Financial institutions have largely been on the cutting edge when it comes to experimenting with and investing in digital technologies and automation—think analytics, cloud computing, and the like. The latest technology on the stage, agentic AI, holds even greater promise for banks’ ability to realize scale efficiencies and simplify complex tasks and process steps.
Agentic AI is a type of AI that can perform a series of tasks independently and end to end across complicated workflows. It marks the latest in automation tools but provides a new wrinkle: While banks have been able to automate discrete processes in the past, some tasks have still been manually driven, with physical handoffs required. Other challenges—limits on system compatibility, issues with data quality, and regulatory constraints—have also prevented full end-to-end automation.
Those challenges remain and may even increase, as regulators closely monitor banks’ integration of AI models to ensure that they’re safely built, clearly explained to key stakeholders, and rigorously documented and controlled.
Now, however, activities once deemed too onerous to automate are getting a second look because of the abilities of AI-driven agents, or groups of agents. They can take over tedious, manual processes; analyze large, unstructured data sets; and produce meaningful insights that remain in line with regulations.
AI-driven agents can check data for quality and consistency and can conduct routine policy checks. Instead of targeting small efficiency gains, banks can now build and assign “multiagentic squads” to facilitate full workflows associated with credit reviews—for instance, conducting financial-risk analyses and business-model-risk analyses at scale. Credit reviews that previously took days could be completed in near real time. Customers could benefit from faster, clearer decisions about their loans and other transactions. And internal teams could gain the ability to make meaningful comparisons across data sets, products, and processes to generate deeper business insights.
“For me, that’s the decisive change—how quickly and efficiently we can now handle processes that used to be too complex to automate using agentic AI,” says Marcus Chromik, chief risk officer (CRO) at Deutsche Bank, one of many global financial institutions that are safely experimenting with agentic AI. “It’s early days, but we’re already seeing significant benefits from the scalability and flexibility that our agentic systems provide” (see sidebar, “AI transformation at Deutsche Bank: A case study”).
AI transformation at Deutsche Bank: A case study
Deutsche Bank Chief Risk Officer (CRO) Marcus Chromik explains to McKinsey how the bank is thinking about agentic AI, the near-term impact from its digital transformation, and longer-term expectations and opportunities from its use of agentic AI. The following has been edited for length and clarity.
McKinsey: Which areas of corporate underwriting could benefit most from agentic AI?
Marcus Chromik: Our experience suggests that agentic AI can support basically any area of credit review that requires drafting, as this task typically involves a core set of data, is repeated often, and accounts for a significant share of effort in the credit process. Among the first applications of agentic AI that we thought of were in the areas of financial analysis and assessment of business model risk.
McKinsey: How has agentic AI helped simplify those processes at Deutsche Bank?
Marcus Chromik: I can think of three ways. The first is speed: Clients no longer wait weeks for a decision; they get a timely and reliable answer. The second is focus: In helping to draft analyses, AI agents serve as a key support for credit officers, which allows officers to be more productive; they can spend their time in those places where their judgment really matters. The third is continuity: Use of agentic AI allows us to increase the consistency of outputs, and because controls are embedded within the drafting process itself, we’re getting early signs of any issues.
The agentic AI system also allows us to capture insights from experienced credit officers across the organization and build up our institutional knowledge. But what makes this even more important is that it’s not just about one tool. It shows us how we can scale gen AI in a responsible and practical way, bringing together technology, governance, and frontline adoption. We can better understand both the opportunities and the limitations—what works, and what doesn’t, and where we need workarounds. Having this clarity is important from a risk management perspective. That’s why this is a personal priority for me. AI improves client service, strengthens our resilience, and sets the direction for the risk function and the bank as a whole.
McKinsey: What advice do you have for others trying to automate workflows?
Marcus Chromik: I have three pieces of guidance. First, leaders in the C-suite must prioritize it and devote the required attention to it—that’s the only way such a transformation will succeed. Second, leaders must expose the whole organization to agentic AI and upskill everyone. In this way, they will be able to see not just all the opportunities in technology but also the possible flaws and failures. Keeping humans in the loop and actively learning from technology—these are very important for safeguarding responsible risk management and building trust with regulators and clients. And finally, from the outset, the transformation should be owned full time by a multidisciplinary group, including data scientists, engineers, user experience designers, project managers, and credit experts. In this way, leaders can create strong buy-in for organization-wide change and usage.
Marcus Chromik is the chief risk officer of Deutsche Bank. This interview was conducted by leaders in McKinsey’s Risk & Resilience Practice.
Comments and opinions expressed by interviewees are their own and do not represent or reflect the opinions, policies, or positions of McKinsey & Company or have its endorsement.
As with other technologies, agentic AI presents a learning curve for most companies. In McKinsey’s 2025 Global Survey on the state of AI, nearly two-thirds of respondents say their organizations haven’t yet begun scaling AI across the enterprise.1
In our work with CROs on a range of digital transformations, we’ve identified several core factors that can help banks successfully navigate that curve. Chief among these is a focus on improving and scaling those areas of the company that have shown the greatest impact—that is, creating a minimally viable solution, learning from that implementation, and ultimately replicating the solution’s success in other areas of the company. Other must-haves include a cross-functional design and implementation team, a strong data set, and governance processes that allow for safe experimentation and clear decision-making.
In this article, we’ll explain how leaders can build flexible, scalable, multiagentic systems that can help them conduct seamless credit review processes. The effort is worth it: In our experience guiding financial institutions through AI transformations, the use of multiagentic AI can create between a 40 and 80 percent productivity uplift per use case, greater consistency of outputs, and increased automation of control coverage.
The agentic AI opportunity
CROs and other senior leaders can use agentic AI to transform financial institutions into true digital organizations led by AI agents but involving a human in the loop at the most critical risk and control points. Some bank CEOs and CROs have begun to deploy AI in creative ways across the credit value chain—for instance, using it to match customers to the right products and to guide them virtually through loan applications. But the biggest benefit may be in redesigning complex manual processes to speed up risk assessments and increase the consistency and auditability of credit reviews end to end.
One financial institution is using a multiagent system, which comprises a collection of AI agents that coordinate with one another, to draft financial-risk assessments for corporate clients. Typically, credit officers would need to spend significant time gathering, examining, and validating all relevant sources of data to try to understand the critical financial-risk drivers and then shape these pieces of information into a structured narrative.
Instead, the multiagent system can help draft financial-risk analyses for the credit officer. It can signal to the credit officer where further focus may be warranted—for instance, where information may still be missing and where adherence to policy guidelines hasn’t yet been achieved. However, it’s still incumbent on the credit officer to review and edit the draft where needed before it can be approved as final.
Through its agentic AI system, the company has reduced by 50 percent the amount of time required to perform financial-risk analyses. What’s more, use of the tool has helped standardize practices across teams and capture information that can serve as a golden source of truth for employees as they transition in and out of teams.
Agentic AI transformation
As mentioned previously, there are a number of success factors required for an agentic system to come together, including a focus on the right use cases, data, team composition, and controls.
Targeting the most effective use cases
Financial institutions should kick off an agentic AI transformation by identifying those domains within the risk function where AI implementation would yield the biggest benefits. That would likely include categories such as the credit review process, reporting (including risk and control self-assessments), and model risk management. And then, within each domain or subdomain, CROs and other senior leaders should prioritize various agentic AI use cases and roll out pilots accordingly.
Relevant agentic AI use cases in the credit review process, for instance, could include financial-risk analysis, business-model-risk analysis, and industry risk analysis—really, any analyses required to reach a credit decision for corporate customers that are currently being conducted manually. As leaders set priorities, they should also actively consider opportunities to scale various use cases.
Building the tech foundation and a minimal viable product
Once priorities have been established, the CRO and risk team can develop a minimum viable product (MVP) for agentic AI aimed at addressing those priorities. The MVP will need access to relevant internal and external data sources and a representation of complex business processes—or a breakdown of tasks and subtasks associated with critical processes. The MVP will also need to be linked to a front-end system or workflow engine so that it can be fully embedded in business processes.
For instance, the financial institution previously cited prioritized building out a multiagent system that could automate draft analyses of P&L performance, cash flow, balance sheets, and other important financial metrics and then put these metrics into perspective versus peers. The financial institution built its agentic system on top of its existing AI infrastructure and AI tools—an advantage that not every organization has, of course, but one that allowed for faster development.
The goal here, as the design and development stages come together, should be to reuse components and draw from the lessons learned during rollout to build a template for the creation of other applications across regions, client portfolios, and use cases. In our experience, an MVP can deliver efficiency increases of 40 percent initially and up to 80 percent in the target state once the complete solution is in place.
Convening a cross-functional team
Developing an agentic MVP or system can’t be treated as an isolated IT project. Financial institutions should establish a cross-functional design and integration team to work on this initiative, pulling in representatives from both IT and business functions. Such a team should include business analysts who can translate business needs into technical requirements and vice versa. It should also incorporate feedback from first- and second-line employees—risk managers and credit officers—who can ensure that the new tool or system reflects their needs and ways of working.
This cross-functional team will set the parameters for what should be included in the MVP (and what shouldn’t) and identify measures of quality, uptime, and other performance factors for the MVP. Its members should meet regularly (at least several times a week) to review plans and outputs so that they can change courses as needed.
Perhaps most important, senior leadership will need to provide clear sponsorship for the use of agentic AI across credit review processes and hold everyone accountable for delivering on targets. The banks that are succeeding with agentic AI tend to boast hands-on senior leaders who can help developers break through talent, technical, and governance roadblocks and launch solutions quickly. In these companies, C-suite leaders insist on defining clear roles and responsibilities among those associated with the agentic AI transformation—assigning, for instance, a large-language-model owner for each business unit and an agentic-system lead in the IT group.
Ensuring that high-quality outputs are in line with policies and standards
For most agentic AI applications, the workflow will replicate existing practices and policies. But the introduction of AI agents offers an opportunity to not just automate but improve processes and outputs: A layer of agents will ingest annual reports, quarterly filings, and external data; calculate crucial metrics; and generate draft narratives. Another layer of subagents, focused on risks and controls, will monitor each step—checking data completeness, logical consistency, and compliance with policies. Yet another layer of synthesis agents will compile the findings into coherent drafts. Credit officers can then access the findings from within the core platform and validate and finalize the outputs.
The critical point here is that CROs need not wait until they have perfect data to keep agentic pilots moving. The agentic AI MVP itself can help leaders correct for less-than-optimal data and still create accurate outputs.
Setting governance and controls
The output from an AI agent or agentic squad needs to be in line with the bank’s policies, and the data being used must adhere to industry and company standards. When building agentic tools or systems, financial institutions should take care to embed appropriate governance and controls within them. This might include building a critic agent within the agentic AI squad to serve as a monitor and provide proactive feedback on data quality and outputs. Or teams might develop a control agent within the agentic AI squad to compare activities against the company’s policy standards and ensure that all actions are being documented and comply with broader regulatory requirements. Of course, there should always be someone to supervise—a human in the loop.
Safeguarding the scalability of agentic AI
Building an MVP in one domain isn’t enough; the true worth of agentic AI for a financial institution comes from its ability to apply AI across multiple use cases, teams, product portfolios, countries, and so on. Indeed, as CROs and risk teams develop individual agentic tools and systems, they will need to be mindful of future opportunities, asking themselves if they will be able to scale this application from one use case to a similar use case and to apply it to different portfolios or regions.
Of course, not every credit-related task will require the same level of automation. Simpler, lower-risk processes can be handled by agentic systems, allowing for instant credit decisions. Moderately complicated cases can be partially automated, where AI prepares the analysis, and credit officers intervene selectively. And the most complex, highest-risk processes will remain mostly manual. But even here, agentic AI can help eliminate repetitive tasks and ensure that decision-makers have access to the most accurate information.
AI transformations take time. For the foreseeable future, banks will still need to retain some human oversight over credit reviews, given the regulatory stakes involved. However, there’s a significant role for agentic AI in the credit risk process: an opportunity to realize both efficiencies and competitive advantage. Leaders can redefine the process from a manual, document-heavy exercise into a continuous, insight-driven workflow. Banks can boost their risk-adjusted returns and capital efficiency. Financial institutions can strengthen their governance and transparency and unlock value from end-to-end automation—now and for years to come.
