Compliance AI Adds Work for More Teams Than It Saves

StarCompliance’s 2026 benchmark found 26% of compliance teams got more work from AI, against 23% who cut manual effort. That split sat at the centre of the first webinar in the firm’s three-part Global Compliance Benchmark Study series, with StarCompliance chief product officer Kelvin Dickenson, NatWest strategic compliance lead Gabby Byrne and Raymond James Financial chief compliance officer Rich Konefal.

The tools can compress hours of gathering into minutes. The calendar does not stay empty.

26% Got More Work, 23% Got Less

On July 14, 2026, StarCompliance, the Rockville, Maryland employee and firm compliance technology provider, released the inaugural study after surveying more than 300 compliance professionals, plus risk and technology staff, worldwide. The function looks busy on paper. 76% said compliance budgets had gone up. 67% said they were already deploying AI or actively piloting it. Another 76% said they now operate across more than one jurisdiction.

Those figures sit next to a smaller, sharper pair. Only 23% of respondents saw reduced manual effort from AI. 26% reported increased workloads instead. That is not a rounding error on a feel-good rollout. It is more people saying the work got heavier than saying it got lighter.

WHAT THE 2026 BENCHMARK FOUND

Chief executive Jennifer Sun tied the spend to a wider brief than policing old rules. New asset classes and new trading models are landing on the same desks that are still stitching together surveillance, conduct checks and evidence for examiners.

AI adoption is accelerating, regulatory expectations continue to evolve, and new risk areas, including digital assets and emerging trading models, are reshaping the compliance landscape.

Jennifer Sun, Chief Executive Officer, StarCompliance

The study also flagged the plumbing. Fragmented systems, disconnected workflows and siloed data still block firms from scaling surveillance, running AI in a way they can defend, and producing consistent, regulator-ready evidence. Money is going in. The files still do not talk to each other.

Speed Fills the Calendar It Was Meant to Clear

The panel’s account of the 26% is mechanical, not mysterious. Faster reporting creates demand for more reporting. Quicker insights raise the speed people expect. When routine tasks leave the queue, staff inherit a heavier load of complex, judgment-heavy cases. The hours do not vanish. They change shape.

Make the gathering step ten times faster and the whole case does not finish ten times sooner. The bottleneck moves to the person who still has to approve the call, check the model, and put a name on the file. Output rises. The number of people qualified to verify it does not. That is how a time-saver becomes extra work rather than spare time.

StarCompliance put that tension in public on September 3, 2026, asking whether faster also means better, and pointing readers to a recap on putting AI to work in compliance. The post framed six lessons on governance, efficiency and human judgment, and repeated the study’s own advice to start with a jammed process rather than a model looking for a home.

AI can make compliance faster. But does it make it better?

Explore 6 lessons on AI governance, efficiency, human judgment and why successful adoption starts with the problem—not the technology.

Read the blog + watch on demand: https://t.co/yyyHycn22q

— StarCompliance (@StarCompliance) September 3, 2026

Access is no longer the argument inside large firms. Staff can already reach the tools. The live question is who is allowed to use them, on which data, with which logs, and with which person still on the hook when the answer is wrong.

What Happens When a Market-Abuse Alert Takes Minutes?

A market-abuse alert that once took hours of gathering trades and context can now be assembled in minutes, the panel said. Analysts then spend the time on the risk call. Quicker files also raise how many files get requested, so leftover work is the judgment the model cannot own.

That specimen is the job, not a side demo. Someone still has to decide whether the pattern is abuse, a coincidence, or a data mess. Someone still has to keep material non-public information and customer data out of a prompt that was never meant to see it. The model will draft the pack. It will not sit in front of the regulator.

Literacy, in the panel’s telling, is not a slide about “being curious.” Staff have to write better prompts, feed the right context, catch bad output, and know when a human has to take the wheel. Without that, a controlled environment is just a faster way to be confidently wrong.

Surveillance has to follow the new channel, the same way it followed email and chat. If the work is happening in an AI window, the monitor has to see that window. Extending existing watch-lists and information barriers into those sessions is unglamorous. It is also the difference between a pilot and a program a CCO will sign.

NatWest’s 60,000-Person Ethics Course Is the Real Cost

Byrne’s employer is already paying the training bill that the 26% implies. In June 2026 NatWest Group said it was rolling an AI and data ethics accreditation for 60,000 colleagues, after a custom course with the University of Edinburgh that first reached about 90 staff across roles and business lines.

HOW NATWEST IS TRAINING THE FLOOR

• Headcount: The bank is extending the programme to 60,000 colleagues across the group.
• First cohort: About 90 staff took the original University of Edinburgh course before the wider rollout.
• Course load: Eight e-learning modules sit against NatWest’s AI ethics principles, plus a half-day session on live cases.
• Calendar: NatWest said the first modules launched in June 2026, with full rollout continuing through October 2026, and that completion should take two to three months.

That is not a lunch-and-learn. It is what “AI literacy” looks like when a bank decides every person who might paste a customer file into a model needs a shared rulebook. Bias in decisions, the way training data bends outcomes, and when to stop and ask are on the syllabus because those are the errors that become conduct issues.

Konefal’s seat on the same panel is a reminder that broker-dealers are in the same bind. Raymond James Financial already has a chief compliance officer in the conversation while the industry is still arguing about access protocols, monitoring and training. The firms sending speakers are not waiting for a perfect framework. They are trying to put a fence around tools their people already have.

Watching AI Channels the Way Firms Watch Email

The panel’s governance list was familiar on purpose: controlled environments, access rules, monitoring, training, and stretching current surveillance over AI channels. The point was that old duties still attach. Material non-public information does not stop being MNPI because a model summarised it. Customer data does not become fair game because the prompt box is inside the firewall.

FINRA has been writing the same warning in a different dialect. On January 27, 2026, Greg Ruppert, executive vice president and chief regulatory operations officer, published observations on AI agents at member firms, building on the 2026 Annual Regulatory Oversight Report. He defined agents as systems that can plan, decide and act without the old rules-based script, and said they lack the tacit knowledge and predictability that ordinary supervision assumes.

FINRA’S AGENT RISKS FOR MEMBER FIRMS

• Autonomy: Agents may act without human validation and approval.
• Scope and authority: They may go past the user’s actual or intended mandate.
• Audit trail: Multi-step reasoning can be hard to trace, which complicates exams.
• Sensitive data: Agents may store, explore, disclose or misuse information they were allowed to touch.
• Domain gaps: General-purpose agents can miss industry-specific facts and still sound sure.

Ruppert also listed misaligned rewards and the usual generative-AI faults, bias, hallucinations and privacy, as still in force for agents. FINRA’s rules stay technology-neutral, he said: the same securities laws apply when a firm uses these tools as when it uses anything else. Human-in-the-loop checks, access logs, action tracking and hard limits on what an agent may do are the practical translation.

For a compliance desk, that maps onto the webinar almost one-for-one. If an agent can pull trading history, it can also pull a restricted deal file. If it can draft an investigation memo, it can also ship that memo to the wrong mailbox. Watching the channel is not a future project. It is how you keep the minutes you saved from becoming an exam finding.

Build Versus Buy Splits a Script From a System

AI has made it easier to stand up an internal helper for a narrow chore. A team can automate a single workflow, paste a better prompt on top of its own data, and look productive by Friday. The panellists drew a line between that and building compliance technology that has to run every day, at scale, with maintenance, integrations and regulatory upkeep attached.

The first kind of build is a script. The second is a product you will still be explaining in three years, when the model has changed, the vendor has changed, and the examiner wants the log. Firms that treat those as the same decision will spend twice: once on the clever internal tool, and again on the system that can actually produce defensible evidence.

Quality still tracks input. Garbage context in, polished garbage out, only faster. The study’s own counsel was to pick a repetitive process or a workflow drowned in information-gathering, test AI against that pain, and grow from there. That is a sequencing choice, not a vision statement. It also assumes someone owns the data. In a bank, four systems can each hold a “customer record,” and the model has to pick one. Committees form around that mess. Deleting the committee does not make the record true.

Buyers still have to ask who keeps the model current when a rule changes, who watches drift, and who can reconstruct a decision six months later. Those questions are dull. They are the difference between a demo and a control.

September 17 Turns to Crypto and Prediction Markets

The August 13, 2026 session, timed at 11:00 AM EST and 4:00 PM BST, was the AI chapter of a longer series. Session two lands on September 17, 2026, on digital assets and prediction markets. Session three is October 22, 2026, on information barriers and MNPI in a more connected firm. The study already said those later topics are weak spots: 40% of respondents felt least prepared for digital assets and crypto, and 34% still had no policy on prediction markets or had not considered them.

THE THREE BENCHMARK WEBINARS

1. August 13, 2026: AI in compliance, from potential to operational impact, with Dickenson, Byrne and Konefal.
2. September 17, 2026: Digital assets and prediction markets, and how employee compliance is supposed to cover them.
3. October 22, 2026: Information barriers and MNPI governance as data moves more freely across desks and jurisdictions.

Regulators are not waiting for that series to finish. The FCA’s landmark review into AI in retail finance, the Mills Review published on July 6, 2026, cited a 2026 survey finding 81% of firms were adopting AI at some level, with 40% at more advanced stages of scaling or transformation. Consumer research for the same review found 20% of people, equal to 11 million UK adults, would likely use AI that can act on its own inside pre-set goals. Sheldon Mills, the FCA executive director who led the work, said AI will transform financial services by 2030. Compliance teams are already living the first version of that, in the alert queue, not in 2030.

The 23% who cut manual effort are not imaginary. A pack that used to take an afternoon can land before lunch, and a good analyst will use that. The 26% are not imaginary either. They are what happens when the afternoon fills with harder cases, extra reports, and a model that still needs a person to say yes. The function bought AI to keep up. It now has to police the same tools, train tens of thousands of users, and show an examiner the log. That is the work that replaced the work that was supposed to disappear.