Look for work involving repeated interpretation of text, documents, conversations, or images, where a useful result can be checked. Automate a defined step before expanding into an entire business process.

These are candidates to investigate, not a ranking of guaranteed returns. An invoice workflow with poor source data and complex approval rules may be a worse first project than a modest inquiry-classification task.

Also check whether AI is needed. A stable calculation, fixed reminder schedule, or exact field-to-field transfer may be better served by ordinary automation. AI becomes relevant when interpretation or variation creates work that fixed rules cannot handle adequately.

Anthropic’s guidance on building effective agents recommends increasing complexity only when needed and distinguishes predefined workflows from agents that dynamically direct their own work. That distinction matters commercially: greater autonomy creates additional implementation and operating responsibilities.

Bring whatever you already know about process volume, handling time, errors, and software costs. Record whether each input is measured, estimated, or unknown.

The goal is one defensible next decision. If three workflows need substantial discovery, use the review to choose what to measure first rather than pretending to finish three investment cases.

Minutes 0 to 2: name the actual workflow

“Automate sales” is too broad to estimate. “Extract the requested service, location, and timing from an inbound email into a draft CRM record” gives the team something measurable.

For each candidate, identify its trigger, input, output, current handler, systems touched, and business owner. State what the initial automation will not do. For example, creating a draft record does not authorize sending a quotation or changing a customer agreement.

Then ask why the work is slow. If delays come from missing approvals or an unresolved policy, faster text generation may leave the bottleneck intact. If the task repeatedly involves reading variable documents and entering the same fields, AI-assisted extraction may address the actual effort.

Choose a task with an observable definition of done. The result should be an accepted record, checked answer, or completed handoff, not merely an AI response that looks convincing.

Minutes 2 to 4: apply the suitability gate

Before estimating a return, ask whether the workflow is a responsible and practical first candidate.

Do not let a large projected saving cancel out a failed permission or safety requirement. A positive spreadsheet does not make unauthorized access acceptable.

Generative AI can produce confident errors. NIST’s Generative AI Profile addresses this risk and cautions against extrapolating performance from narrow or anecdotal assessments. Review is itself work, so a “human checks everything” plan belongs in both the operating design and the cost estimate.

Minutes 4 to 6: estimate usable volume and net effort

Use cases per month, not the number of people who could theoretically benefit.

Expected AI-assisted cases = total cases × eligible share × adoption within eligible cases.

Eligibility asks which cases the proposed workflow can appropriately handle. Adoption asks how many of those cases will actually use it. Keep the definitions separate so the same reduction is not applied twice.

Released hours = AI-assisted cases × (baseline minutes − assisted minutes) ÷ 60.

Both times should cover comparable cases and the same work boundary. Assisted minutes include preparation, checking, corrections, exception handling, fallback work, and any new handoff effort attributable to the change.

Include unsuccessful attempts. Counting only easy successful cases makes the average look better than the workflow the business will operate. If the assisted process takes longer, retain the negative time difference.

Count each business case once in the volume estimate. Retries add effort and cost; they do not create extra completed business outcomes.

Distinguish active effort from elapsed time. An eight-hour wait for approval is not eight hours of employee work. Faster turnaround may have separate business value, but it is not automatically a labor saving.

For a rough estimate, use a representative sample and a range. After a pilot, replace assumptions with observed volume, usage, and effort. Microsoft’s agent-value guidance warns against theoretical time savings disconnected from adoption and business outcomes.

Minutes 6 to 8: count the full cost

Business process automation ROI should include the resources required to make the workflow useful and keep it useful.

Separate incremental cash spending from internal time assigned an economic value. Both can matter to the decision, but they answer different questions.

Do not count per-case review time twice: once in assisted minutes and again as a separate review bill. Count additional batch audits separately only when they sit outside that per-case work boundary.

AWS’s cost guidance includes development, ongoing operation, and management expenses in total cost of ownership. A model API estimate is therefore one input, not the whole project budget.

Minutes 8 to 10: test the downside and choose a next step

Change the assumptions most likely to alter the decision: eligible share, adoption, human review effort, usage cost, integration effort, and the value of the resulting capacity.

If a small change makes the return disappear, identify what the pilot must prove. Do not hide uncertainty inside a weighted score that gives every workflow an apparently precise ranking.

Use one of four next steps:

• Pilot: The task passes the suitability gate and has plausible value worth testing.
• Gather evidence: Missing volume, effort, or cost information prevents a useful comparison.
• Simplify or use ordinary automation: A process change, existing product feature, or deterministic integration better addresses the need.
• Defer: Required data, controls, ownership, or recoverability are not available for the proposed scope.

The process with the highest estimated ROI is not automatically the first choice. A smaller benefit that can be verified and delivered through a manageable integration may be a better starting point.

For a defined period, use:

ROI (%) = (total attributable benefit − total incremental cost) ÷ total incremental cost × 100.

Use a positive cost denominator and consistent units. Include setup costs once and recurring costs across the selected period. State whether the estimate uses cash benefits, assigned economic values, or both.

The formula is straightforward. Establishing the benefit is the difficult part.

Released capacity can be useful: employees may clear a backlog, serve more customers, or perform other valuable work. Name that intended use. A wage-based valuation is an estimate of resource value, not proof that all freed minutes create equivalent profit.

Agree the hourly valuation with the business’s finance owner. Employer compensation costs can inform it, but allocating fixed overhead to an hour does not mean that overhead will disappear when the task gets faster.

The UK Government Efficiency Framework illustrates the distinction between reduced spending and monetizable benefits that do not reduce spending. Its rules concern public-sector reporting; the distinction is also useful when interpreting a business automation estimate.

Use one credible benefit path for each improvement. Do not value the same hours as both released capacity and avoided payroll. If those hours also support additional sales, do not add the full sales benefit without checking overlap and attribution.

For a substantial investment, extend the screen into a month-by-month model with rollout delays, benefit ramp-up, recurring commitments, and the financial treatment your business uses. The simple calculation below is undiscounted and excludes tax and financing effects.

Consider a hypothetical services business receiving emailed work requests. Employees read each message, identify requirements, check missing details, and prepare an internal intake record.

The proposed AI workflow extracts information and prepares a draft. A person checks the result before it enters the operating process. Quoting, payments, and contractual commitments are outside the scope.

All figures below are illustrative assumptions, not Devine Globe client results, supplier prices, or promised savings.

The example assumes usable source data, a fixed maintenance bill, and no additional internal setup effort. Add that effort and any other startup costs to a real case. The three-minute handling assumption includes unsuccessful attempts and fallbacks, not only accepted drafts.

Calculate monthly effort and operating cost

Eligible, adopted requests: 2,000 × 60% × 80% = 960.

Released handling capacity: 960 × (8 − 3) ÷ 60 = 80 hours.

Assigned gross capacity value: 80 × $30 = $2,400.

Monthly operating resources:

• Variable usage: 960 × $0.25 = $240.
• External monitoring and maintenance: $300.
• Additional audit effort: 4 × $30 = $120.

Total monthly operating cost, including assigned internal audit time: $660.

The four audit hours are additional to per-case handling, so the net internal capacity released is 76 hours. After variable usage and external maintenance, the estimated net monthly value is $2,400 − $660 = $1,740.

Calculate the return over the first 12 operating months

Assume the base usage and costs apply from the first operating month, with no ramp-up or seasonal change.

This is a return calculated using the assigned value of released capacity. It is not a demonstrated cash return. The payback assumes that capacity has the stated value from launch; implementation time and a gradual rollout would change the calendar result.

If payroll is unchanged and no additional cash benefit is established, there is no demonstrated cash payback. Under this example’s assumptions, incremental cash spending over the same period is $6,000+12×($240 + $300) = $12,480. The $120 monthly internal audit allocation is excluded from that cash total because it does not create additional payroll spending here.

The next question is how the business will use the released capacity and whether that use justifies the spending. “We saved 80 hours” is an operating estimate, not the end of the business case.

There is no reliable universal percentage for your business. Results depend on task mix, adoption, review, errors, integration, and what changes economically after implementation.

The research paper Generative AI at Work studied an assistant used by 5,172 customer-support agents and reported an average 15% increase in issues resolved per hour, with substantial variation across workers. It studied a particular assistive workflow. That productivity result is not a forecast of your company’s cash savings.

For the intake example, change usage and net handling-time improvement while keeping the hourly value, fixed costs, and setup bill unchanged. Variable usage costs still scale with attempted requests.

The downside case has a negative monthly result even before recovering setup cost. The upside is a scenario to validate, not a sales promise. If higher usage requires more maintenance or auditing, increase those costs too.

At the base volume of 960 cases, covering the assumed $6,000 setup cost and $660 monthly operating cost over 12 operating months requires approximately 2.42 handling minutes saved per assisted request, using the assigned $30 hourly value. With an eight-minute baseline, average assisted handling needs to be about 5.58 minutes or less under those assumptions.

That threshold gives the pilot a practical question to answer: can the workflow sustain sufficient net time reduction while meeting its quality requirements?

Also test the valuation. If only half the base gross capacity value can be justified, the monthly net value falls from $1,740to**$540** with the same costs. More accurate extraction alone does not establish that employees can use every released minute productively.

HM Treasury’s appraisal guidance uses sensitivity analysis to test uncertain assumptions. Its public-sector appraisal rules are not being prescribed here; the same analytical habit makes this private-business screening exercise more useful.

Choose among candidates that pass the suitability gate. Compare conservative benefit, setup and operating burden, evidence quality, and the time required to validate the result.

For example, a hypothetical business considering three options might reach this decision:

This does not establish that inquiry processing always wins. It shows how scope, evidence, and implementation burden can change the choice.

Preserve a short decision card:

The card is the output of the 10-minute framework. It gives an implementation partner a problem to investigate and gives the business owner a basis for evaluating the proposal.

A useful pilot compares the current and proposed process on representative work. Include routine inputs, incomplete information, difficult cases, and relevant failures. Protect customer and employee information through the same permissions and handling requirements that apply to the intended workflow.

Agree on success criteria before interpreting the results. These should include acceptable quality, net human effort, cost, usage, and the response to failures. The threshold should reflect the business and consequences; there is no universal accuracy percentage that makes every workflow suitable.

Measure the whole process. A faster draft can still increase downstream correction or customer follow-up. Where practical, use a comparison group or a controlled rollout, and account for changes in workload, staffing, or task difficulty. Log workflow attempts rather than counting only completed successful runs.

Microsoft’s business-value measurement guidance includes adoption, response time, errors, cost, and user feedback. AWS’s production-value guidance likewise connects operating performance with ongoing business-value monitoring.

After the pilot, replace assumptions with observed inputs, record limitations, and decide whether to expand, revise, or stop. Continue measuring after launch: the input mix, tool costs, model behavior, and user adoption can change.

The best first project is one whose usefulness can be established through a bounded implementation and honest measurement. It should have an owner, a checked outcome, manageable failure consequences, and a credible route from improvement to business value.

Devine Globe Technologies provides AI strategy, workflow automation, and software integration. To discuss a potential first project, share the recurring task, its current volume, and the systems involved. Those details provide a concrete starting point for assessing scope, feasibility, and the evidence a pilot should produce.