Article
Payments
4 min read
Dondi BlackHead of Industry Advisory for Payments

In brief

  • AI governance must cover the full workflow, including outputs that shape decisions even when humans make the final call.
  • Effective human oversight requires access to evidence, sources and exceptions, plus authority to challenge or escalate AI recommendations.
  • Reusable controls and workflow redesign can help institutions scale AI while measuring decision quality, explainability and productivity.

AI governance is becoming a workflow engineering challenge.


Across financial services, AI is already helping organizations interpret policies, monitor activity, create documentation, analyze compliance requirements and explain decisions. It is also supporting employees as they make increasingly complex judgements. What began as a way to summarize information or draft content is moving into the workflows that shape financial outcomes and customer experiences.

That shift is happening across fraud investigations, merchant underwriting, disputes, servicing, compliance, payment exceptions and treasury operations. As AI takes on a larger role in these processes, governance has to move with it.

 

Recent research on generative AI in financial institutions highlights an important distinction: risk depends not only on whether a workflow uses an AI model, but also on how close the model’s output is to a business decision and what harm could follow if that output is wrong.

 

That is a useful lens for financial services leaders. A model that summarizes an internal document has a different risk profile from one that drafts adverse-action language, interprets a monitoring alert or recommends how to resolve a fraud investigation. The underlying technology may be similar. The workflow determines how the output is used, who relies on it and what happens next.

Three implications for financial institutions

 

  1. 1. Governance needs to follow the decision

    Traditional model governance has established disciplines for validation, monitoring, change management and oversight. Generative and agentic AI complicate where those disciplines should apply because AI can influence a decision without making the final calculation or taking the final action.

  2.  
  3. Consider merchant underwriting. AI might summarize business information, interpret policy, identify missing documentation and draft an analyst recommendation. A deterministic risk model may still produce the score, and a human underwriter may still approve or decline the application. Yet AI has shaped the information the underwriter sees and may have influenced the explanation behind the decision.

  4.  
  5. Governance should therefore cover the full decision workflow. Leaders need to understand where AI retrieves information, interprets policy, generates content, invokes tools and hands work back to employees. They also need to know what data and sources inform the output, how changes to the workflow are managed, and where responsibility sits when the output is incomplete or incorrect.

  6.  
  7. This broader view helps institutions see AI’s influence even when it does not appear in the final decision record. It also makes it easier to identify where existing controls apply, where they need to be extended and which decisions require additional review.

 

  1. 2. Human oversight has to be engineered

  2.  
  3. “Human in the loop” is often treated as if it were a complete control. In practice, the presence of a reviewer is only useful if that person has the information, time and authority to challenge the system.

  4.  
  5. The key question is: what is the person actually expected to review?

  6.  
  7. In a fraud investigation, an AI-enabled workflow could query transaction history, connect related entities, summarize alerts, build a timeline and draft a case narrative. These are natural applications for generative AI, but each step can affect what an investigator notices and how the case is understood.

  8.  
  9. A reviewer needs more than a polished final paragraph. They need visibility into the underlying evidence, the sources used, any exceptions or missing information, the system’s confidence and the policy interpretation behind its recommendations. They also need to see what actions have already been taken and know which steps they can reverse or escalate.

  10.  
  11. Effective oversight depends on deliberate workflow design: where judgement is required, what evidence is presented, which actions may proceed automatically and what conditions trigger escalation. When those boundaries are clear, human review can focus on meaningful decisions instead of checking AI output without enough context to assess it.

 

  1. 3. Governance can become an accelerator

  2. Strong governance should make responsible deployment easier to repeat. If every team has to create controls from scratch, AI adoption will be slow and uneven. Shared patterns for identity, access, data retrieval, monitoring, auditability, human review and escalation can give teams a consistent foundation.

  3.  
  4. Institutions can then adapt those patterns to the risk of each workflow. A document-summary tool and an AI-assisted fraud investigation may use some of the same safeguards, but they will not need identical approval points or escalation rules.

  5.  
  6. This is where the technology ecosystem matters. Cloud, data and model-monitoring platforms from providers such as AWS, Google, Microsoft, Databricks and Snowflake can supply components of a control environment. But the platforms alone cannot determine how those capabilities should work within a bank’s policies, risk tolerance, regulatory obligations and customer commitments. That requires engineering the controls into the workflows where AI is used.

Three priorities for the next six months

 

  1. 1. Map AI’s influence, not just the inventory

  2.  
  3. Identify where AI is already interpreting, recommending, drafting, monitoring or initiating actions inside critical workflows. Include tools embedded in existing platforms as well as systems developed specifically for AI use.

  4.  
  5. Then assess how close each output is to a customer, financial or compliance decision. This gives leaders a more useful picture of exposure and helps them prioritize where stronger controls or clearer ownership are needed.

 

  1. 2. Establish reusable control patterns


  2. Define common requirements for data access, grounding, traceability, monitoring, approvals, escalation and evidence retention. Make clear what must be consistent across the organisation and what teams can configure for a specific workflow.

  3.  
  4. These patterns are especially useful for agentic workflows, where work may cross systems, trigger actions and move through repeated hand-offs. Clear controls can help teams understand what an AI agent is authorised to do, when it must pause for review and how its actions can be reconstructed afterwards.

 

  1. 3. Redesign workflow priorities end-to-end

  2.  
  3. Choose a small number of high-value processes, such as fraud investigations, merchant underwriting, disputes or payment exceptions. Map how work happens today, identify where AI can help and design the controls alongside the new capabilities.

  4.  
  5. Evaluate more than productivity. Measure decision quality, explainability, exception rates and the level of human intervention alongside time saved or processing volume. Those measures help leaders understand whether AI is improving the workflow as a whole, not simply accelerating one step within it.

The opportunity ahead

 

Financial services has spent decades engineering control into payments, credit, risk and compliance. AI introduces a new layer of intelligence across those same processes, bringing opportunities to find patterns sooner, reduce manual effort and help employees make better-informed decisions.

 

The institutions that create lasting value will connect that capability to workflows designed to make decisions observable, explainable, controllable and accountable. That means treating governance as part of how work gets done: built into the flow of information, decisions, approvals and actions.

 

That’s why the next chapter of AI governance will require more than policy. It will require workflow engineering.

 

Read more about the latest shifts in financial services and payments—including AI-native operations—in our executive guide, Payments at the Operating Core.

FAQs

AI can shape a decision even when it does not make the final calculation or take the final action. Governance should account for how AI retrieves information, interprets policy, generates recommendations and hands work to employees.