Artificial intelligence in financial services is rapidly evolving. 52% of the financial services firms surveyed for the University of Cambridge's 2026 Global AI in Financial Services Report are already adopting agentic AI. Of those, 29% are running pilots, and 23% have reached scaling or transforming. Adoption of AI in some form reaches 81%.
That 29-point gap is where most compliance functions currently sit. A model that summarizes a policy is useful. An agent that reads the same policy, inspects the verification flow it applies to, and drafts a replacement configuration is a different proposition because it acts on systems instead of producing text about them.
That shift raises one governing question: can every action an agent takes be authorized in advance and reconstructed afterwards?
What is an AI agent in compliance?
An #AIagent is an autonomous software system that uses a model to pursue a defined goal. The official OpenAI documentation describes agents as “applications that plan, call tools, collaborate across specialists, and keep enough state to complete multi-step work.”
In an AI compliance agent workflow, an underlying large language model (LLM) interprets task instructions and reasons, while the agent framework and connected compliance platform provide the data, permissions, and tools that allow the agent to act. The agent might read an anti-money laundering (AML) policy, identify its requirements, inspect an existing know your customer (KYC) flow, use an API to create a draft configuration, and present the result for approval.
Autonomy exists on a spectrum. An agent may complete low-risk steps independently but stop before changing a live workflow, rejecting an applicant, or submitting a regulatory report.
How agentic AI works: Agent skills and MCP explained
Agents need more than a language model to complete operational work reliably. They require a goal, instructions, access to relevant systems, and rules governing what they can do.
Model Context Protocol (MCP) is an open standard for connecting AI applications to external systems. An MCP server exposes tools with defined inputs, allowing an agent to discover permitted actions, supply the required arguments, and receive the results.
While MCP connects an agent to external systems, skills provide the procedural knowledge needed to complete a task. The open Agent Skills specification describes a skill as a folder containing instructions and, optionally, scripts, references, and other resources. An AI agent skill might teach an agent how to analyze a regulation or configure a KYC workflow. Skills and MCP AI agents can convert a broad request into a repeatable workflow.
The table below explains the function of each component:
From chatbots to autonomous compliance agents
A traditional chatbot mainly responds to prompts in front of it. It can explain a rule or summarize a policy, but it generally does not translate analysis into action. Generative AI compliance tools expanded these capabilities by drafting reports, summarizing case files, and extracting information from documents.
Agents coordinate actions instead. A request is broken into steps: tools are selected, results inspected, and the next step adjusts to what came back. Asked about onboarding documents – a chatbot lists them. An agent can open the company's current verification flow, find the checks missing against a given standard, build a proposed replacement, and put it in front of someone for sign-off.
Why compliance teams are adopting AI agents
Compliance work involves a great deal of translation. A requirement written in legal language becomes an operational rule, then a platform setting, which must be documented and evidenced. One policy change can mean separate updates across several products and jurisdictions, each done by hand, each an opportunity for the implementations to drift apart.
Agents can extract the requirements and follow a defined procedure without a specialist retyping every step. Consistency is the real gain here: the fifth jurisdiction gets the same treatment as the first.
The cost of getting this wrong scales the same way. Access rules, approval thresholds, and activity logs are what keep an agent's actions attributable to the person who authorized them. Without them, automation only increases the speed and scale of mistakes.
AI agents in AML compliance and KYC workflows
In AML, an agent can turn policy text into a proposed configuration ready to launch in a dashboard, summarize an applicant's review history, or assemble the applicant and transaction data an investigator needs before opening a case. It can also compare an updated policy against the current configuration and flag the controls that no longer match.
KYC work is mostly setup and retrieval: onboarding flows per country or risk band, questionnaires, verification links, applicant status lookups. An agent can sequence a workflow that combines document, address, biometric, and database checks. The verification platform still performs the checks.
Rejecting a customer and filing a suspicious activity report are different in kind. Both require human approval, no matter how smoothly the rest of the workflow runs.
Building an agent-operated compliance setup
Two things must be in place before any of this runs: access to the relevant policies and permission to inspect or modify platform settings. The sequence then looks like this:
- Supply the source material, whether a regulation, a risk assessment, or an internal policy
- Extract the obligations, risk factors, scoring rules, and conditional requirements, with ambiguities flagged for a person rather than resolved by the agent
- Map each requirement to something the platform actually has: a verification level, questionnaire, risk rule, monitoring control, or workflow
- Produce a configuration plan for the compliance team to review before building anything
- Build the proposed configuration in a sandbox or test environment
- Compare the built configuration against the source policy, then test it with representative cases and known exceptions
- Have an authorized user approve deployment, then record changes and monitor outcomes
Mapping agent skills to compliance tasks
Each AI agent skill should correspond to a defined compliance task and guide how the agent should approach it. A policy-analysis skill might instruct the agent to read a document and produce a configuration plan. A configuration skill could guide the agent through creating verification levels in a sandbox, while a case-review skill could outline how to retrieve and analyze applicant data.
This distinction is important for AI agents in financial services, where accessing customer data, preparing draft settings, and changing live controls can create significant risk exposure. Skills themselves should not be treated as security or permission boundaries. Instead, the connected platform must enforce what the agent can read or modify, with changes to live verification and monitoring settings restricted through appropriate access controls and authorization mechanisms.
AI agents in banking and financial services
Plausible banking uses include gathering material for periodic KYC reviews, retrieving the transactions that triggered a monitoring rule, and drafting regulatory reports for a human to check.
Fragmentation is the practical obstacle. Identity records sit in a KYC platform, transactions in a payments or core banking system, alerts in a monitoring tool, and investigation notes in case management software. The 2026 Cambridge report found that 46% of surveyed AI vendors ran into legacy or siloed systems at client sites, and 41% reported data-sharing restrictions.
Each connection needs its own answer to two questions: what the agent may read, and what it may change. A periodic review is read-heavy. An agent can collect verification status, recent transactions, prior alerts, and past review decisions, then hand the assembled file to an analyst.
AI risks, governance, and human oversight
An agent may misread a policy, invent a requirement, or select the wrong tool. Malicious instructions hidden in a document or connected data source could manipulate its behavior, and excessive permissions may allow a small error to alter live configurations or customer records.
Reviewers may also place too much trust in an agent’s output, while incomplete records can make errors difficult to investigate. Model updates may change agent behavior, and dependence on external providers can create operational risks.
To mitigate the risks, effective AI governance and compliance controls should include:
- access limited to the information and actions required for each task, with separate permissions to read information, create drafts, and change live settings
- approved tools and data sources
- testing in a sandbox or test environment before changes are applied to live systems
- human approval for actions affecting customers, risk thresholds, or regulatory filings, as well as cases where requirements are unclear
- records of the instructions given, sources consulted, tools used, and changes made
- a contingency process if the agent or an external provider becomes unavailable
