AI automation ROI: make the business case before development

AI automation ROI is the financial value of a specific workflow after all build and operating costs are counted. Start with the current process, price the automated version per completed task, and test uncertain assumptions in a small pilot. If the team cannot measure the baseline or define a correct result, it does not yet have a defensible business case.

I work on digital products and advise teams on what to validate before committing to development. I start with the workflow itself: what it costs today, what a pilot can measure, and which result would justify more work.

What should an AI business case answer?

It should answer one decision: should the team test, build, change, or stop a proposed workflow? That requires a comparison over a defined period, such as a pilot window or the first year after launch.

The familiar ROI formula is:

ROI = (measurable value during the period − total cost during the period) / total cost during the period × 100%.

The calculation is only as useful as its definitions. “Value” may mean direct costs avoided or additional contribution margin that can be attributed to the workflow. Faster work is not automatically a cash saving. If staff keep the same hours and handle the same volume, the business has gained capacity. Record it as capacity until the team can connect it to a real operating or revenue measure.

Pick one unit of work, such as a reviewed document or a qualified enquiry. A role may contain many tasks, and only some may suit an LLM. MIT Sloan recommends breaking work into tasks, weighing the cost of reaching the required level of correctness, and evaluating pilots rather than assuming a model can do the whole job (MIT Sloan on choosing generative AI use cases).

How do you establish a baseline?

Describe the current workflow before changing it. For a set period, log how many units arrive, how long people spend processing and checking them, how often work returns for correction, and how long a request stays in the queue.

Baseline measureEvidence to collectDecision it supports
Completed volumeSystem events or a reconciled operational reportWhether there is enough repeat work to matter
Staff time per unitA sample of active handling and review timeCurrent labor effort, separate from cash savings
Rework and errorsCorrection records, causes and verified direct costsQuality-adjusted cost per completed unit
Waiting timeTime from intake to a completed outcomeCapacity, service level and backlog questions
ExceptionsCases that need a specialist or a second systemExpected human review and escalation effort

Use the same definition of a completed unit before and after the pilot. If a team changes the intake form, staffing or process rules at the same time, a before-and-after comparison will not isolate what the AI workflow contributed.

Where records are missing, do not fill the gap with a confident-sounding estimate. Add a small measurement step first. A manual time sample or event log may be enough to decide whether an AI pilot is worth testing.

If the workflow itself is still vague, first define the MVP scope. For support automation, decide how the assistant hands a conversation to a human before estimating review effort.

For an inbound lead workflow, the completed unit might be an enquiry saved with an owner and a next action. See the related lead capture automation example for the operational steps that need to be counted.

Which costs belong in the model?

Price the end-to-end workflow, not just model calls. The cost list usually includes:

  • Implementation: discovery, prototyping, data preparation, integration, access controls, testing and deployment.
  • Model and tool usage: input and output tokens, retries, search or other tools, storage and infrastructure. Provider prices vary by model and request type. Check current terms when doing the calculation, such as the OpenAI API pricing page.
  • Human review: the time to check, correct, approve or take over an answer.
  • Failure handling: repeated work, escalation, recovery and the cost of a wrong action.
  • Operations: monitoring, prompt or data updates, regression testing, integration maintenance and incident response.
  • Adoption: training, new procedures and the process owner's time.

For a workflow where a mistake can cause real harm, include privacy, security and specialist oversight. NIST's generative AI profile treats risk management as work that continues across the AI lifecycle, rather than a single pre-launch check (NIST AI 600-1). That can change both the cost and the pilot design.

How do you estimate the cost per successful task?

Use a like-for-like unit:

Automated cost per completed unit = model and tool usage + remaining staff time + expected rework.

Then compare it with the current cost per completed unit. Multiply the difference by a volume you can support with records. Subtract fixed monthly expenses and one-time implementation costs for the period you chose.

This calculation should include non-happy-path work. A task might trigger several model calls, fail a validation rule or require a person to review the result. Count those events. The useful number is not the cheapest successful request; it is the cost of completing a unit at the quality level the business needs.

You can keep the decision sheet compact:

InputCurrent workflowPilot or forecastSource / assumption
Work units per periodMeasuredMeasured or cappedCRM, queue or event log
Staff minutes per unitSampledReview plus exception timeTime sample
Error and rework costRecordedMeasured in the pilotCorrection log
Model and tool costNot applicableActual request logs and current pricesProvider records
Build and integrationNot applicableOne-time estimateScoped work plan
Ongoing supportCurrent process costMonitoring, updates and ownershipOperating plan

For a rough payback estimate, divide one-time implementation cost by positive net monthly cash benefit. If that monthly benefit is zero or below, the current assumptions do not show a payback. A hoped-for future increase in volume is a separate scenario, not proof that the project already works.

What should a pilot prove?

Keep the pilot narrow enough that the team can inspect its outputs. Before development, name the task, the people who will judge the result, the quality floor, and the cases that must go to a human.

Track cost per completed unit alongside the business measure that matters. Depending on the workflow, that could be time to resolve, queue size, rework, or a directly attributable margin measure. Acceptance rate alone is not a financial return. Nor is a model's benchmark score a substitute for testing against the team's actual inputs.

Use a stop rule. Set a maximum cost, error rate, review burden, or risk boundary that would make expansion a bad decision. A pilot that fails that threshold is still useful if it tells the team what to redesign or why to stop.

For high-consequence decisions, keep the person accountable for the outcome. NIST's risk framework provides a way to organize testing, measurement and monitoring, but it does not supply a financial ROI result for a particular company. The company still needs its own process data and decision criteria.

A product analytics change that sharpened my view of AI costs

When reviewing analytics for a subscription digital product, I found that an AI cost figure is useful only when its event, owner, period and data coverage are clear. Before estimating ROI, check whether the system records the relevant model usage and whether human review effort can be tied to the workflow being evaluated. Missing, cached or unlogged calls may limit what the data can establish. This is a measurement principle, not evidence that a particular feature increased revenue or paid for itself.

When is it too early to build?

Wait if the team has not agreed on what a successful unit looks like, the process changes every week, or nobody can estimate its present workload. A manual pilot can reveal the steps that repeat and the exceptions that need judgment.

AI may also be the wrong tool. A deterministic rule, a better form or a simpler handoff can solve a workflow with fewer failure modes. Rare cases with expensive errors and heavy human oversight often make poor first pilots. Start with the business decision; add a model only if it helps.

When the workflow is frequent, measurable and bounded, use a pilot to test the few assumptions that dominate the spreadsheet. Then make a specific decision: expand, redesign the task, or stop. A negative result before a large build is a useful product decision.

Frequently asked questions

#### Can I calculate AI automation ROI before building?

You can model a range and identify the assumptions that matter. The actual result needs evidence from a pilot or operating system. Label forecasts clearly and keep them separate from observed savings.

#### Does saved employee time count as a financial return?

Only when the business converts that time into lower spend or measurable additional output. Otherwise, call it released capacity and report it separately from cash savings.

#### Should the model include human review?

Yes. Count review, correction, approval and escalations. If a person must check every output, that effort is part of the cost per completed unit.

#### What if there is no baseline data?

Instrument one workflow first. Record volume, handling time, corrections and exceptions before testing a model. A small measurement exercise may answer whether an AI pilot is worth funding.

#### When has an AI project paid back?

When verified financial value over the agreed period exceeds implementation and operating costs. State the period, cost boundary and evidence source before presenting a payback figure.

Sources and limits of the case