Baseline before you deploy anything

You cannot report hours reclaimed without knowing what the hours were. The baseline is a per-workflow time audit taken before agents touch anything: how many human minutes a unit of work costs today (per report produced, per lead triaged, per contract reviewed) multiplied by weekly volume. Capturing it as minutes-per-unit rather than gross hours matters, because volume grows. A workflow that doubles its throughput while holding human time flat is a win that a gross-hours comparison would report as zero.

This is a week-one job, and it is the single most common thing teams skip. Skip it and every later number is a guess dressed up as a metric. The baseline does not need to be exhaustive. The top five workflows by suspected time cost cover most of the value, and telemetry will tell you within a month whether you picked the right five. Timebox it: unit times from work history or a small sample of non-users, not a company-wide time-tracking mandate that dies in week two.

Three ways to measure reclaimed hours: use all of them

Telemetry is what actually happened: sessions, active days, tasks completed, exceptions escalated. Usage analytics from your AI platform are cheap to collect and impossible to argue with, but they measure engagement, not hours saved. Their real job is catching contradictions. A team self-reporting eight hours a week of savings on two sessions a week of usage is telling you about the survey, not the savings.

Cohort comparison is your best causal evidence. If part of the company uses agents and part does not, you have a natural control group: compare cycle time and throughput for matched roles on countable units of work. It is immune to self-report bias and it is what finally convinces a skeptical CFO. It needs about a quarter of data and only works where output is countable, which is exactly where you should concentrate measurement effort anyway.

Task-anchored surveys fill the gaps, with one hard rule: never ask “how much time does AI save you?” That answer inflates roughly two-fold. Ask about the last concrete instance (“the last contract you reviewed: how long did it take, and did you use the agent?”) and compute the delta against the baseline yourself. Five minutes, quarterly, a rotating sample of a third of your users. Where the three methods disagree, telemetry-weighted cohort data wins and the survey gets discounted.

Three worked examples: sprawl, saturation, and partial adoption

Case one: a 30-person company where every single person uses Claude, plus ChatGPT, Gemini, and a rotating cast of niche tools. This is tool sprawl, the earliest maturity stage: AI as personal productivity, not as a function. Telemetry is fragmented across vendors and mostly unavailable, there is no control group, and half the “agent operations” denominator is hidden in evaluation churn: hours spent trying, comparing, and re-prompting across tools. The measurement advantage is size: at 30 people the survey is a census, not a sample. Interview everyone, task-anchored, in a week. The AI BizOps job here is consolidation before optimization: find which tools earn their seat on reclaimed minutes, kill the rest, and pick the two or three workflows worth graduating from ad-hoc prompting to an owned, agent-run process. Leverage is measured per tool as much as per workflow, and the honest first report is usually “high engagement, thin leverage, too many denominators.”

Case two: a 100-person company where most people use Claude. Adoption is saturated on one platform, which changes what you can measure and what you should. There is no meaningful control group left (the handful of non-users are non-comparable by definition), so cohort comparison is off the table and your evidence rests on per-workflow baselines and telemetry, which a single-vendor stack actually makes tractable. The maturity question shifts too. The company is past “does AI help,” so leverage stops being an adoption argument and becomes an operations report: which workflows are actually owned by agents end to end, where residual review time is falling quarter over quarter, and which copilot-style usage is ready to graduate into agent-run workflows. The denominator is usually diffuse here (no dedicated team, ops time scattered across managers), and pinning it down, often to roughly one FTE-equivalent, is half the measurement work.

Case three: a 600-person company where 100 people use Claude. This is partial adoption, and the 500 non-users are the asset: cohort comparison becomes your primary instrument, and the leverage number carries a different job: it is the business case for the next 200 seats. Measure like an early-maturity function even if the users are sophisticated: baseline the top workflows first, run matched-cohort reads on countable output, and report per-segment leverage so the expansion argument names exactly which teams to onboard next and why. The ops denominator is usually legible here (a small named team runs the program), so the ratio is cleaner, and suppose it lands at 400 hours a week reclaimed against two FTEs of operations: 5x leverage, with coding workflows near 8x and ad-hoc chat near 2x.

Same metric, three maturities, three burdens of proof. The sprawl company measures consolidation (which tools and workflows deserve to survive) because focus is the question. The saturated company measures depth (residual minutes falling, workflows crossing from assisted to agent-owned) because expansion is no longer the question. The partial-adoption company measures spread (verified deltas between users and non-users) because expansion is the only question. Confusing them wastes the effort: cohort machinery produces nothing where everyone already uses AI, and depth-audits answer questions nobody asked where the real problem is thirty people running five tools each.

Validate reallocation, or say you did not

Reclaimed hours that turn into slack are not leverage. The metric promises strategic hours, so the reclaimed time has to show up somewhere: the deferred backlog project finally staffed, more discovery calls, faster shipping. At scale you cannot audit a hundred calendars, so validate by proxy: did the team’s output mix shift, and does a sample of managers confirm it?

Where you cannot confirm reallocation, report the hours on a separate line, labeled reclaimed with reallocation unconfirmed, rather than silently counting them. This split is not pedantry; it is what keeps the whole number trustworthy. The first time someone catches leverage inflated by hours that quietly became longer lunches, every future report is discounted, and the function loses the only asset it has: a number the business believes.

Report it like finance reports revenue

Monthly, per segment: the leverage ratio, the trend, and the top three workflows by hours reclaimed. Annotate methodology changes the way finance annotates restatements. If the survey instrument changed or a baseline was corrected, the report says so. Precision claims should be honest too: cohort-measured workflows can get tight, but ad-hoc chat usage will never be better than roughly ±30%, and pretending otherwise costs credibility you cannot buy back.

The quarter-one sequence: instrument telemetry in week one, baseline the top five workflows in weeks two through four, run the first survey wave in week four, and take the first cohort read at week twelve. Until then, report engagement and baselines only, and resist the pressure to report leverage early. One quarter of honest “not yet measurable” beats a year of numbers nobody trusts. The discipline is the point: AI BizOps earns its seat next to sales and finance by owning a number with the same rigor they own theirs.