AI ROI Calculation for UK SMEs

In short

A defensible AI ROI calculation has three legs: labour savings (hours × rate × volume), revenue lift (conversion rate × AOV × visitors, or pipeline → close rate × deal size), and risk reduction (avoided cost of errors). Wingenious produces this as the Feasibility Study deliverable: £3,950, 2–3 weeks.

What’s in the model

• Baseline state: current cost/time/error rate without AI
• Future state: projected with AI (best case, expected case, worst case)
• One-off cost: the build + integration
• Ongoing cost: API + tooling + maintenance
• Payback period: months to break even
• 3-year NPV: defensible against finance scrutiny
• Sensitivity analysis: which variables matter most

What we won’t do

• Pretend the upside is certain when it isn’t
• Use industry-average ROI numbers where your data exists
• Ignore the integration cost (which is usually 30-50% of total)

The three places ROI calculations go wrong

Most ROI numbers that arrive on an SME’s desk are wrong in predictable ways. Three failure modes account for the bulk of them.

The first is the upside-only model. A vendor quotes the ROI assuming everything goes well and quoting from enterprise benchmarks. Integration cost is missing because the vendor does not see it. Maintenance is absent because the vendor does not pay for it. Adoption risk is glossed over because the vendor’s sales process is incentivised to assume adoption is solved. The model looks attractive on paper and is wrong by a factor of two or three by the time the build is twelve months old.

The second is the single-number model. ROI quoted as “5x return inside 12 months” with no sensitivity analysis. The number depends on a dozen assumptions and any one of them being off by 30 percent moves the result dramatically. Without sensitivity analysis the board cannot interrogate the assumptions that drive the headline.

The third is the boil-the-ocean model. The ROI includes every conceivable benefit, including ones that are difficult or impossible to attribute back to the build. “Improved customer experience” gets counted as £50,000 of retained revenue without a defensible mechanism. The board sees through the inflation and discounts the whole model, including the parts that were honest.

The Wingenious approach avoids all three. The model is honest about downside, transparent about assumptions, and conservative about benefits that cannot be directly attributed.

The three legs of a defensible model

The framework Wingenious uses has three categories of benefit, each modelled separately so the board can weight them according to their own appetite for soft numbers.

1. Labour savings. Hours saved multiplied by fully-loaded hourly rate multiplied by volume. The most defensible category. If automating an invoice workflow saves 12 minutes per invoice across 800 invoices a month at a fully-loaded rate of £28 per hour, the annual saving is £53,760. The model documents each assumption: minutes per invoice (with range), volume (with seasonality), fully-loaded rate (including pension, NI, overheads). Sensitivity is explicit.
2. Revenue lift. Conversion rate, average order value, customer lifetime value, win rate, or pipeline-to-close rate. Modelled against your current baseline, not against industry benchmarks. The lift is bracketed by a conservative, base and optimistic scenario. Where the lift compounds over time (customer retention improvements, segment-aware personalisation), the model uses a decay or growth curve rather than a flat percentage.
3. Risk reduction. Errors prevented, late fees avoided, compliance fines dodged, reputational damage averted. Modelled using historical incident rate multiplied by typical cost per incident, then discounted for uncertainty. The board sees the calculation rather than just a number; they can argue with each input.

The three categories combine into the headline ROI figure. They are also reported separately so the board can decide which categories they trust most for their own decision shape.

The costs that get missed

The cost side of the model includes more than the build invoice. Six categories appear consistently.

• Build cost. The headline number from Wingenious, the vendor, or the internal team.
• Integration cost. Usually 30 to 50 percent of build cost. The work of wiring the new capability to existing systems. The single most under-estimated category.
• Ongoing licence cost. API calls, vendor subscriptions, seat fees. Modelled at expected volume, not at trial volume.
• Maintenance cost. Updates, monitoring, occasional fixes. Typically 15 to 25 percent of build cost per year for an ongoing custom build; lower for vendor products.
• Internal time. The hours your team spends working with the new tool, training on it, providing oversight. Usually invisible in the procurement conversation but real in the running cost.
• Opportunity cost. What the leadership team, finance team and operations team are not doing while they are doing this. Sometimes material, particularly for SMEs running multiple AI initiatives simultaneously.

The model includes all six. Hiding any one of them produces a misleadingly attractive headline.

Payback realism

Realistic payback periods vary by category of build.

• Operational AI (workflow automation, document processing, customer support automation, data cleaning): typically 6 to 12 months. Labour savings appear quickly and compound predictably.
• Strategic AI (segmentation, personalisation, pricing optimisation, lead scoring): typically 12 to 18 months. Revenue lifts appear gradually and the model has to learn against your specific data before the benefit shows up.
• Differentiating AI (custom builds at the heart of the SME’s competitive edge): typically 18 to 36 months. Higher build cost, larger benefit, longer payoff curve.

Builds projecting payback inside 90 days are usually optimistic, and the model documents this honestly. Builds projecting payback beyond 24 months usually need a stronger differentiation story to justify the patient capital they require.

The Feasibility Study itself

The ROI calculation is the centrepiece of the Feasibility Study at £3,950, two to three weeks. The Study also produces the business case, the built-versus-buy analysis, the vendor shortlist, the risk register and the recommended next step.

The model is delivered as an editable spreadsheet (Excel or Google Sheets) with documented assumptions. The SME’s finance team can rerun it with updated volumes, costs or vendor pricing as conditions change. Quarterly refresh is straightforward.

What happens after the model is built

Three patterns.

1. Approve and deliver. The board signs off, the build proceeds via Implementation Sprint or Quick Win. The model becomes the benchmark against which the live workflow gets measured in the quarterly review cycle.
2. Pause and address. The model surfaces a prerequisite (data cleaning, governance, team capability) that needs to land before the main build. The SME addresses the prerequisite first, the model gets refreshed, and the build proceeds in the next quarter.
3. Kill. The model shows that the build does not pay back at any realistic assumption set. The Feasibility Study saves the SME from a wrong investment that would have cost many times the £3,950 fee.

Around a quarter to a third of Feasibility Studies recommend not proceeding with the specific build proposed. That is the diagnostic working as intended.

How the model survives the build

A model that gets written, approved and forgotten is half a model. The Wingenious version is built to live alongside the production workflow.

• Quarterly refresh. As part of the quarterly AI review, the actual performance gets compared against the model. Deviations get explained. The model gets refreshed with the new data.
• Annual revalidation. A full re-model annually catches structural shifts (new vendor pricing, regulatory changes, business model evolution). Available to past clients at half the original fee.
• Decision archive. The model and its assumptions sit in the SME’s shared workspace. When the next AI investment is considered, the team has a reference point for what assumptions held and what did not.

What the finance director cares about

Finance directors interrogating an AI business case typically want three things the model has to deliver explicitly.

The first is conservative base-case numbers. A model whose base case is the same as the optimistic case is suspicious by default. The Wingenious version makes the base case the modal expected outcome, with optimistic and conservative bands clearly separate. The finance director can decide which case to plan against.

The second is downside transparency. What happens if the build delivers half the projected benefit. What happens if the volume forecast is 30 percent below plan. What happens if the integration takes twice as long. The sensitivity analysis surfaces the bad-case numbers honestly.

The third is comparable-investment context. How does the projected ROI on this AI build compare to the ROI on the SME’s other recent investments. The number that looks impressive in isolation might look ordinary against the SME’s typical capex return. The model includes the comparison where the historical data supports it.

Common assumptions that get pushed too hard

Three assumptions tend to be over-played in optimistic ROI models, and the Wingenious version pushes back on each.

The first is adoption. Models often assume 100 percent of the affected team uses the new tool from day one. Reality is usually 60 to 80 percent in month one rising to 90 percent by month three. The model uses the realistic adoption curve, not the wishful one.

The second is volume. Models often assume the workflow runs at peak volume from day one. Reality is usually a ramp as the team gets comfortable. The model uses the realistic volume curve.

The third is benefit retention. Models often assume the benefit, once captured, persists indefinitely at full strength. Reality is usually that some erosion occurs over years as the business changes shape and the tool drifts from purpose. The model includes a modest decay assumption to reflect the truth.

Pushing back on these three assumptions usually reduces the headline ROI by 20 to 40 percent and increases its credibility by considerably more. Finance directors trust a number that has been argued with.

Where the model lives after delivery

The ROI model is a working document, not a one-off report. Three practical homes for it.

The first is the SME’s shared workspace. A version on Google Sheets or Excel that the finance team can update with current figures. Versioning matters: the original projection should be preserved alongside the live revision so the team can see the drift.

The second is the Quarterly AI Review document. The current actuals against the projection feed directly into the workflow performance section. Where the model is meeting projection, the workflow is green. Where it is missing, the section explains why.

The third is the next business case. The assumptions that held and the assumptions that did not become inputs to the next AI investment decision. A leadership team that has built one honest ROI model is better at building the second one because the first set of assumptions is a real reference point.

Related capabilities

AI business case · Built-vs-buy analysis · Vendor shortlisting · AI use case identification

Related

Sectors where ROI calculation matters most: manufacturing, accountants.