On this page

A mid-size bank with slow, manual business onboarding has three practical moves: measure the process as it runs today, automate it step by step according to customer risk, and prove each gain with a metric an auditor can trace. This article sets out how to do that. It defines the KPIs, gives planning ranges for targets, provides an ROI model with a worked example, and lays out a 30/60/90-day plan for taking the case to internal stakeholders.

The approach is regulatory-first because that's where mid-size banks tend to stall. Compliance teams hesitate to automate when they can't show how an automated decision was reached. Operations teams hesitate to measure when the numbers look bad. A KPI framework that is both audit-ready and commercially meaningful resolves both objections.

Manual onboarding is a measurable defect, and three regulatory sources define what "acceptable" automation looks like

Three documents shape how a European bank can automate business onboarding. None sets a speed target. All of them make automation conditional on governance, which is why the KPIs below include control metrics next to speed metrics.

The EBA remote onboarding guidelines set governance expectations for digital identification

The European Banking Authority published its final "Guidelines on the use of remote customer onboarding solutions" (EBA/GL/2022/15) on 22 November 2022. The guidelines are technology-neutral. They expect institutions to assess a remote onboarding solution before rollout, monitor it on an ongoing basis, and keep governance and accountability internal even where a third party supplies the technology. The EBA publishes the guidelines as a PDF with an accompanying compliance table spreadsheet. The landing page itself is a download hub with little interpretation, so translating the text into operating metrics is left to the bank.

The practical consequence: "ongoing monitoring of the solution" is a KPI requirement in disguise. A bank that can't report discrepancy rates, failure rates, and manual override rates can't demonstrate it.

The AMLR moves customer due diligence toward a single EU rulebook

Regulation (EU) 2024/1624 (the AMLR) was adopted on 31 May 2024 and published in the Official Journal on 19 June 2024, as listed on EUR-Lex. It applies from 10 July 2027 and sets directly applicable customer due diligence requirements, including for beneficial ownership. Banks building automation now are building for those rules, so workflows should be policy-configurable instead of hard-coded to one national interpretation. Dutch banks, for example, currently work under the Wwft as supervised by DNB, using KVK data and the UBO register as inputs. The Wwft will give way to the directly applicable EU rules.

FATF frames digital identity as a risk-based CDD tool

FATF's "Guidance on Digital Identity" applies Recommendation 10 (customer due diligence) to digital ID systems used for identification and verification. Its central principle is proportionality: assurance levels and controls should reflect the risk of the customer and the product. That principle is the basis for the risk-tiered targets later in this article. A single onboarding SLA or STP target across all customers would contradict it.

Automation coverage needs a step-level definition before anything else

"We automated onboarding" isn't a measurable claim. Define coverage per step, then report it as a percentage of cases in which that step ran without a manual touch.

Step

What "automated" means

Coverage formula

Data capture

Customer enters data once; registry data prefilled

Cases with prefilled registry data / total cases

Document verification

Extraction and validity checks run without analyst input

Documents auto-verified / documents received

Ownership and UBO checks

Ownership structure resolved and reconciled to registers

Cases with UBO resolved automatically / total cases

Sanctions and PEP screening

Screening run and low-risk alerts dispositioned by rule

Alerts closed by policy / total alerts

Decisioning

Policy applied and outcome recorded without manual review

Cases decided by policy / total cases

Escalation

Exceptions routed to the right queue with context attached

Escalations routed automatically / total escalations

Report coverage per step and per risk tier. A bank that automates document checks for 90% of cases but resolves ownership structures manually will see its cycle time barely move, because ownership resolution is usually where the long tail sits.

The KPI set covers six areas, and each one maps to a regulatory or financial question

Each KPI below has a formula, so two analysts calculating it get the same number. Compute them per entity type (for example sole proprietorship, limited company, foundation) and per risk tier.

Time and throughput

  • Time-to-first-decision: timestamp of first system or analyst decision minus timestamp of complete submission. Use the median and the 90th percentile, not only the average.

  • Time-to-complete CDD: timestamp of account-ready status minus timestamp of application start.

  • SLA attainment: cases completed within SLA / total cases.

  • Bottleneck contribution: elapsed time per step / total elapsed time. Include waiting time for customer responses as its own step.

Analyst workload

  • Analyst minutes per case: logged handling time summed per case. Where time isn't logged, sample 50 cases per tier and time them.

  • Manual review rate: cases with at least one analyst touch / total cases.

  • Escalations per 100 cases.

  • Rework rate: cases sent back to the customer or re-opened because of missing or incorrect data / total cases.

Quality and compliance

  • Straight-through processing (STP) rate: cases approved or rejected by policy with no analyst touch / total cases. Report it by tier, since high-risk STP should sit near zero by design.

  • Screening false-positive rate: alerts closed as no-match after review / total alerts.

  • False-reject rate: applications rejected and later approved on appeal or reapplication / total rejections. This is the hardest KPI to measure and the most revealing.

  • Discrepancy rate: cases where submitted data conflicts with register data / total cases.

  • Evidence completeness: cases whose file contains every item required by policy / total cases audited.

Conversion and revenue

  • Completion rate: completed submissions / started applications.

  • Step-level drop-off: abandonments at step N / entries into step N.

  • Approval rate by risk segment.

  • Reboarding success rate: customers re-verified within deadline / customers due for re-verification.

Duna's own analysis of business onboarding finds that each follow-up request loses roughly 15% of submitted customers. That figure comes from Duna's published research on onboarding as a conversion problem, and your own baseline may differ. It does explain why rework rate belongs in the conversion section as well as the workload section: every avoidable round-trip is both an analyst cost and a drop-off event.

Control effectiveness

  • Decision traceability: decisions linked to a specific policy rule and version / total decisions.

  • Human-in-the-loop coverage: required human reviews completed / required human reviews. The target is 100%, and the metric exists to prove it.

  • Exception handling timeliness: exceptions resolved within their own SLA / total exceptions.

Data governance

  • Structured-field quality score: valid, complete, correctly formatted KYB fields / total fields.

  • Data acceptance rate: submissions passing validation on first attempt / total submissions.

  • Registry data error rate: prefilled or API-sourced values later corrected / total prefilled values.

Baseline first: ninety days of data decide which targets are realistic

Targets set without a baseline are guesses. Pull the last 90 days of closed cases and segment them by entity type and risk tier before setting any target.

  1. Export case-level timestamps from the case management or workflow system: application start, submission, each handoff, decision, account activation.

  2. Join analyst activity data. If the bank doesn't log handling time, run a two-week time-study sample of at least 50 cases per risk tier.

  3. Tag every rework event with its cause: missing document, incorrect data, unclear request, registry mismatch.

  4. Pull alert-level screening data and the disposition of each alert.

  5. Sample 30 to 50 completed files per tier and score them for evidence completeness against written policy.

Expect the exercise to surface surprises. Banks commonly discover that waiting on customers, not analyst work, dominates elapsed time, and that handling time varies several-fold between analysts on similar cases. Those findings point to different fixes: intake design for the first, policy codification for the second.

Target ranges by maturity level are planning assumptions, not industry benchmarks

No credible, neutral, bank-size-specific benchmark set exists for KYB metrics, and published vendor figures aren't comparable across institutions. The ranges below are planning assumptions for a mid-size bank. Treat them as hypotheses to validate against your measured baseline, then adjust.

KPI

Level 1: early automation (intake, document extraction, registry prefill)

Level 2: risk-based optimization (policy-driven decisioning, tiered routing)

Level 3: near-STP for low risk

Median time-to-decision reduction vs. baseline

25% to 40%

50% to 65%

70% or more

Analyst minutes per case reduction

20% to 35%

40% to 55%

60% or more

STP rate, low-risk tier

15% to 30%

40% to 60%

70% to 85%

Screening alert volume escalated to analysts

10% to 20% lower

30% to 50% lower

50% or more lower

Rework rate reduction

20% to 30%

40% to 50%

50% or more

SLA attainment

90%

95%

97% or more

Evidence completeness (audited sample)

95%

98%

100%

Evidence completeness is the one line that shouldn't be traded off. A bank can accept a slower cycle-time improvement. It can't accept an audit file that gets thinner as automation expands.

Targets must differ by risk tier because the regulators' own logic is risk-based

FATF's digital identity guidance ties assurance and controls to risk. The EBA guidelines likewise expect the solution's design and the scope of verification to reflect the ML/TF risk of the customer relationship. One blended target hides the problem: a 60% overall STP rate could mean 90% of low-risk cases and 0% of high-risk cases, or a worrying mixture.

Dimension

Low risk

Medium risk

High risk

Automation expectation

Decision by policy for clean cases

Automated data gathering and screening; analyst decides on flagged cases

Automation prepares the file; analyst and approver decide

STP target (Level 2 planning range)

40% to 60%

10% to 25%

Near zero by design

Human-in-the-loop

Sampled quality review

Required for any flag

Required, with four-eye approval

Time-to-decision priority

Minutes to hours

Same day to two days

Quality and depth over speed

Primary KPI emphasis

STP, completion rate

Escalation rate, false positives

Evidence completeness, traceability

Setting different goals per tier also gives compliance a reason to support the project. The efficiency gain is concentrated where risk is lowest, and analyst time is redirected toward the cases that carry real exposure.

The executive dashboard fits on one page, with a drill-down for compliance operations

Executives need six measures, trended monthly and split by risk tier: median time-to-decision, SLA attainment, cost-to-serve per onboarded customer, completion rate, analyst minutes per case, and exception rate.

Compliance operations needs the drill-down: step-level bottleneck contribution, escalation reasons, false-positive rate by screening list, rework causes, discrepancy rate by registry source, and the control view (traceability, human-in-the-loop coverage, exception timeliness). Add a regression panel that flags any control KPI falling below its threshold, so a speed gain never hides a quality loss.

The ROI model needs eight inputs and every output must map back to a KPI

Inputs

Input

Source KPI

Applications started per month

Completion rate denominator

Baseline and target completion rate

Completion rate

Completed cases per month

Derived

Loaded analyst cost per minute

Finance

Baseline and target analyst minutes per case

Analyst minutes per case

Approval rate

Approval rate by segment

First-year gross profit per approved business customer

Finance or product

Implementation cost (amortized) and annual run cost

Vendor and internal estimates

Formulas

  • Labor savings = future completed cases × (baseline minutes − target minutes) × cost per minute

  • Incremental approved customers = started applications × (target completion − baseline completion) × approval rate

  • Conversion value = incremental approved customers × first-year gross profit per customer

  • Annual cost = amortized implementation + annual run cost

  • Net annual benefit = labor savings + conversion value − annual cost

  • ROI = net annual benefit / annual cost

  • Payback (months) = upfront implementation cost / ((labor savings + conversion value − run cost) / 12)

Worked example with placeholder values

The figures below are illustrative planning assumptions in generic currency units (CU), not industry averages.

  • 1,000 applications started per month; completion rate rises from 62% to 70%. Completed cases move from 620 to 700 per month.

  • Analyst cost: 1 CU per minute (60 CU per hour, loaded). Minutes per case fall from 150 to 70.

  • Approval rate: 90%. First-year gross profit per approved customer: 600 CU.

  • Implementation: 450,000 CU amortized over three years (150,000 CU per year). Annual run cost: 220,000 CU.

Calculation:

  • Labor savings: 700 × 12 months × 80 minutes × 1 CU = 672,000 CU per year

  • Incremental approved customers: 1,000 × 12 × 8% × 90% = 864 per year

  • Conversion value: 864 × 600 = 518,400 CU

  • Annual cost: 150,000 + 220,000 = 370,000 CU

  • Net annual benefit: 672,000 + 518,400 − 370,000 = 820,400 CU

  • ROI: 820,400 / 370,000 ≈ 222%

  • Payback: 450,000 / ((1,190,400 − 220,000) / 12) ≈ 5.6 months at full run-rate

Run the sensitivity case that finance will ask for. With conversion value set to zero, net annual benefit is 302,000 CU, ROI is about 82%, and payback is roughly 12 months. The case holds on labor savings alone, and conversion is upside.

Two cautions apply. First, a bank that doesn't plan to redeploy or avoid hiring won't realize the labor line as cash, so state whether savings come from avoided hires, overtime, or capacity absorbed by growth. Second, ramp-up matters: model benefits at 25%, 60%, and 100% of target across the first three quarters.

Proof requirements

Every ROI line needs a KPI that evidences it after go-live. Labor savings map to analyst minutes per case and manual review rate. Conversion value maps to completion rate, step-level drop-off, and rework rate. Cost avoidance on false positives maps to alert escalation volume. Quality assurance maps to evidence completeness and discrepancy rate. If a benefit can't be mapped, take it out of the model.

Two illustrative vignettes show how the KPIs move in practice

These are template-style composites for planning purposes, not reports on specific institutions. For documented examples of lifecycle automation at scale, see Duna's perpetual KYB case studies, including the re-verification of 40,000 partners at bol.

Vignette one: intake and document verification

  • Bottleneck: analysts re-key registry data and check incorporation documents by hand. Rework runs high because customers receive multiple separate requests for missing items.

  • Automation steps: registry prefill, single-pass document request driven by entity type, automated extraction and validity checks, discrepancy flags where submitted data conflicts with the register.

  • KPI movements to expect: rework rate and analyst minutes per case fall first, followed by completion rate as follow-up requests drop. Discrepancy rate rises initially, because the system now detects conflicts that manual review missed. Don't read that as a quality regression.

  • Lessons: log which fields were prefilled and which were customer-supplied so the file shows the provenance of every data point. Keep a sampled analyst review of auto-verified documents until the error rate is stable.

Vignette two: risk-based decisioning and escalation

  • Bottleneck: every case goes through the same analyst queue regardless of risk, and screening alerts on low-risk customers consume disproportionate time.

  • Automation steps: policy rules assign a risk tier at intake, route cases by tier, close low-risk alerts that meet documented criteria, and send high-risk cases to a four-eye queue with the evidence pack attached.

  • KPI movements: STP rate in the low-risk tier and escalations per 100 cases move first. Time-to-decision for high-risk cases may barely change, which is the correct outcome.

  • Lessons: version every policy rule and record which version produced each decision. Approve rule changes through the same governance route as model changes, and back-test them on historical cases before activation.

The business case for stakeholders should address each function's actual objection

A single document can serve four audiences if its sections are separable.

  • C-suite and COO: cost-to-serve, time-to-revenue, and the conversion effect. Duna's analysis of why onboarding belongs on the COO agenda makes the same argument that onboarding drives conversion, cost, and productivity, not only compliance.

  • Compliance: alignment with the EBA remote onboarding guidelines, the FATF risk-based principle, and the AMLR timeline; the control KPIs; the human-in-the-loop design.

  • Operations: the baseline findings, bottleneck map, and expected change in analyst workload.

  • IT and security: integration scope, data residency, access controls, and the outsourcing and vendor-risk position.

Governance has to be explicit in the document

The EBA guidelines keep accountability with the institution, so the case should name an owner for each of these:

  • Approval workflow for policy rules and any AI or scoring component, including validation before release and periodic revalidation.

  • Human-in-the-loop design: which decisions are never automated, which are sampled, and what triggers mandatory review.

  • Vendor accountability: the bank's outsourcing assessment, exit plan, fallback procedures if a data provider fails, and audit rights.

  • Evidence retention: what is stored, for how long, and in what format, aligned to Wwft or equivalent national retention requirements.

Capability matters here because it determines whether governance is practical. A policy engine that expresses compliance logic as versioned rules, with an audit trail of interactions, decisions, and policy changes, gives the bank something to show a supervisor. Duna's platform is built on that structure, and its reported outcomes (10.6x faster onboarding, a 37% conversion increase, and 4.8x analyst efficiency) are the company's own customer figures, so validate any vendor claim against your baseline in a pilot. Banks running a formal selection can use the criteria in the bank-grade KYB RFP overview, and Duna's guide to AI in KYC and AML covers how to reduce manual work without losing control.

A 30/60/90-day plan gets from baseline to a scaled pilot

Days 1 to 30: baseline and process map

  • Complete the 90-day data pull and sampling described earlier.

  • Map the process by entity type and risk tier, with elapsed time per step.

  • Fix KPI definitions in writing and have compliance sign off on them.

  • Rank bottlenecks by contribution to elapsed time and analyst minutes.

Days 31 to 60: pilot design and build

  • Choose one entity type and one risk tier, typically low-risk limited companies, where volume is high and policy is clearest.

  • Codify the policy for that segment as explicit rules, including escalation triggers.

  • Configure intake, registry prefill, document checks, and screening disposition rules.

  • Set the dashboard up and test traceability on sample cases.

Days 61 to 90: run, measure, decide

  • Run the pilot in parallel with the manual process on a defined share of cases, or with 100% analyst review of automated decisions for the first weeks.

  • Report KPIs weekly against the baseline.

  • Present a scale-or-stop recommendation to the steering group.

Acceptance criteria combine improvement targets with no-regression ceilings

Category

Criterion (planning assumption)

Speed

Median time-to-decision down at least 25% in the pilot segment

Workload

Analyst minutes per case down at least 20%

Conversion

Completion rate up, or rework rate down at least 20%

Evidence

Audit-sample completeness at or above baseline, and at least 95%

Traceability

100% of automated decisions linked to a rule and version

Human review

100% of required human reviews completed

Discrepancies

Rate within a ceiling agreed with compliance (rising detection is acceptable if explained)

False rejects

No increase against baseline

If the pilot meets the improvement targets but breaks any no-regression line, it doesn't pass. That condition is what makes the project defensible to a supervisor.

The baseline decides where automation goes, and the control metrics decide whether it stays

Manual, slow business onboarding rarely has a single cause. Waiting on customers, re-keyed data, one-size-fits-all review queues, and alert noise each contribute, and the 90-day baseline shows how much. Automate those steps by risk tier, hold the control KPIs fixed, and report speed and conversion gains against the same definitions every month.

The EBA's remote onboarding guidelines, the AMLR, and FATF's digital identity guidance all point to the same operating model: automation that is risk-based, governed, and traceable. A mid-size bank that builds that model now, with KPIs it can show a supervisor, has also built the business case it needs. For a view of how different platform architectures fit banks of different sizes, see Duna's KYC and KYB orchestration platform guide.

A mid-size bank with slow, manual business onboarding has three practical moves: measure the process as it runs today, automate it step by step according to customer risk, and prove each gain with a metric an auditor can trace. This article sets out how to do that. It defines the KPIs, gives planning ranges for targets, provides an ROI model with a worked example, and lays out a 30/60/90-day plan for taking the case to internal stakeholders.

The approach is regulatory-first because that's where mid-size banks tend to stall. Compliance teams hesitate to automate when they can't show how an automated decision was reached. Operations teams hesitate to measure when the numbers look bad. A KPI framework that is both audit-ready and commercially meaningful resolves both objections.

Manual onboarding is a measurable defect, and three regulatory sources define what "acceptable" automation looks like

Three documents shape how a European bank can automate business onboarding. None sets a speed target. All of them make automation conditional on governance, which is why the KPIs below include control metrics next to speed metrics.

The EBA remote onboarding guidelines set governance expectations for digital identification

The European Banking Authority published its final "Guidelines on the use of remote customer onboarding solutions" (EBA/GL/2022/15) on 22 November 2022. The guidelines are technology-neutral. They expect institutions to assess a remote onboarding solution before rollout, monitor it on an ongoing basis, and keep governance and accountability internal even where a third party supplies the technology. The EBA publishes the guidelines as a PDF with an accompanying compliance table spreadsheet. The landing page itself is a download hub with little interpretation, so translating the text into operating metrics is left to the bank.

The practical consequence: "ongoing monitoring of the solution" is a KPI requirement in disguise. A bank that can't report discrepancy rates, failure rates, and manual override rates can't demonstrate it.

The AMLR moves customer due diligence toward a single EU rulebook

Regulation (EU) 2024/1624 (the AMLR) was adopted on 31 May 2024 and published in the Official Journal on 19 June 2024, as listed on EUR-Lex. It applies from 10 July 2027 and sets directly applicable customer due diligence requirements, including for beneficial ownership. Banks building automation now are building for those rules, so workflows should be policy-configurable instead of hard-coded to one national interpretation. Dutch banks, for example, currently work under the Wwft as supervised by DNB, using KVK data and the UBO register as inputs. The Wwft will give way to the directly applicable EU rules.

FATF frames digital identity as a risk-based CDD tool

FATF's "Guidance on Digital Identity" applies Recommendation 10 (customer due diligence) to digital ID systems used for identification and verification. Its central principle is proportionality: assurance levels and controls should reflect the risk of the customer and the product. That principle is the basis for the risk-tiered targets later in this article. A single onboarding SLA or STP target across all customers would contradict it.

Automation coverage needs a step-level definition before anything else

"We automated onboarding" isn't a measurable claim. Define coverage per step, then report it as a percentage of cases in which that step ran without a manual touch.

Step

What "automated" means

Coverage formula

Data capture

Customer enters data once; registry data prefilled

Cases with prefilled registry data / total cases

Document verification

Extraction and validity checks run without analyst input

Documents auto-verified / documents received

Ownership and UBO checks

Ownership structure resolved and reconciled to registers

Cases with UBO resolved automatically / total cases

Sanctions and PEP screening

Screening run and low-risk alerts dispositioned by rule

Alerts closed by policy / total alerts

Decisioning

Policy applied and outcome recorded without manual review

Cases decided by policy / total cases

Escalation

Exceptions routed to the right queue with context attached

Escalations routed automatically / total escalations

Report coverage per step and per risk tier. A bank that automates document checks for 90% of cases but resolves ownership structures manually will see its cycle time barely move, because ownership resolution is usually where the long tail sits.

The KPI set covers six areas, and each one maps to a regulatory or financial question

Each KPI below has a formula, so two analysts calculating it get the same number. Compute them per entity type (for example sole proprietorship, limited company, foundation) and per risk tier.

Time and throughput

  • Time-to-first-decision: timestamp of first system or analyst decision minus timestamp of complete submission. Use the median and the 90th percentile, not only the average.

  • Time-to-complete CDD: timestamp of account-ready status minus timestamp of application start.

  • SLA attainment: cases completed within SLA / total cases.

  • Bottleneck contribution: elapsed time per step / total elapsed time. Include waiting time for customer responses as its own step.

Analyst workload

  • Analyst minutes per case: logged handling time summed per case. Where time isn't logged, sample 50 cases per tier and time them.

  • Manual review rate: cases with at least one analyst touch / total cases.

  • Escalations per 100 cases.

  • Rework rate: cases sent back to the customer or re-opened because of missing or incorrect data / total cases.

Quality and compliance

  • Straight-through processing (STP) rate: cases approved or rejected by policy with no analyst touch / total cases. Report it by tier, since high-risk STP should sit near zero by design.

  • Screening false-positive rate: alerts closed as no-match after review / total alerts.

  • False-reject rate: applications rejected and later approved on appeal or reapplication / total rejections. This is the hardest KPI to measure and the most revealing.

  • Discrepancy rate: cases where submitted data conflicts with register data / total cases.

  • Evidence completeness: cases whose file contains every item required by policy / total cases audited.

Conversion and revenue

  • Completion rate: completed submissions / started applications.

  • Step-level drop-off: abandonments at step N / entries into step N.

  • Approval rate by risk segment.

  • Reboarding success rate: customers re-verified within deadline / customers due for re-verification.

Duna's own analysis of business onboarding finds that each follow-up request loses roughly 15% of submitted customers. That figure comes from Duna's published research on onboarding as a conversion problem, and your own baseline may differ. It does explain why rework rate belongs in the conversion section as well as the workload section: every avoidable round-trip is both an analyst cost and a drop-off event.

Control effectiveness

  • Decision traceability: decisions linked to a specific policy rule and version / total decisions.

  • Human-in-the-loop coverage: required human reviews completed / required human reviews. The target is 100%, and the metric exists to prove it.

  • Exception handling timeliness: exceptions resolved within their own SLA / total exceptions.

Data governance

  • Structured-field quality score: valid, complete, correctly formatted KYB fields / total fields.

  • Data acceptance rate: submissions passing validation on first attempt / total submissions.

  • Registry data error rate: prefilled or API-sourced values later corrected / total prefilled values.

Baseline first: ninety days of data decide which targets are realistic

Targets set without a baseline are guesses. Pull the last 90 days of closed cases and segment them by entity type and risk tier before setting any target.

  1. Export case-level timestamps from the case management or workflow system: application start, submission, each handoff, decision, account activation.

  2. Join analyst activity data. If the bank doesn't log handling time, run a two-week time-study sample of at least 50 cases per risk tier.

  3. Tag every rework event with its cause: missing document, incorrect data, unclear request, registry mismatch.

  4. Pull alert-level screening data and the disposition of each alert.

  5. Sample 30 to 50 completed files per tier and score them for evidence completeness against written policy.

Expect the exercise to surface surprises. Banks commonly discover that waiting on customers, not analyst work, dominates elapsed time, and that handling time varies several-fold between analysts on similar cases. Those findings point to different fixes: intake design for the first, policy codification for the second.

Target ranges by maturity level are planning assumptions, not industry benchmarks

No credible, neutral, bank-size-specific benchmark set exists for KYB metrics, and published vendor figures aren't comparable across institutions. The ranges below are planning assumptions for a mid-size bank. Treat them as hypotheses to validate against your measured baseline, then adjust.

KPI

Level 1: early automation (intake, document extraction, registry prefill)

Level 2: risk-based optimization (policy-driven decisioning, tiered routing)

Level 3: near-STP for low risk

Median time-to-decision reduction vs. baseline

25% to 40%

50% to 65%

70% or more

Analyst minutes per case reduction

20% to 35%

40% to 55%

60% or more

STP rate, low-risk tier

15% to 30%

40% to 60%

70% to 85%

Screening alert volume escalated to analysts

10% to 20% lower

30% to 50% lower

50% or more lower

Rework rate reduction

20% to 30%

40% to 50%

50% or more

SLA attainment

90%

95%

97% or more

Evidence completeness (audited sample)

95%

98%

100%

Evidence completeness is the one line that shouldn't be traded off. A bank can accept a slower cycle-time improvement. It can't accept an audit file that gets thinner as automation expands.

Targets must differ by risk tier because the regulators' own logic is risk-based

FATF's digital identity guidance ties assurance and controls to risk. The EBA guidelines likewise expect the solution's design and the scope of verification to reflect the ML/TF risk of the customer relationship. One blended target hides the problem: a 60% overall STP rate could mean 90% of low-risk cases and 0% of high-risk cases, or a worrying mixture.

Dimension

Low risk

Medium risk

High risk

Automation expectation

Decision by policy for clean cases

Automated data gathering and screening; analyst decides on flagged cases

Automation prepares the file; analyst and approver decide

STP target (Level 2 planning range)

40% to 60%

10% to 25%

Near zero by design

Human-in-the-loop

Sampled quality review

Required for any flag

Required, with four-eye approval

Time-to-decision priority

Minutes to hours

Same day to two days

Quality and depth over speed

Primary KPI emphasis

STP, completion rate

Escalation rate, false positives

Evidence completeness, traceability

Setting different goals per tier also gives compliance a reason to support the project. The efficiency gain is concentrated where risk is lowest, and analyst time is redirected toward the cases that carry real exposure.

The executive dashboard fits on one page, with a drill-down for compliance operations

Executives need six measures, trended monthly and split by risk tier: median time-to-decision, SLA attainment, cost-to-serve per onboarded customer, completion rate, analyst minutes per case, and exception rate.

Compliance operations needs the drill-down: step-level bottleneck contribution, escalation reasons, false-positive rate by screening list, rework causes, discrepancy rate by registry source, and the control view (traceability, human-in-the-loop coverage, exception timeliness). Add a regression panel that flags any control KPI falling below its threshold, so a speed gain never hides a quality loss.

The ROI model needs eight inputs and every output must map back to a KPI

Inputs

Input

Source KPI

Applications started per month

Completion rate denominator

Baseline and target completion rate

Completion rate

Completed cases per month

Derived

Loaded analyst cost per minute

Finance

Baseline and target analyst minutes per case

Analyst minutes per case

Approval rate

Approval rate by segment

First-year gross profit per approved business customer

Finance or product

Implementation cost (amortized) and annual run cost

Vendor and internal estimates

Formulas

  • Labor savings = future completed cases × (baseline minutes − target minutes) × cost per minute

  • Incremental approved customers = started applications × (target completion − baseline completion) × approval rate

  • Conversion value = incremental approved customers × first-year gross profit per customer

  • Annual cost = amortized implementation + annual run cost

  • Net annual benefit = labor savings + conversion value − annual cost

  • ROI = net annual benefit / annual cost

  • Payback (months) = upfront implementation cost / ((labor savings + conversion value − run cost) / 12)

Worked example with placeholder values

The figures below are illustrative planning assumptions in generic currency units (CU), not industry averages.

  • 1,000 applications started per month; completion rate rises from 62% to 70%. Completed cases move from 620 to 700 per month.

  • Analyst cost: 1 CU per minute (60 CU per hour, loaded). Minutes per case fall from 150 to 70.

  • Approval rate: 90%. First-year gross profit per approved customer: 600 CU.

  • Implementation: 450,000 CU amortized over three years (150,000 CU per year). Annual run cost: 220,000 CU.

Calculation:

  • Labor savings: 700 × 12 months × 80 minutes × 1 CU = 672,000 CU per year

  • Incremental approved customers: 1,000 × 12 × 8% × 90% = 864 per year

  • Conversion value: 864 × 600 = 518,400 CU

  • Annual cost: 150,000 + 220,000 = 370,000 CU

  • Net annual benefit: 672,000 + 518,400 − 370,000 = 820,400 CU

  • ROI: 820,400 / 370,000 ≈ 222%

  • Payback: 450,000 / ((1,190,400 − 220,000) / 12) ≈ 5.6 months at full run-rate

Run the sensitivity case that finance will ask for. With conversion value set to zero, net annual benefit is 302,000 CU, ROI is about 82%, and payback is roughly 12 months. The case holds on labor savings alone, and conversion is upside.

Two cautions apply. First, a bank that doesn't plan to redeploy or avoid hiring won't realize the labor line as cash, so state whether savings come from avoided hires, overtime, or capacity absorbed by growth. Second, ramp-up matters: model benefits at 25%, 60%, and 100% of target across the first three quarters.

Proof requirements

Every ROI line needs a KPI that evidences it after go-live. Labor savings map to analyst minutes per case and manual review rate. Conversion value maps to completion rate, step-level drop-off, and rework rate. Cost avoidance on false positives maps to alert escalation volume. Quality assurance maps to evidence completeness and discrepancy rate. If a benefit can't be mapped, take it out of the model.

Two illustrative vignettes show how the KPIs move in practice

These are template-style composites for planning purposes, not reports on specific institutions. For documented examples of lifecycle automation at scale, see Duna's perpetual KYB case studies, including the re-verification of 40,000 partners at bol.

Vignette one: intake and document verification

  • Bottleneck: analysts re-key registry data and check incorporation documents by hand. Rework runs high because customers receive multiple separate requests for missing items.

  • Automation steps: registry prefill, single-pass document request driven by entity type, automated extraction and validity checks, discrepancy flags where submitted data conflicts with the register.

  • KPI movements to expect: rework rate and analyst minutes per case fall first, followed by completion rate as follow-up requests drop. Discrepancy rate rises initially, because the system now detects conflicts that manual review missed. Don't read that as a quality regression.

  • Lessons: log which fields were prefilled and which were customer-supplied so the file shows the provenance of every data point. Keep a sampled analyst review of auto-verified documents until the error rate is stable.

Vignette two: risk-based decisioning and escalation

  • Bottleneck: every case goes through the same analyst queue regardless of risk, and screening alerts on low-risk customers consume disproportionate time.

  • Automation steps: policy rules assign a risk tier at intake, route cases by tier, close low-risk alerts that meet documented criteria, and send high-risk cases to a four-eye queue with the evidence pack attached.

  • KPI movements: STP rate in the low-risk tier and escalations per 100 cases move first. Time-to-decision for high-risk cases may barely change, which is the correct outcome.

  • Lessons: version every policy rule and record which version produced each decision. Approve rule changes through the same governance route as model changes, and back-test them on historical cases before activation.

The business case for stakeholders should address each function's actual objection

A single document can serve four audiences if its sections are separable.

  • C-suite and COO: cost-to-serve, time-to-revenue, and the conversion effect. Duna's analysis of why onboarding belongs on the COO agenda makes the same argument that onboarding drives conversion, cost, and productivity, not only compliance.

  • Compliance: alignment with the EBA remote onboarding guidelines, the FATF risk-based principle, and the AMLR timeline; the control KPIs; the human-in-the-loop design.

  • Operations: the baseline findings, bottleneck map, and expected change in analyst workload.

  • IT and security: integration scope, data residency, access controls, and the outsourcing and vendor-risk position.

Governance has to be explicit in the document

The EBA guidelines keep accountability with the institution, so the case should name an owner for each of these:

  • Approval workflow for policy rules and any AI or scoring component, including validation before release and periodic revalidation.

  • Human-in-the-loop design: which decisions are never automated, which are sampled, and what triggers mandatory review.

  • Vendor accountability: the bank's outsourcing assessment, exit plan, fallback procedures if a data provider fails, and audit rights.

  • Evidence retention: what is stored, for how long, and in what format, aligned to Wwft or equivalent national retention requirements.

Capability matters here because it determines whether governance is practical. A policy engine that expresses compliance logic as versioned rules, with an audit trail of interactions, decisions, and policy changes, gives the bank something to show a supervisor. Duna's platform is built on that structure, and its reported outcomes (10.6x faster onboarding, a 37% conversion increase, and 4.8x analyst efficiency) are the company's own customer figures, so validate any vendor claim against your baseline in a pilot. Banks running a formal selection can use the criteria in the bank-grade KYB RFP overview, and Duna's guide to AI in KYC and AML covers how to reduce manual work without losing control.

A 30/60/90-day plan gets from baseline to a scaled pilot

Days 1 to 30: baseline and process map

  • Complete the 90-day data pull and sampling described earlier.

  • Map the process by entity type and risk tier, with elapsed time per step.

  • Fix KPI definitions in writing and have compliance sign off on them.

  • Rank bottlenecks by contribution to elapsed time and analyst minutes.

Days 31 to 60: pilot design and build

  • Choose one entity type and one risk tier, typically low-risk limited companies, where volume is high and policy is clearest.

  • Codify the policy for that segment as explicit rules, including escalation triggers.

  • Configure intake, registry prefill, document checks, and screening disposition rules.

  • Set the dashboard up and test traceability on sample cases.

Days 61 to 90: run, measure, decide

  • Run the pilot in parallel with the manual process on a defined share of cases, or with 100% analyst review of automated decisions for the first weeks.

  • Report KPIs weekly against the baseline.

  • Present a scale-or-stop recommendation to the steering group.

Acceptance criteria combine improvement targets with no-regression ceilings

Category

Criterion (planning assumption)

Speed

Median time-to-decision down at least 25% in the pilot segment

Workload

Analyst minutes per case down at least 20%

Conversion

Completion rate up, or rework rate down at least 20%

Evidence

Audit-sample completeness at or above baseline, and at least 95%

Traceability

100% of automated decisions linked to a rule and version

Human review

100% of required human reviews completed

Discrepancies

Rate within a ceiling agreed with compliance (rising detection is acceptable if explained)

False rejects

No increase against baseline

If the pilot meets the improvement targets but breaks any no-regression line, it doesn't pass. That condition is what makes the project defensible to a supervisor.

The baseline decides where automation goes, and the control metrics decide whether it stays

Manual, slow business onboarding rarely has a single cause. Waiting on customers, re-keyed data, one-size-fits-all review queues, and alert noise each contribute, and the 90-day baseline shows how much. Automate those steps by risk tier, hold the control KPIs fixed, and report speed and conversion gains against the same definitions every month.

The EBA's remote onboarding guidelines, the AMLR, and FATF's digital identity guidance all point to the same operating model: automation that is risk-based, governed, and traceable. A mid-size bank that builds that model now, with KPIs it can show a supervisor, has also built the business case it needs. For a view of how different platform architectures fit banks of different sizes, see Duna's KYC and KYB orchestration platform guide.