Because the math is doing exactly what you told it to do: every saved hour priced at full loaded cost equals full dollar value. Real deployments don't work that way - some saved time is absorbed by Parkinson's Law, some gets re-spent reviewing AI output, and adoption never hits 100% on day one. Use the realization rate slider (60% is a good default) and add an implementation cost if you're rolling out at scale. Published benchmarks (Forrester TEI on Copilot, IDC enterprise AI surveys) cluster in the 100-400% range for first-year ROI - anything above that usually means the model is excluding real costs.

For typical knowledge work, 40-65% is defensible. Roles with highly structured workflows (drafting routine emails, summarizing meetings, formatting reports, light data entry) sit at the high end. Roles requiring deep judgment, original research, or face-to-face relationships sit at the low end. Engineering tends to be 50-75% on coding tasks but lower on architecture. Support tends to be 60-80% on tier-1 tickets but drops on complex escalations. If you find yourself entering 80%+ for a role, ask whether you're modeling what AI can do (capability) or what it will do for this specific team (adoption).

Three approaches: (1) Self-report - ask the team. People tend to underestimate. (2) Calendar audit - look at a week of work and tag tasks by automation potential. More accurate, more effort. (3) Benchmark - McKinsey research suggests knowledge workers spend 19% of their week on information gathering and 14% on communication, much of which is automatable. For most office roles, 8-15 hours per week is a credible starting estimate. Higher for content-heavy roles (marketing, research), lower for hands-on roles (sales calls, lab work).

2,080 hours/year (40 × 52) is the standard FTE basis used in nearly every published productivity ROI study, including Forrester TEI. We keep it fixed so your numbers stay comparable to those benchmarks. The "working weeks per year" setting only affects how many weeks of saved hours we count - it does not change the per-hour cost. If you'd prefer to divide loaded cost by actual productive hours (which would inflate value modestly), that's a defensible alternative but harder to defend to a CFO who's seen the standard methodology.

How many months of net benefit (value recovered minus tool cost) it takes to recoup one year of tool cost plus any one-time implementation cost. If net benefit is negative, payback shows "Never" - the tool will not pay for itself at current assumptions. If the tool is free (cost = $0) and value is positive, payback is "Immediate."

Presets reflect publicly listed 2026 list prices for the standard team / business tier of each tool. They do not include required base licenses - Microsoft 365 Copilot requires a qualifying Microsoft 365 plan ($12-57 extra per user); Gemini for Workspace requires Google Workspace ($6-18 extra per user). For an apples-to-apples comparison at your organization, use Custom and enter the true per-seat all-in cost from your procurement quote.

Notably missing from the model: (1) adoption ramp (year 1 typically delivers ~50% of steady-state benefit), (2) ongoing training and change management cost (often $50-200/user/year), (3) productivity dispersion across users (top quartile gets 3-4x the value of bottom quartile), (4) risk costs (data exposure, IP leakage, AI hallucinations in customer-facing outputs), (5) downstream revenue effects (faster sales cycles, better products). The first three depress the headline number; the last two can amplify it. For a defensible CFO conversation, lean conservative on inputs and acknowledge the rest qualitatively.