Banks are putting their AI agents on a budget — and under audit
82% of AI-investing leaders worry about token costs. The answer emerging this week: budgets, audit trails, and agents that cannot improvise.
- 82% of senior leaders at organisations investing in AI are concerned about token usage and costs, while 64% actively monitor usage and have clear budget guardrails. Self-reported m:…3438
- 94% of portfolio managers and analysts use AI or machine learning in their alternative data research process. Self-reported m:…3315
- N26 uses Rasa for regulated banking customer service, deflecting 30% of routine contacts. Self-reported m:…2373
- BDO announced a global agreement with Legora after an initial pilot earlier in 2026 in Spain and Norway. Self-reported m:…9237
The governance turn
Banking's AI conversation has moved from what the technology can do to how an institution runs it. Professor Luis Maldonado, delivering the opening address at the AI in Financial Services Regulatory Forum in Madrid on 6 November 2026, argued that the greatest productivity gains from AI occur when organisations redesign processes around new technologies rather than replacing existing tools. Institutions need to embed AI into standardised processes with clear mechanisms for testing, approval, deployment and measurement, he said, and AI sovereignty requires understanding dependencies across the technology stack, including infrastructure, computing capacity, models and applications. AI is transforming banking through value creation, operational efficiency, and risk and compliance — and it introduces new risks including deepfake fraud, systemic dependencies and model concentration. Regulations such as the Digital Operational Resilience Act place greater emphasis on operational resilience and third-party risk. m:…3810
Regulators are tightening the frame around that shift. Databricks notes that regulators like ESMA and the FSB are sharpening expectations around AI governance, model transparency and data lineage. The adoption numbers explain the urgency: 94% of portfolio managers and analysts use AI or machine learning in their alternative data research process, and 71% of managers say combining data from different sources is their most frustrating challenge. The award Databricks won — Best AI Solution for Data Intelligence and Governance at the second annual AI in Capital Markets Award 2026, voted on by practitioners across the industry — is itself a sign of what the market now rewards: not the cleverest model, but the best-governed one. m:…3315
The budget
The most concrete new control this week is a budget. Manifold has launched budget and spend tracking features for AI agents, and its survey data explains why: 82% of senior leaders at organisations investing in AI are concerned about token usage and costs, while 64% actively monitor usage and have clear budget guardrails. Manifold allows users to set one budget per person that covers multiple coding agents, with policies that can warn employees at 80% of their budget and block prompts at 100%, and steer routine work to cheaper models. A jump in AI agent spend is often the first sign an agent has gone off course, the company says — which makes the budget a safety instrument as much as a cost control. m:…3438
The work itself
The work itself is being rebuilt to fit inside those controls. Rasa argues that voice agents built only on a large language model can improvise, which is a risk in regulated settings, and that a well-built banking voice agent combines language models for understanding with defined logic for decisions. A voice agent for banking handles phone conversations about accounts, payments, cards, fraud and lending, and self-hosting or deployment in a bank's own cloud keeps customer data within the bank's environment. The results Rasa cites are concrete: N26 uses Rasa for regulated banking customer service, deflecting 30% of routine contacts, and nib Group uses Rasa agents for insurance, leading to a 26% increase in self-service. Financial services buyers prioritise compliance, deployment control and transparency, Rasa says. m:…2373
The same pattern shows up in the deals. RBC Capital Markets built Aiden QuickTakes, a proprietary AI agent on Databricks that can auto-generate draft research reports up to 95% faster, and LSEG partnered with Databricks to bring AI-ready financial data natively into analytics, AI apps and agents; LSEG AI agents can query more than 350,000 investments using plain English. And BDO announced a global agreement with Legora, an agentic operating system for legal professionals, after an initial pilot earlier in 2026 in Spain and Norway. The agreement provides BDO member firms with a common framework to adopt Legora, which includes permissions, ethical walls and an audit trail, and supports research, document review, drafting, summarisation and contract analysis. Legora is used by more than 2,500 law firms and in-house legal teams, across more than 80 markets, by more than 140,000 legal professionals — and BDO, with 2025 fee income of $10.8 billion, is now standardising on it. m:…3315 m:…9237
What is still unresolved
- Whether budget controls slow agent adoption or speed it by giving risk teams something to approve.
- Whether voice agents pass bank audit at scale, or stay limited to routine contacts.
- Whether model concentration — many banks depending on the same few models — becomes a systemic risk, as Maldonado warns it could.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.