AI in finance: why it’s failing now—and how to win

Only 15% of executives said AI improved profit margins over the last year, and just 5% reported “widespread value” from AI programs. That gap—between belief and results—is the most useful thing you can pay attention to in 2026.

If you work in an Australian bank, lender, insurer, or fintech, you’ve probably lived this: a promising generative AI pilot, an impressive demo, then reality—hallucinations, inconsistent answers, compliance anxiety, and a workflow that quietly drifts back to humans.

Here’s the stance I’ll take: AI in finance is absolutely the future, but most implementations are failing because they’re being treated like software instead of a risk-managed operating model. The winners in fraud detection, credit scoring, and algorithmic trading won’t be the firms that “try AI.” They’ll be the firms that build reliable systems around it.

The real problem isn’t AI hype—it’s reliability debt

AI projects stall because organisations accumulate reliability debt: the hidden cost of making probabilistic systems behave like deterministic software.

The RSS story nails the pattern. A wine app built an “AI sommelier” that was too polite to tell users they’d hate a bottle—classic model “sycophancy.” A rail company tried to summarise a 100-page safety rulebook and got inconsistent results—forgetting, misinterpreting, and inventing details. A payments firm scaled an AI support agent, then had to admit humans still matter when situations get emotional or complex.

Finance feels this harder for three reasons:

1. Your edge cases are the business. Fraud, disputes, vulnerable customers, and unusual trading conditions aren’t rare—they’re where money is won or lost.
2. Your tolerance for error is low. A wrong wine recommendation is annoying. A wrong credit decision can be illegal.
3. Your data is messy on purpose. Product lines, acquisitions, legacy cores, and channel-specific definitions create a swamp where models see patterns that don’t exist.

The practical takeaway: If your AI plan doesn’t include a plan for reliability, you don’t have an AI plan. You have a demo plan.

“Jagged frontier” is the default—design like it

Researchers call it the “jagged frontier”: models can be brilliant at one task and strangely bad at another. In finance, that means an LLM might write a decent policy summary but fail at something “simple” like applying a specific exception, interpreting time windows (“last week”), or respecting internal definitions (“active customer”).

Where this hits finance teams first

• Fraud detection operations: LLMs can summarise a case file well, but may miss the one line that matters (a device fingerprint mismatch) or fabricate rationale.
• Credit scoring and underwriting: Models can draft adverse action notices, but can’t be trusted to infer or “fill in” missing data—because regulators won’t accept “the model assumed.”
• Algorithmic trading and research: LLMs can accelerate research notes, but are unreliable at factual claims unless grounded in approved datasets and constrained tool calls.
• Customer service in banking: AI is fine for balance queries and password resets; it struggles with hardship, bereavement, scams, and disputes—where empathy and judgment are the product.

The better approach is to treat AI as a variable-precision component. Use it where being “mostly right” is still valuable, then wrap it with controls so it can’t harm customers or create regulatory exposure.

A simple design rule that works

Use generative AI to draft, triage, and explain—not to decide, approve, or execute without guardrails.

That one sentence prevents a lot of expensive mistakes.

Data formatting isn’t busywork—it’s the bottleneck to ROI

One of the most overlooked lines in the source piece: financial data comes from many sources with different formats, which can prompt AI to “read patterns that don’t exist.” That’s not theoretical. It shows up as:

• contradictory customer identifiers across systems
• inconsistent transaction descriptors
• different time zone handling
• merchant/category mappings that vary by channel
• free-text notes that contain critical context but no structure

AI teams often pitch “we’ll put a chatbot on top of our knowledge base.” Then the chatbot starts answering with confidence—using outdated policy PDFs, duplicate versions, or content that contradicts current product terms.

If you want AI ROI in banking, you almost always need a data readiness sprint before you scale anything:

1. Define canonical entities (customer, account, device, merchant, case, interaction).
2. Create a “gold” knowledge layer for policies and procedures (versioned, owned, testable).
3. Fix the top 20% of data issues that cause 80% of model failures (formats, missing fields, duplications).
4. Add retrieval and grounding so the model cites your approved sources, not its memory.

This is why some firms will delay planned AI spending: not because AI is useless, but because their foundations can’t support it yet.

High-impact, low-lift AI use cases in fraud, credit, and trading

AI vendors are increasingly encouraging companies to start smaller—“high impact but low lift.” For finance, that’s the right instinct, as long as you pick use cases that are measurable and controllable.

Fraud detection: faster decisions without auto-approving anything

Start with fraud analyst copilots, not autonomous fraud agents.

Good first wins:

• Case summarisation: turn multi-system evidence into a 10-line brief with links to source fields.
• Reason-code drafting: propose a rationale that analysts can accept/edit (and that you can audit).
• Queue triage: group cases by pattern and urgency (new mule account behaviour, first-party fraud signals).

Success metric examples:

• reduce average handling time per case by 20–40%
• increase analyst throughput per shift
• reduce “re-opened case” rate

Credit scoring: documentation and consistency, not “black-box approvals”

You can create value without touching the score itself.

• Application data validation: flag missing or inconsistent fields before underwriting.
• Policy Q&A for underwriters: grounded answers from the latest credit policy and exceptions matrix.
• Adverse action notice drafting: consistent wording aligned to regulated requirements (human review required).

Success metrics:

• fewer manual back-and-forth requests
• reduced decision cycle time
• improved audit outcomes (consistency + traceability)

Algorithmic trading: research acceleration with strict guardrails

For trading and investment research, the temptation is to ask a model for ideas and treat the output like insight. Don’t.

Instead:

• use AI to summarise research packets
• generate first drafts of commentary tied to approved data
• produce scenario checklists (what to verify, what can break)

Success metrics:

• faster production of research notes
• fewer manual formatting tasks
• improved coverage breadth without sacrificing controls

Why humans are “back” in the loop—and why that’s good news

The source story describes organisations returning to humans in customer service after chatbot rollouts. That’s not a failure; it’s a correction.

In financial services, the human role doesn’t disappear—it changes:

• Humans become exception handlers for complex cases.
• Humans become quality controllers who review AI drafts and catch weak reasoning.
• Humans become process designers who tune prompts, workflows, and escalation paths.

A practical model I’ve found works well is tiered autonomy:

1. Assist: AI drafts and summarises; humans decide.
2. Recommend: AI proposes actions with confidence signals; humans approve.
3. Act with limits: AI executes low-risk actions (e.g., routing, templated comms) with strict thresholds.
4. Act broadly: rare in finance; requires mature controls and strong governance.

If your organisation tries to jump from 0 to tier 4, you’ll end up like many early adopters: big pilots, limited production value.

A finance-ready checklist for AI that actually works in 2026

If you’re planning budgets right now, use this as a sanity check before committing to a major generative AI program.

1) Governance that matches financial risk

• Documented model purpose and limits
• Clear ownership (product + risk + compliance)
• Audit logs for prompts, sources retrieved, outputs, and human approvals
• Testing standards (including red-teaming for fraud/social engineering)

2) Grounding and source control

• Retrieval-augmented generation tied to approved documents and datasets
• Versioned knowledge base (policy, fees, terms, procedures)
• Output citations to internal sources (even if not shown to customers)

3) Evaluation beyond “it looks good”

• Golden test sets for fraud cases, credit edge cases, and support scenarios
• Measurement of hallucination rate and policy adherence n- Monitoring drift over time (policy updates, product changes, seasonality)

4) Workflow-first delivery

• Embed AI into the tools teams already use (case management, CRM, underwriting workbench)
• Design escalation to humans with clear triggers
• Train staff on “how it fails,” not just “how to prompt it”

5) Vendor support that isn’t just a contract

The source article points out vendors are building “forward deployed” teams and applied AI experts to embed with clients. In finance, that matters.

You want partners who will:

• map real workflows end-to-end
• co-design controls with risk teams
• help you instrument measurable outcomes

If a vendor can’t explain how they’ll reduce hallucinations in your fraud or credit workflows, walk away.

The competitive edge: invest now, but invest in the right layer

AI spending may be delayed in some industries, but finance doesn’t get the luxury of waiting. Scam volumes, real-time payments, open banking data flows, and customer expectations aren’t slowing down in 2026. If you pause too long, you’ll be competing against institutions that have already built reliable AI operations—especially in fraud detection, credit decisioning support, and AI-driven customer service triage.

The winning move isn’t to “buy more AI.” It’s to build the pieces that make AI dependable: clean knowledge, controlled data, measurable evaluations, and workflows that keep humans in the right places.

If you’re planning your next quarter’s roadmap for AI in finance and fintech, what’s the one process—fraud cases, underwriting, disputes, trading research—where you can ship a reliability-first AI copilot and prove ROI in 90 days?