based assistants to point staff to the right policy clause and procedure.
   • Don’t ask the model to invent policy interpretations.

2. Marketing and product analytics support

   • Drafting campaign variants, analysing feedback themes, generating experiment ideas.
   • Low downside when governed.

Medium-confidence use cases (require stronger design)

These can work, but only with more engineering and governance:

• Personalised financial insights inside apps (needs safe wording + suitability rules)
• Collections optimisation (avoid harmful tone; ensure hardship processes are respected)
• SME credit scoring enhancements (careful around bias, explainability, drift)

Low-confidence use cases (avoid for leads/brand safety)

These are the ones that create impressive demos and painful incidents:

• Fully autonomous “financial advisor” chatbots without suitability controls
• AI that generates final credit decisions without explainability and governance
• Models that summarise long regulatory obligations without verification loops

If you’re hunting for leads, this distinction matters: prospects don’t want AI theatre. They want the short list of use cases that survive contact with audit, risk, and customers.

The practical playbook: make AI reliable before you make it big

Answer first: The fastest path to ROI is a small, controlled deployment that improves one workflow end-to-end, with measurable metrics and human-in-the-loop design.

The source story mentioned AI vendors building “forward deployed” teams and offering more handholding. Finance teams should take that as a cue: treat AI as an implementation program, not a software purchase.

1) Pick a workflow, not a department

Choose one process with clear inputs/outputs and pain you can quantify.

Good examples in Australian banking and fintech:

• Disputes: classify, summarise, draft outcomes, route exceptions
• Merchant onboarding: extract documents, flag missing fields, produce a checklist
• AML investigations: summarise entity networks and prior case history

Bad example: “Make the whole bank more productive with genAI.” That’s how you end up with scattered pilots and no business case.

2) Design for “safe failure,” not perfection

AI will be wrong sometimes. Your job is to ensure that when it’s wrong, it’s wrong in a way that doesn’t create customer harm.

Practical controls that work:

• Confidence thresholds (auto-apply only above X)
• Mandatory citations from approved sources (policy docs, product T&Cs)
• Restricted actions (draft-only; no sending without approval)
• Audit logging of prompts, outputs, and user actions

3) Fix the data plumbing that blocks you every time

If you’re serious about AI adoption in finance, allocate budget and time to:

• A canonical data model for key entities (customer, account, transaction, product)
• Field-level data quality checks
• Standardised taxonomies (issue types, outcomes, reasons, hardship flags)

This is where many AI programs quietly die—because teams try to build “AI on top of mess” and then blame the model.

4) Measure what the CFO actually cares about

Track metrics that translate to dollars, risk, or capacity:

• Fraud: time-to-disposition, false positives, losses prevented
• Contact centre: containment rate with CSAT guardrails, handling time, escalations
• Credit ops: turnaround time, rework rate, exceptions per application
• Compliance: incident rate, policy adherence, audit findings

If you can’t measure it, you can’t defend next year’s AI budget—especially as more firms delay spending.

“People thought AI was magic.” Finance can’t afford that belief

Answer first: The winning posture for 2026 is pragmatic optimism: assume AI will deliver, but only after you constrain it, teach it your domain, and operationalise governance.

The RSS article’s best line is the simplest: AI isn’t magic. In financial services, magical thinking shows up as two expensive mistakes:

1. Over-automation: handing customer conversations or credit outcomes to systems that can’t guarantee correctness.
2. Over-piloting: endless proofs of concept with no integration, no controls, and no path to production.

There’s a better way to approach this. Start with one production-grade use case, build the guardrails once, and reuse the pattern.

If you’re planning AI investments for 2026, pressure-test each initiative with three questions:

• What decision or task gets faster, safer, or cheaper—specifically?
• What data does it rely on, and how messy is it right now?
• What happens when the model is wrong, and who catches it?

That’s the gap between AI hype and AI value. Where is your organisation still expecting the “easy button,” and where could you commit to a workflow that actually ships?