Every year I get asked some version of the same question: how is the industry changing? I’m glad to say that in recent years the number of answers to this question have been growing. Even so, 2026 has been a big year. Between conversations with banks, projects I’ve worked on, and what I’m seeing across the industry, this has been the year agentic AI stopped being a slide in a deck and started being something compliance teams actually run; the year the goalposts for what ‘success’ even means started to move. Here are the shifts that stood out to me most.

The Investigator's Job is being Rewritten

I’ve spent a lot of time this year thinking about how the skills essential for the role of AML investigator are shifting in the age of Agentic AI. A few changes I keep coming back to:

• An Increased focus on risk evaluation rather than data collection and box-ticking.

• The three-tiered investigation process will start to disappear. L1 and L2 will increasingly be handled by AI, with humans reserved for the riskiest, most complex cases.

• Agent management is becoming more of a priority; spot-checking decisions, catching where the reasoning breaks down, and feeding that back before the same mistake repeats at scale.

I find the last change the most interesting. It’s exactly the kind of conversation I want to keep having with peers: how else does the investigator’s role change from here, and how comfortable is our industry, really, with handing more of the risk decision itself to AI?

Compliance Becomes a Competitive Advantage

I’ve spoken with banks this year who are looking to expand into new markets and customer segments but are being held back because their investigations can’t scale fast enough to support that growth. What’s changed is that I’ve also watched Agentic AI let teams ramp up productivity to absorb the alert spikes that come with that growth. When a compliance function can enable the business to enter new markets safely while still managing financial crime risk, FCC transforms from cost center to a genuine competitive advantage. That reframing, from blocker to enabler, has been one of the most consistent threads in my conversations with customers this year.

The Definition of ROI is Shifting

Cost reduction used to be the gold standard for measuring the success of an AI project. Working with customers this year, I’ve noticed that standard start to shift. A more common goal now is cost avoidance — gaining capacity to do more with existing resources. The ability to enable business expansion or keep pace with new regulatory expectations is increasingly seen as more valuable than cost-cutting alone. If you’re still building your business case for AI purely on headcount savings, I think you’re measuring the wrong thing in 2026.

MRM is Getting a Seat at the Table Earlier

Banks are looping Model Risk Management in earlier, and the difference is immediate: clearer requirements, fewer delays, and a smoother path to production. The approach enables strong documentation and transparency to set the tone from the start, and partners who genuinely understand MRM help teams move faster with less friction. Early alignment keeps AI projects steady and gives institutions the confidence to scale, which matters more than ever now that governance is more of a structural requirement. The EU AI Act’s sandbox provisions and a wave of internal AI governance committees have made the same point from the regulatory side this year. AI in FCC has to be treated as a regulated system, with board-level attention, not just a tool bolted on and forgotten.

Data: The Unglamorous but Foundational Trend

None of the above works without data you can trust. Ask an AI model to draw a watch and it’ll draw the hands at 10:10, because that’s what nearly every watch ad has ever taught it to show. Financial crime models trained on skewed or limited data make the same kind of mistake, just with higher stakes: over-flagging low-risk activity while missing the typologies that don’t look like the training set. It shows up clearly in global name screening, where a single name can appear a dozen legitimate ways once you account for transliteration and cultural naming conventions. This is where rule-based systems either drown in false positives or miss the match entirely. The banks making real progress this year are the ones treating data quality and linguistic coverage as core infrastructure, rather than an edge case.

An Unlearned Belief

Early in my career, a data scientist newly assigned to financial crime once told me he didn't need me to tell him what to look for, just give him the data, and the patterns would surface themselves. I dismissed it outright. I'd spent years learning what risk looked like, and I wasn't ready to believe a model with no domain experience could find something I'd missed. Two years later, it had, repeatedly, and I've spent most of my career since unlearning that instinct.

This is the instinct that Agentic AI is asking the industry to unlearn in 2026. Holding these systems at arms length until they’re proven beyond doubt is the responsible choice, but the lesson I keep coming back to is that this technology isn't just faster and more scalable. It can surface blind spots and uncover patterns that experience alone would never have found.

Patrick Kirwin is Head of Product Management at Silent Eight. He will be speaking about agentic AI governance at the ICA FinCrime & Banking Forum in London this October 8th at 11:00am GMT.