Key takeaways

‍

• KYC (Know Your Customer) reviews and client onboarding are the strongest first process for AI in banking: costly, rules-based, spread across many systems, and ending in a human risk decision.

• Banks have adopted AI tools widely, yet most KYC review work is still manual, so adoption on its own hasn't changed how the work gets done.

• Choose the steps to automate from measured, step-level data on how the work runs, since time spent in a tool can't show which steps are ready.

• Agents gather data, screen names and draft the memo, while compliance officers keep risk decisions, approvals and oversight.

‍

AI in Banking: Why Operations Teams Should Start With KYC

‍

The results of AI in banking on overall output is a mixed bag.

The ECB's annual data on large European banks shows more than 85% already use AI in some form, yet PwC's 2026 Survey shows customer due diligence (CDD) as the single largest operational bottleneck. Most institutions still lack AI tools reliable enough to check whether a client's documents are genuine.

Banks bought the tools, but the review work remains mostly manual.

‍

The step forward is agentic AI in banking. These agents carry out multi-step work, such as pulling a client’s records from several systems and drafting the file, instead of only answering questions. It’s agentic process automation applied to the back office, automating cases where one review passes through many systems and many hands.

‍

This guide covers where to start, why KYC (Know Your Customer) is a sensible first choice for automation, and how to efficiently implement KYC and other steps to save time while mitigating risk.

‍

What is the first banking process to replace with an AI agent?

‍

Start with KYC reviews and client onboarding. Identity verification and risk-assessment are both costly and rules-heavy, with much of the time going into gathering and re-keying the same client data across systems. An AI agent can take these over while compliance staff retain ownership of the key risk decisions.

Choose which KYC steps to automate from measured data on how the work is actually done, rather than guessing based on the tools your team spends the most time in.

‍

KYC is the strongest first candidate for automation for four reasons:

‍

• It's expensive. Financial crime compliance, often heavy with manual work, can consume up to 5% of total banking costs according to BCG’s KYC appraisal. Every repeated lookup costs significant drain.

• It's rules-based. The rules are written down, so you can tell an agent what to follow and show an examiner how each file was checked.

• It involves many systems. Client data sits in several internal systems and on registry sites, which means analysts lose time navigating to and typing in details. An agent can take over those rote steps without touching a decision.

• It ends in a human decision. A compliance officer still makes the risk call, so accountability stays where regulators expect it.

‍

According to McKinsey’s look at financial crime across the industry, most financial institutions start their financial crime AI work in KYC or transaction monitoring, with banks commonly assigning 10 to 15% of their FTE to KYC and AML (anti-money laundering) alone.

‍

Two other processes share the pattern and make natural next candidates: payment exceptions and dispute handling.

Payment exceptions  run through a set of checks before a person releases or rejects the payment, while dispute handling pulls records together before a person rules on the case.

‍

Where AI in banking gets attention vs. where the cost sits

‍

Coverage of AI use cases in banking tends to start with easily seen customer-facing tools: a chatbot, a fraud alert, a personalized offer, etc. Since clients interact with those tools and they offer clear feedback, they draw most of the attention.

‍

Operation costs are less immediately visible. Onboarding is a trail of documents, periodic reviews refresh client details on a cycle, and exceptions get investigated one case at a time. Every handoff between front office, operations, and compliance means waiting for a resolution.

Picture a corporate client file moving from the relationship manager to operations to compliance and back. It waits at every stop, whether for a missing document, an answer from another team, or simply to be opened up. A single question from compliance can hold up the process for days. The result is a slow onboarding for the client, and immediate distrust.

That makes impactful banking automation a back-office story first. A chatbot is judged on the answers it gives, while a back-office process is judged on how long a file waits and how many times the same data gets entered. The former is a quick UX fix, the latter is a transformation.

‍

It also changes what you measure.  After running workforce analytics for financial services, you can see where operations time really goes. You can see also the clearest places to hand routine work to an agent.

‍

Why most banks pick the wrong first process

‍

Intelligent automation in banking starts with a distinction that’s easy to blur.

AI adoption is staff using AI tools in their daily work, which you could call Human + AI. An analyst asking a chat assistant to summarize a client’s registry filing is adoption.

AI implementation is handing an existing process to AI while people review and approve, or AI + Human. Here an agent assembles a client file from the registries and the bank’s own systems, and an analyst reviews it before it moves on.

‍

The data is clear on which is winning: tool use has spread wildly while most review work stays manual. AI adoption is running well ahead of implementation. Inputs that are easily visible on a usage dashboard look like a win, while the less visible review queue is easily ignored, and piles up.

‍

The common mistake is measuring how long people spend in a tool and calling it a day. Hours in a screening tool might be real review work, or a window being open while waiting for a decision. The number looks the same either way.

Hours in core banking might be account setup or data typed in again from an earlier step. Time in a tool can’t tell you which process is ready for an agent. Step-level data can, because it follows one client file through the specific activity and interactions surrounding each step to triage where the time went.

‍

What a KYC review really looks like, step by step

‍

Say a corporate client has agreed to open an account and the file reaches your KYC team. It typically goes through eight steps before the account is live:

‍

‍1.  Document requests. A KYC analyst emails the client for incorporation papers and ownership documents, then again for the items that don’t arrive.

‍

‍2. Ownership checks. The analyst works out who ultimately owns and controls the client. Any unclear items return to the client while the file waits.

‍‍

3. Data from internal systems and registries. The analyst looks the client up in the CRM, core banking, and public registries, then keys in the same details manually.

‍

4. Screening. The client and its owners are run through the screening tool, which returns a list of possible matches

‍‍

5. Clearing alerts. Each possible match is reviewed one by one, with many turning out as false positives.

‍

6. The risk memo. The findings and a proposed risk rating go into a memo, often rebuilt from one written for a similar client.

‍

7. Approvals. A compliance officer reviews the memo and approves or rejects the risk rating, while the file sits in the queue until they do.

‍‍

8. Account setup. Operations enters the client’s data into core banking, much of it captured earlier.

‍

Almost every step is a candidate for KYC automation, especially the repeated gathering and entering of client data. The cost of waiting on making that call is high. PwC's survey, which covered 531 financial institutions, had nearly a third expecting compliance costs to rise 10–30% over the next two years.

‍

Slow onboarding needs particular care. ‍

‍

‍Fenergo's 2025 survey found that 70% of firms lost clients in the past year due to inefficient onboarding, up from 67% in 2024 and 48% in 2023. In Celent's 2025 onboarding research, 72% of bankers said there was still more they could do to automate.

‍

Most banks know about the problem, they just don't have the data to fix it: in PwC's survey, up to 89% of respondents named data quality as the top barrier to AI adoption.

The practical question is which of these steps an agent can take on efficiently, and how far.

‍

The Process Savings Matrix for KYC and onboarding

‍

This matrix for banking operations automation sets two readings of KYC and onboarding side by side. The left side is what surface-level tool data shows: where the hours went, by application. The right side is what step-level process data shows: what happened inside each step.

‍

‍

The last two columns decide where an agent can help, and where your people stay in charge.

A high rating means the step is mostly work an agent can carry. A medium rating means the agent prepares the work while a person makes the call.

Three rows show how that plays out:

‍

• Request client documents: high. Clients are asked for the same documents again and again, so an agent can send and follow up on the requests while the relationship stays with your team.

• Screen names and clear alerts: medium. Many alerts are false positives, so an agent can triage and summarize them, and deciding which matches are real stays with an analyst.

• Write the risk memo: medium. Similar clients get similar memos, so an agent can draft one, and the compliance officer approves the risk rating.

‍

A rating describes the step in general. What a step actually saves depends on its volume and how often it goes wrong. Only measurement shows both.

‍‍

Measure the process before you automate it

‍

Measuring a process used to mean hiring consultants that spend two to three weeks alongside your staff writing the process down by hand. Steps still get missed, because a few weeks only covers the files that happen to move within that window, and one observer can follow only so many at once.

‍

A better starting point for AI in banking operations is to record the work itself at bottom level. Workflow Intelligence captures how the work actually happens across apps, step by step, for every team. For a KYC review, it gives full context on where files wait, which steps get redone, and why.

‍

The measurements are deterministic: they come from activity, interactions, and context captured from your own data, not from AI guesswork. In a regulated industry you need to show where every figure came from.

After measurement, opportunities are sorted into a prioritized list of recommendations, so person can review each recommendation before acting.

‍

Measured data changes decisions, and transforms businesses. A leading US bank uncovered $2.5 million in contractor savings by discovering it was being billed 25% more than the hours worked.

‍

Five measures show which KYC steps are ready for data-driven banking:

‍

• Time per step, which shows the slow steps.

• App switches, where a high count can mean the same details are being entered in more than one system.

• Handoffs between teams, which show where a file stalls as it passes from one team to the next.

• Wait time, which separates slow work from a file waiting on a client or an approver.

• Rework, which marks the steps that had to be repeated because the first pass fell short.

‍

Where the human stays in the loop

‍

AI agents in banking don’t make risk decisions. The agent does the repetitive groundwork: it gathers client data, re-keys it, screens names, triages alerts, and drafts the memo.
‍

The compliance officer makes the call and signs off. The audit trail, approvals, and model oversight stay human.

Expect three questions about any agent:

‍

• Who approves? Every risk rating has a named compliance officer behind it.

• What is logged? Every action the agent takes is recorded, so the file can be reconstructed afterward.

• How are exceptions handled? Anything outside the agent’s remit goes to a person.

‍

Automating away repetitive low-risk tasks frees analyst capacity for higher-risk cases. The time once spent re-entering client data or working through a long list of alerts goes to files that need judgment.

‍

Start with a 14-day audit

‍

A 14-day work audit captures exactly how your bank’s KYC and onboarding work runs across teams and systems to learn how each step can be best-served by an agent.  Clients doing the audit are often shocked at how accurately the data matches what their people are doing daily.
‍

For now the audit is offered only to eligible companies, who receive their first report in as little as 14 days. Request your 14-day audit now to begin your ROI journey.
‍

Frequently asked questions

‍

Which banking process should I automate with AI first?

KYC reviews and client onboarding. Both follow clear rules, draw client details from many systems and end in a human risk decision. An agent can do the gathering while compliance officers make the call, and measured data decides which steps go first.

‍

Will AI replace banking jobs?

AI takes over data gathering and repetitive steps, while risk decisions, approvals and client relationships stay with people. Analysts spend less time re-entering data and more on the cases that need judgment, so the work shifts toward that judgment and toward clients.

‍

Is AI replacing RPA in banking?

No. RPA in banking runs fixed, scripted steps, while AI agents take on the steps that vary and summarize what they find. The two work together: a scripted step moves the data, and an agent handles the cases that don’t fit the script.

‍

How is automation used in banking?

Automation in banking covers onboarding, KYC, payments, disputes and reporting. Each area has steps that repeat and rules that stay the same, which is where automation fits. KYC is the strongest place to start, because the same steps repeat for every client.

‍

How much does a KYC review cost?

Fenergo’s KYC in 2024 research estimates a review of a medium-risk corporate client at about $2,250, and BCG puts financial crime compliance at up to 5% of total banking costs. Your own cost depends on how your team does the work, so measure the steps before you automate any of them.

‍

Can an AI agent make KYC risk decisions?

No. The agent assembles the file and flags anything that needs a closer look, and a compliance officer decides. Its actions are logged so the decision and what led to it can be reviewed, and model oversight stays with people too.

‍

How do I measure a banking process before automating it?

Measure time per step, app switches, handoffs, wait time and rework. Take the numbers from real work instead of interviews, which describe how the process is supposed to run. Collect them across the whole team, so no single analyst’s habits shape the picture.

‍