AI Automation Cost and ROI: A Baseline-to-Decision Worksheet

Calculate AI automation ROI from an observed workflow baseline, the full cost of ownership, and a bounded benefit range—not from vendor time-saving claims. Separate recoverable labor capacity, avoided error cost, incremental contribution, and risk reduction. Record low, expected, and high cases with their assumptions. If the result depends on unmeasured volume, perfect adoption, unrestricted autonomy, or benefits the organization cannot capture, the investment case fails.

Start with a decision statement

Write one sentence before opening the worksheet:

We are deciding whether to change this workflow for this group, over this decision horizon, because this observed baseline problem may justify this operating investment.

The blanks matter. “Adopt AI” is not a decision statement. A specific example might concern request triage, document extraction, follow-up preparation, or reconciliation between systems, but the worksheet should use the organization’s actual process and records.

Name the comparison options as well: keep the current process, remove a step, configure an existing product, add deterministic automation, or add a bounded AI component. ROI is a comparison, not a property of the technology by itself.

Worksheet part 1: capture the baseline

Use a representative measurement window and keep the raw counts. Do not begin by asking employees how much time they think AI will save. Record how work behaves now.

Use a fully loaded internal cost rate only if the organization can explain what it includes and who supplied it. Keep time units consistent. If salary or staffing data is sensitive, finance can provide an approved aggregate rate rather than exposing personal compensation.

A useful baseline formula is:

baseline handling cost = case volume × active handling time per case × approved cost per unit of time

That formula is not the whole business case. Waiting time may matter to customers without creating labor cost. Some employees perform valuable work during the delay. Some handling time cannot be removed because review remains necessary.

The methodology page explains why baselines, measurement windows, and confounding changes should accompany any later outcome claim.

Worksheet part 2: include the full cost of ownership

Implementation is only one cost category. Build the decision horizon in periods meaningful to finance and list each cost with an owner and source.

Initial costs

• workflow discovery and process documentation;
• data inventory, cleanup, labeling, and migration;
• product configuration or custom application work;
• API and system integration;
• identity, permissions, security, privacy, and legal review;
• evaluation-set creation and acceptance testing;
• training, documentation, and change management;
• parallel operation during the pilot; and
• decommissioning or transition work for the old process.

Recurring costs

• product licenses, model or API usage, infrastructure, and storage;
• connector subscriptions and data-provider charges;
• human review that remains in the target process;
• monitoring, evaluation runs, alerting, and incident response;
• content, policy, and source maintenance;
• model, prompt, workflow, and integration changes;
• security updates, access reviews, backup, and recovery testing;
• support coverage and vendor management; and
• audit, reporting, and retention obligations.

Risk and exit costs

• downtime and manual fallback capacity;
• incorrect actions, messages, or records that escape review;
• duplicate or missed transactions during retries;
• vendor price, availability, or product changes;
• data export and migration if the product is replaced; and
• time required to pause, investigate, correct, and notify after an incident.

The NIST AI Risk Management Framework treats governance, mapping, measurement, and management as continuing work. NIST’s Generative AI Profile extends that risk perspective to generative systems. Those activities belong in operating cost when the proposal depends on AI.

Add the categories using the organization’s finance convention:

total cost of ownership = initial cost + recurring cost over the decision horizon + internal operating time + expected risk and exit allowance

Do not hide internal time because no invoice is issued. It remains capacity consumed by the project.

Worksheet part 3: calculate only benefits the organization can capture

Separate benefit mechanisms. Combining everything into one savings figure makes weak assumptions hard to find.

Recoverable capacity

Estimate the active minutes removed from eligible cases, then reduce the result for exceptions, review, correction, and adoption. Add a capture factor that answers a practical question: what will the organization actually do with the released capacity?

capacity benefit = eligible volume × net time removed per case × capture factor × approved cost per unit of time

A capture factor should not default to complete conversion. Time saved in small fragments may improve responsiveness without reducing paid hours. That can still be valuable, but it should be described as service capacity rather than booked as cash savings.

Avoided error and rework

Define the exact error first. Then use observed frequency and remediation cost.

avoided error benefit = baseline error count × expected reduction × average verified consequence per error

Use a range when rare errors have very different consequences. Do not multiply a speculative error rate by the worst incident and call the result expected value.

Incremental contribution

Additional completed sales, bookings, renewals, or collections may create benefit, but revenue is not profit and attribution is difficult.

incremental contribution = attributable additional completions × contribution per completion

The measurement plan needs a comparison group, pre/post design, or other defensible counterfactual where practical. Conversation count, generated drafts, clicks, and model confidence are not incremental contribution.

Risk reduction and service quality

Faster ownership, clearer audit trails, improved recovery, and more consistent policy use may be worthwhile even when they do not reduce payroll. Record these separately with an observable measure. Do not assign a currency value merely to force every benefit into the same cell.

Worksheet part 4: make uncertainty visible

For every material input, record:

• the source and measurement period;
• who owns the number;
• whether it was measured, quoted, modeled, or assumed;
• a low, expected, and high value grounded in available evidence;
• what could make it change; and
• when it will be checked again.

Run the calculation for at least the low and expected cases. The low case should not be a token discount from the preferred answer; it should reflect plausible adoption, integration, quality, and maintenance problems.

Then run sensitivity analysis. Change one input at a time and identify which assumptions control the result. Common swing variables include eligible volume, review rate, adoption, connector work, error reduction, capture of released time, support coverage, and ongoing maintenance.

If one weak assumption determines whether the case is positive or negative, that assumption becomes a pilot question. It is not a reason to average the uncertainty away.

Worksheet part 5: compute the decision metrics

Use the finance definitions already accepted by the organization. At minimum, show:

net benefit = total captured benefit − total cost of ownership

ROI = net benefit ÷ total cost of ownership

payback point = first period when cumulative captured benefit exceeds cumulative cost

Keep cash, capacity, risk, and service-quality outcomes visible rather than forcing them into one blended percentage. A project can have a credible operational benefit and still fail an investment threshold. It can also have a positive modeled ROI and remain unacceptable because a high-consequence failure is insufficiently controlled.

If finance uses discounted cash flow, hurdle rates, depreciation, or a required decision horizon, use those rules. The worksheet should fit the organization’s capital process rather than invent a parallel one for AI.

Add the cost of authority

An automation that drafts a recommendation is not the same investment as one that changes money, access, customer commitments, or business records. Greater authority increases the need for authentication, authorization, validation, approvals, audit logs, idempotency, reconciliation, monitoring, and recovery.

OpenAI’s function-calling documentation illustrates a model proposing calls to application-defined tools. The application still owns permission and execution. Those controls, test cases, and staff interfaces are part of total cost.

User messages, documents, and tool outputs are also untrusted input. OWASP’s prompt-injection guidance explains a central failure mode in systems that let models interpret external content. Limiting tools, validating actions, isolating secrets, and reviewing consequential steps add real engineering and operating work. Omitting them does not make the system cheaper; it makes the estimate incomplete.

Explicit fit and non-fit criteria

The investment case may fit when:

• the workflow is repeated or consequential enough to measure;
• baseline records are usable;
• the proposed benefit has a named capture mechanism;
• existing configuration and deterministic options were considered;
• integration access is confirmed;
• a process owner will maintain the system;
• acceptance criteria include quality and failure recovery; and
• the low case remains acceptable or the pilot can resolve its key uncertainty.

The case fails when:

• volume, handling time, or error cost is guessed and materially drives the result;
• all saved time is counted as cash with no capacity plan;
• projected revenue has no defensible attribution method;
• benefits depend on perfect adoption or no exceptions;
• the product cannot access the required systems under supported terms;
• review, maintenance, security, and fallback are excluded;
• the organization cannot define a correct result;
• a high-consequence action lacks an affordable control; or
• a clearer process, configured feature, or deterministic workflow provides the benefit with less ownership.

The frequently asked questions explain how we approach ownership, security, integrations, and project scope. Our business automation service covers deterministic workflow changes; AI agent development applies when context-dependent tool use is genuinely required; and data intelligence may be the earlier need when no trustworthy baseline exists.

Turn the worksheet into a pilot contract

A pilot should resolve the few assumptions that dominate the decision. Define the eligible case set, comparison period, review rules, acceptance threshold, failure stop, data owner, and decision date before launch. Measure correct completion, human correction, exception rate, operating cost, and recovery—not only model output quality.

At the end, replace assumed cells with observed values and rerun the low and expected cases. Continue only if the operating team can own the process and the decision still clears its agreed thresholds. Book an AI automation baseline and ROI working session.