All insights
Fintech AI ConsultingJuly 11, 20265 min read

AI Agents in Financial Services: Why Most Never Leave the Pilot

80% of organisations report returns from AI agents, yet Gartner expects over 40% of agentic AI projects to be cancelled by 2027. What separates the two, and what UAE leaders in production do differently.

In this article
AI Agents in Financial Services: Why Most Never Leave the Pilot

Two numbers define applied AI in 2026, and at first glance they contradict each other.

  • 80% of organisations say their AI agent investments are already delivering measurable economic returns, according to Anthropic's 2026 State of AI Agents report, a survey of more than 500 technical leaders.
  • More than 40% of agentic AI projects will be cancelled by the end of 2027, according to Gartner, because of rising costs, unclear business value or weak risk controls.

Both are true. The gap between them is where most of the value sits, and where most of the wasted budget goes.

The bottleneck isn't the model anymore

In the same Anthropic report, 46% of organisations named system integration as their biggest barrier. Model intelligence is no longer the main constraint. Gartner has also warned about agent washing: of the thousands of vendors claiming agentic AI, it estimates only about 130 offer the real thing. And Capgemini found that trust in fully autonomous agents fell from 43% to 27% in a single year, as organisations moved from experiments to real deployments.

In plain terms: the demo works. What kills projects is everything around it. Connecting the agent to core banking, the KYC stack, the payments hub and the approval chain, then proving to a risk committee that it won't do something expensive.

What UAE leaders are doing

UAE banks have moved further than most markets, and they're publishing numbers.

  • Emirates NBD has more than 150 AI initiatives underway, 65 of them in production. It became the first bank in MENA to run OpenAI's GPT models on in-country infrastructure, and it uses frontier models defensively to probe its own systems for weaknesses.
  • First Abu Dhabi Bank has moved more than 30 agentic AI use cases into its production pipeline, across trade and payments, client operations, compliance and engineering. That work runs through a central AI Innovation Hub that supplies governance and reusable components.
  • Commercial Bank of Dubai estimates that AI agents could handle 55–60% of operational tasks, with people still approving outcomes. It also intends to own the agents it builds and the IP behind them.
  • ruya, a digital Islamic bank, has agentic AI live in production with Dubai-based Magure.

The pattern is consistent: a central function that owns governance and reusable building blocks, a portfolio of narrow use cases, human approval where it matters, and clear ownership of data and IP.

The demo works. What kills projects is everything around it.

Fraud: the use case that forces the issue

If you want a use case where AI clearly pays for itself, look at fraud. The attackers are already using AI.

  • In the UK, which publishes some of the most detailed fraud data, authorised push payment (APP) fraud losses rose 19% to £576.4 million in 2025, according to UK Finance.
  • Sumsub recorded a 94% rise in deepfake fraud attempts in the UK during 2025.
  • Real-time rails shrink the window to act. Aani's transfers grew sixfold year on year, and instant payments settle before traditional controls can step in.

The defensive results are just as concrete. HSBC's adaptive machine-learning system reportedly cut false positives by 60% while finding two to four times more suspicious activity across roughly 980 million transactions a month.

Fraud has every trait of a good AI deployment: a clear P&L line (losses, false positives, analyst hours), high data volume, people already reviewing decisions, and a regulator who wants it done well.

A five-point playbook for getting to production

  1. Start with a workflow that has a number attached. Cost per case, time to decision, loss rate, false-positive rate. If you can't measure it before, you can't prove it after.
  2. Do the integration work first. Budget more for APIs, data access and identity than for the model. Integration, not intelligence, is what stalls almost half of deployments.
  3. Build governance into the architecture. Give each agent scoped permissions and a full audit trail. Require human approval for anything that moves money or changes a customer record, and build a kill switch. Governance added after a production incident is how projects end up in Gartner's 40%.
  4. Evaluate continuously. Build test sets from real cases, measure weekly and alert on drift. A model that scored well in the pilot can degrade quietly in production.
  5. Own what matters. Data residency, in-country hosting where required, and ownership of the prompts, workflows and fine-tuned models you build. Emirates NBD and CBD have both made this a priority.

The bottom line

In 2026 the question isn't whether AI agents work. It's whether your organisation can run them: integrated, governed, measured and owned. Narrow and production-grade beats broad and impressive every time.

Sources
  1. 5 takeaways from the 2026 State of AI Agents report — Arcade (summarising Anthropic's report)
  2. 5 surprising truths from the 2026 State of AI Agents report — Collabnix
  3. Why most agentic AI projects stall before they scale — CIO, Feb 2026 (Gartner forecast)
  4. Gartner: almost half of agentic AI projects will be scrapped — Process Excellence Network
  5. Agentic AI in banking UX — Onething Design (cites Capgemini)
  6. Inside the AI revolution reshaping Dubai's banking sector — Fortune, Nov 2026
  7. FAB shifts from AI pilots to enterprise-wide deployment — Middle East AI News, Feb 2026
  8. UAE digital bank ruya deploys agentic AI — Khaleej Times, Aug 2026
  9. UK fraud losses and deepfakes — IBTimes UK (cites UK Finance Fraud Report 2026)
  10. AI in the fight against fraud — Payment Expert, Apr 2026 (cites Sumsub)
  11. UAE fintech boom and instant payments — Khaleej Times
  12. Fintech fraud prevention in 2026 — Pulse / HedgeThink, May 2026
How relevant and useful is this article for you?
From Ripae

Ripae Labs builds applied AI for credit, fraud and operations, with privacy and auditability designed in from the first line.