Count the full cost of an AI tool

An AI tool’s cost includes the subscription and the work around it. Count setup, training, checking, corrections, and upkeep, then compare those costs with the task you’re trying to improve.

Keep cash spending separate from the value assigned to staff time. A tool that gives someone an hour back may create useful capacity without reducing payroll. The free worksheet makes those categories visible instead of turning every saved minute into profit.

Define the task and the comparison period

Choose the job and a period suited to the decision. A first-year comparison can include setup and recurring work. A shorter pilot can establish whether the output is useful before you model a full year.

Record the current process and its observed effort. Include checking and error correction. Mark any estimate explicitly. Comparing a perfect demo with your most difficult manual case would give a misleading result.

Before modeling benefits, define the done checks. Anthropic’s prompting guidance recommends success criteria and empirical tests. A cheaper output that fails the task isn’t a useful saving.

Count cash and labor separately

Use current vendor terms for seat, usage, hosting, and connection costs. Save the plan, date, billing period, and any minimum commitment. Leave missing prices unknown rather than supplying a likely number.

Category	What to include	Evidence status
Recurring cash	Subscription, usage, hosting, and required connections	Current quote or plan, with date
One-time cash	External setup and purchases	Quote or actual expense
Setup and learning time	Internal configuration and training	Measured or estimated hours
Recurring effort	Review, corrections, monitoring, and maintenance	Comparable observed work
Exit effort	Export, migration, or replacement	Tested requirement or estimate

Use a loaded hourly rate only when you have an agreed basis for it. Otherwise report hours and cash separately. Don’t charge the same hour twice as both setup and learning.

Data requirements can change implementation effort. OpenAI’s API data documentation distinguishes training policy from retained logs and application state. Inspect the chosen service’s configuration instead of assuming the least expensive plan meets the task’s requirements.

Work through teaching arithmetic

This example uses fictional inputs to demonstrate the calculation, not vendor prices or customer results. Suppose a task takes 3 hands-on hours per month before a trial and 2 afterward, including review. The observed difference in that teaching scenario is 1 hour per month.

If the tool also has 4 setup hours, the first 4 months of that simplified capacity comparison would offset the setup hours, assuming the monthly difference stays the same. It wouldn’t establish cash payback, because no expense reduction was supplied.

Add actual recurring cash costs, and keep them alongside the hours. If the process requires more checking later, update the comparison. The worksheet’s instructions show formulas and a worked row without pretending those teaching values are typical.

Evaluate benefit without forecasting revenue

Use measured changes where you can: task time, corrections, error count, or an expense that genuinely changed. Record the observation period and comparable workload. Keep proposed benefits in an assumptions column until tested.

OpenAI’s evaluation guidance recommends tests tied to the task. Repeating representative examples helps you check whether an apparent improvement persists when inputs change.

Revenue belongs in the comparison only when you have a defensible observed basis. An extra article, a draft, or a freed hour doesn’t automatically create a sale. Don’t add speculative income merely to produce a favorable ROI.

If net cash benefit is zero or negative, cash payback isn’t established. Leave it undefined instead of dividing by zero or labeling a capacity estimate as cash. If you do calculate payback from a positive supported monthly net cash benefit, disclose the inputs and assumptions. There isn’t a universal acceptable number of months for every business.

Include maintenance and failure costs

Account changes, new data, and broken connections create work after launch. Name who handles that work and whether it is included in the current engagement. The NIST AI Risk Management Framework treats evaluation and risk management as ongoing concerns, rather than a one-time installation.

Record the fallback when the tool fails and the consequence of a bad result. A task with an easy manual fallback has different costs from an external action that needs correction. The automation guide helps specify that recovery work.

Make a bounded decision

Complete the cash inputs, hours, source dates, and unknowns. Decide whether to continue the pilot, adopt the tool, change the process, or keep the manual method. Use the tool-comparison guide if another option deserves the same test.

You can prepare this comparison in any spreadsheet with the free files. If a specific marketing system needs to be built and handed over, Custom AI Builds can scope it before work starts. Bring the measured task and assumptions so the quote responds to an actual job.

Provider guidance checked October 8, 2026.