AI ROI in Financial Services: Costs, Cash Flow and Risk

An AI business case for a bank or insurer needs a defined comparison, attributable financial benefits and the full cost of running the proposed service. Separate expenditure that can be avoided from staff capacity, and show expected loss reductions separately from cash already received. Calculate returns over the same period as the costs.

This guide supplies no sector-wide ROI benchmark. Every numerical business example below was newly invented for this guide. None describes a customer, an anonymised engagement, a supplier quotation or an observed banking result. The examples explain calculation choices; they do not predict investment performance.

Establish a comparison that survives scrutiny

Choose a workflow, its eligible population and an observation period. For payment investigations, record the number and complexity of cases, human handling time, rework, external processing charges and service quality. Compare the proposal with continuing the current process and with a feasible alternative, such as improving existing rules or forms.

Measure the complete workflow. A faster first draft can still require more checking or produce additional complaints. Include exceptions, escalations and corrections. In a pilot, use a comparable control group where feasible and record changes in customer mix, staffing, policy and transaction volume. Otherwise, an improvement caused by a quieter queue could be attributed to AI.

For lending, compare borrower groups, approval volumes, exposure, observation windows and economic conditions. A lower observed default rate after rejecting more applicants does not establish a better return on the same lending activity. For insurance, compare claim complexity and final settlement outcomes as well as processing speed.

In lending, additional interest or fee revenue needs the associated funding, servicing and expected loss costs deducted before it becomes a contribution estimate. Avoid counting a credit-loss reduction again if it is already included in that contribution. Additional lending can also require funding and working-capital cash flows that a simple revenue table omits.

Estimate the spending needed to operate the service

Build costs from tasks, contracts and usage estimates. Record implementation, data preparation, integration, testing, training, assurance, ongoing support and eventual exit. Allocate shared infrastructure once across the programme. Show internal staff effort even where it does not create another payroll payment.

For a decision about future spending, distinguish incremental cash flows from money already spent. Include backfill or a measurable financial effect of displaced work where relevant. ACCA explains the treatment of relevant cash flows and consistent tax and inflation assumptions in its investment-appraisal guidance.

A bank’s regulatory obligations do not establish a standard price for model validation or resilience testing. Estimate the actual work required for the entity, system and intended use. Specify what a contingency covers and whether the forecast assumes it is spent.

Hypothetical example: preparing payment-investigation files

Consider an invented project that helps staff prepare files for payment investigations. Staff still review the material and make decisions. The model assumes the required quality and service standards are maintained. It assigns no benefit to faster settlement, fewer fraud losses or better compliance.

Newly invented operating assumptions

The workflow-use percentages already include the year’s rollout and uptake. The separate 60% assumption identifies released hours that eliminate expenditure. The other hours remain available capacity and are excluded from the cash total.

Invented cash-cost budget

The €2,000 is an actual annual payment to an external provider for export readiness and exit support, including final exit work in Year 3. No further exit payment is assumed. The other operating amounts are additional cash expenditure. Human review is already included in the nine-minute handling time and is not charged again.

The model is before tax, in constant euros, with an illustrative real discount rate of 8%. It excludes inflation, financing flows, working-capital changes, residual value and displaced-work costs. All annual benefits and operating payments occur at year end; implementation is paid at time zero. A real proposal must add any applicable items and use a discount rate consistent with its cash flows.

Calculate the cash benefit and return

Released hours = 72,000 × (12 − 9) ÷ 60 × Workflow-use share
Cash benefit = Released hours × 60% × €45
Net annual cash flow = Cash benefit − €26,000

Three-year cash benefits total €194,400. Costs total €146,000: €68,000 initially and €78,000 during operation. Using cumulative net benefit divided by all included costs:

Three-year ROI = (€194,400 − €146,000) ÷ €146,000 = 33.2%
NPV at 8% = −€68,000 + €22,600/1.08 + €46,900/1.08²
            + €46,900/1.08³ = approximately €30,366

The 33.2% result is cumulative over three years. It is not an annual return. NPV incorporates the specified payment timing. Under the year-end assumption, simple payback occurs at the end of Year 2, or 24 months. Annual totals do not justify a more precise date within that year. The general AI ROI guide explains these measures in more detail.

Check an adverse scenario

The downside keeps volume, handling-time reduction, the 60% cash-realisation share and the €45 hourly amount unchanged. Workflow use falls and annual operating expenditure rises by €5,000, all assigned to additional support. The initial outlay stays at €68,000.

These scenarios have no assigned probabilities. The downside identifies two assumptions to test: how much work will use the system and how much support it needs. A probability-weighted forecast would need a justified set of mutually exclusive scenarios, including failure or abandonment and the costs incurred in those outcomes.

Hypothetical credit-loss calculation: define the percentage change

Use probability, exposure and severity together when estimating credit losses. The Basel Framework’s CRE35.2–35.3 distinguishes the expected-loss rate from the amount obtained by multiplying it by exposure at default under the internal ratings-based approach. The following simplified planning example is separate from that regulatory calculation and from the operations project above.

All inputs in this credit example are also newly invented. It considers one year, a comparable pool of performing loans with €6,000,000 of exposure at default (EAD), an exposure-weighted probability of default (PD) of 2.5%, and loss given default (LGD) of 35%. The LGD is assumed to include recoveries and recovery costs. The alternative holds exposure and borrower mix constant and assumes PD falls to 2.3%, with unchanged LGD.

Simplified expected loss amount = EAD × One-year PD × LGD
Before: €6,000,000 × 2.5% × 35% = €52,500
After:  €6,000,000 × 2.3% × 35% = €48,300
Expected loss reduction over one year = €4,200

That PD change is 0.2 percentage points, or 8% relative to the starting PD. It is not a 0.2% relative reduction. For the same invented exposure and LGD:

These amounts are before project costs and tax. They are not project ROI, guaranteed annual savings or an IFRS 9 provisioning calculation. No receipt date is assumed, so this example does not calculate NPV. A cash-flow appraisal would need the costs of the proposed change and the timing of defaults, recoveries and other payments. A provision release must not be added as a second cash benefit for the same underlying loss reduction.

Validate any forecast PD change against a suitable observed comparison. Test sensitivity to exposure, LGD, approval rates and economic conditions. Neither this example nor a lower back-test error establishes that an AI model will reduce losses in production.

Budget regulatory work for the actual use

The legal sources below were checked on 14 September 2026. Identify the entity, jurisdiction, intended purpose and provider or deployer role before estimating obligations. A banking or insurance label alone does not classify an AI system.

AI Act. Article 6 and Annex III 5(b)–(c) address natural-person creditworthiness and credit scores, and risk assessment and pricing for life and health insurance. Fraud detection is excluded from the creditworthiness entry; this is not a blanket exemption from other rules. Apply the classification conditions to the actual use. See the current consolidated AI Act.

Timing. Under the amended Article 113, Chapter III Sections 1–3, except Article 6(5), apply to Annex III high-risk systems from 2 December 2027. Other provisions have different dates, and Article 111 includes transitional conditions. The Commission’s AI Omnibus entry-into-force notice confirms the changed timetable. Budget preparation against the applicable provisions and planned deployment date.

DORA. Articles 2 and 4 establish scope and proportionality; Article 24 covers risk-based resilience testing, with qualifications including microenterprise treatment. Article 28 addresses ICT third-party risk. Articles 50–51 leave administrative penalties and their exercise to national legal frameworks. They do not establish a universal €10 million or 5% ceiling. Cost the required work and identify the relevant national and entity regime. See DORA.

Record applicable data-protection, sectoral and supervisory requirements with the people responsible for them. This guide assigns no standard regulatory price per model. Keep unsubstantiated avoided fines out of the base cash case; statutory maximums do not establish an event probability or a loss estimate. Required controls still need funding when a speculative penalty-avoidance figure is unavailable.

Keep the decision tied to observed results

For each material assumption, record its definition, evidence, owner, date and range. Give finance the requested funding, alternatives, annual cash flows and downside. Give operations the quality thresholds and staffing implications. Shared costs and overlapping benefits need a single consistent treatment across the portfolio.

During rollout, compare actual usage, complete handling time, paid expenditure and service outcomes with the forecast. Explain a variance before changing the baseline. Set a review date and a decision rule for continuing, narrowing or stopping the investment.

Frequently Asked Questions

What ROI should a financial institution expect from AI?

There is no verified sector-wide target in this guide. Compare a specific investment with its alternatives, using the same horizon, cost scope and risk assumptions. The hypothetical 33.2% result illustrates one calculation and supplies no forecast for a bank or insurer.

Can all time savings be counted as cost savings?

No. Record hours released after checking and rework. Include cash savings where the change removes paid expenditure, or estimate a separate financial use for the capacity. An unchanged salary bill does not fall when a task takes less time.

Should credit-loss reductions be included in ROI?

A supported expected-loss estimate can inform the investment decision. State PD, exposure, LGD, period and comparison. Convert the effect into appropriately timed cash flows and include project costs before calculating a cash-based return. Keep accounting provision movements distinct and remove overlap with lending contribution.

How should compliance costs and possible fines be handled?

Estimate applicable implementation and recurring control work in the cost budget. A statutory maximum fine is not a benefit forecast. Establish the relevant obligation, jurisdiction and loss assumptions before assigning an expected value, and keep unsupported fine avoidance out of the base cash case.

How quickly should the investment pay back?

Use the project’s dated cash flows and funding requirements. The fictional operations project recovers its outlay at the end of Year 2 under year-end timing; its downside does not recover costs within three years. Neither result establishes a normal banking payback period.

Related reading and contact

Read the AI ROI calculation guide and the financial-services overview.

Contact.