The Worklytics charts in this article are example views from its AI adoption model, included to show what each measurement looks like in practice, not audited client benchmarks. The research figures cited (St. Louis Fed, NBER, HBS, METR and Anthropic) are independent primary sources.
Most leadership teams can tell you how much they spend on AI. Very few can tell you where those dollars turn into recovered hours. That gap is the whole problem. A seat count proves you bought access. A login proves someone was curious. Neither proves a quota-carrying rep finished her call prep faster or that an engineer shipped a feature a day sooner.
This article lays out a practical way to answer that question honestly. Measuring AI productivity is not about one adoption number. It is an efficiency map that shows where AI is actually returning time, team by team and task by task. We use the lens most executives reach for first, AI sales team effectiveness, and then apply the same method to the rest of the company. The Worklytics figures are illustrative, the external studies are not, and every layer of the map ties back to a specific measurement you can put in place.
The best current evidence comes from two kinds of studies: large surveys that ask workers how much time AI saves them, and field experiments that measure output directly.
Economists Alexander Bick, Adam Blandin and David Deming run a national survey that asks workers how many extra hours they would have needed without generative AI. In their November 2024 survey, users saved an average of 5.4% of their work hours, about 2.2 hours a week for a 40-hour worker. Spread across all workers, including people who do not use AI, that was 1.4% of total hours. Frequency matters: 33.5% of people who used AI every workday said it saved them four hours or more in a week, compared with 11.5% of people who used it on one day (St. Louis Fed).
The tracker is now published quarterly. Time saved across all employed adults rose from 1.8% of work hours in Q3 2025 to 2.1% in Q3 2026 (FRED: Generative AI time savings, employed adults). These are self-reported numbers, which is their main weakness.
The pattern across these studies: AI saves real time on tasks it handles well, the gains are largest for less experienced people, and self-reported savings can run ahead of measured ones. That is why a company needs its own measurement, not a borrowed average.
Most companies need method 2 to see where AI is used and method 3 to prove the time turned into output. The rest of this article shows how to put them together.
The claim gets repeated in board decks long before anyone checks it. The reason is that three separate things keep getting collapsed into one: access, activity, and outcome. A Copilot or ChatGPT license report answers access. A weekly active-user chart answers activity. Only a link between AI use and work output answers outcome, and that link is the one almost nobody puts in place.
The distribution matters more than the average. The Anthropic Economic Index, which analyzed over four million Claude conversations, found that AI use concentrates in a narrow set of tasks rather than spreading evenly across all work, with software development and writing together making up nearly half of all usage. If real usage clusters like that, a company-wide "40% adoption" headline is close to meaningless. It averages a team saving four hours a week against a team saving four minutes, and hides the very thing a leader needs to act on. The fix is resolution: measure at the level of team and task, which is what a productivity analytics platform is built to do.
An AI efficiency map is a team-by-team, task-by-task view of where AI returns time, in place of one company-wide adoption percentage. A single number is a vanity metric. A useful map is layered, and each layer answers a different question. Worklytics organizes this as a maturity progression that moves from counting who uses AI to proving what that use is worth.
Read from the bottom up, the map has four working layers:
Each layer only makes sense on top of the one below it. Measuring productivity without knowing proficiency tells you little, because you cannot separate a power user's results from an occasional user's. The rest of this piece walks each layer with the data that lives there.
Adoption is the floor of the map, and the first surprise usually lives here. When Worklytics breaks AI use down by function, the ranking rarely matches intuition. Customer Support and Engineering tend to sit near the top. Sales, Marketing, and HR often sit at the bottom. Worklytics' published 2025 adoption benchmarks by department put real ranges on each function, so you can compare your own floor against them.
That ordering carries a direct message for revenue leaders. Sales sits closest to the money yet often shows the smallest AI footprint, which flips the assumption that customer-facing teams adopt fastest. It also means the single largest lever on AI sales team effectiveness is not squeezing more from current users. The baseline is low, so it has room to move. A team in the low double digits has far more upside than one already past two-thirds.
You can only act on that if the adoption view is trustworthy and current. Worklytics builds this layer by connecting usage metadata (records of how often each tool is used, never the content of messages) from the AI tools already in your stack, including ChatGPT, Claude, Microsoft Copilot, Google Gemini, Slack, and Zoom, into one AI adoption dashboard instead of a stack of separate exports from each tool. Because it reads real usage rather than self-reported surveys, the map reflects what people do, not what they say in a training follow-up.
Adoption is a yes or no. Proficiency, meaning how well, and how often, someone uses a tool, is a spectrum, and that distinction is where most dashboards quietly mislead. A rep who pastes one prompt a month and a rep who runs two dozen AI actions a week both register as “active.” Treating them as equal guarantees you will misread the productivity layer above.
Worklytics separates them on two axes. The first is intensity, which sorts users into heavy, light and occasional tiers so the map can weight results by how much someone actually relies on the tool. The second is task type. Worklytics classifies AI activity into work categories such as coding, research, analysis, summarization, drafting and email creation, without storing the content of prompts.
This layer earns its place through inference, not decoration. Once you know that a sales team's AI activity is mostly email drafting and very little research, you can predict its productivity ceiling before you measure output, because drafting returns less time than analysis or preparation. Task-level proficiency is what turns a flat usage number into a diagnosis. Worklytics explains which adoption metrics to track in its guide to the AI adoption KPIs that matter, and the same measurement extends beyond AI to how teams collaborate, focus and communicate.
This is the layer executives actually want, and it is where the sales lens pays off. The pattern in Worklytics data is consistent: AI does not lift every activity equally. It compounds where it removes preparation and drafting friction ahead of a human conversation, which is exactly the shape of front-line sales work.
Start with outreach volume. In a representative benchmark, sales development reps who use AI heavily contact about 12.2 unique clients in a week, against 8.5 for reps who use no AI, roughly 30% more accounts reached.
The same shape holds in Customer Success. Reps who use AI heavily book about double the external customer meetings of peers who use none, roughly six a week versus three. The mechanism is not magic. AI absorbs the account research, recap drafting, and follow-up writing that used to sit between meetings, so more of the week is spent in front of customers instead of preparing to be.
The pattern extends past sales. In another representative benchmark, customer support reps who use AI heavily resolve about 80% more tickets a week, 10.5 versus 6, with the largest gains among newer staff. That skew is not unique to Worklytics. The peer-reviewed Generative AI at Work study found a 14% average productivity gain among support agents, rising to 34% for the least experienced. The through-line for revenue teams is the same. AI sales team effectiveness improves most where the tool clears the research, drafting, and preparation that sit in front of a customer conversation, and the reps who gain most are the ones still building their instincts.
Productivity gains are only real if you can price them. The top layer of the map turns task-level time savings into a financial figure with a simple, defensible method: estimate the typical time each task takes, multiply the hours saved by a fully-loaded hourly rate, and add it up.
The figure that should change a leader's behavior is the last one. The gap between value captured today and value available at fuller adoption reframes the conversation. The question stops being “is AI working” and becomes “how much return is stranded in the teams that have not adopted yet.” That is a management problem with a named owner, not a licensing question for procurement.
Impact is only honest when you also account for spend, which is why AI cost tracking belongs in this layer. Worklytics can export usage data into your data warehouse or reporting tool, where it joins each AI tool's billing exports so spend can be attributed to the team that generated it. Read that way, a team's productivity gain sits right next to what it costs to produce.
For the full method, from cost baselines to payback, see the guide to tracking and improving ROI from your AI investments.
Once the map exists, the next question is how to shift the low-adoption teams. The data gives a clear, slightly uncomfortable answer: adoption travels downhill from managers. In a representative sample, teams whose manager is a heavy AI user adopt AI at 37%, nearly double the 20.5% on teams led by a manager who rarely uses it.
The practical takeaway is that a company-wide enablement email is the weakest lever available, and manager behavior is the strongest. To raise AI sales team effectiveness, the fastest path is not another all-hands demo. It is getting sales managers to use AI in their own workflow, because their teams follow. Worklytics breaks AI adoption down by manager as well as by team and role, so you can see which managers' teams have stalled. For why the dynamic exists, see Worklytics on leadership's role in AI adoption. The same platform measures manager effectiveness more broadly, from how often managers hold one-on-ones to how much focus time their teams keep.
None of this works if employees experience it as surveillance, so how the map is built matters as much as what it shows. Worklytics runs on usage metadata, such as how often tools are used and by which team, rather than the content of prompts, messages or documents. Personal identifiers are pseudonymized (hashed with SHA-256) before analysis, and results are only shown for groups above a minimum size, typically 5 to 8 people.
This is the difference between workplace measurement that leaders can defend and monitoring that erodes the trust it depends on. Aggregate, de-identified records give you the resolution to find stranded value without turning the tool into a reason for people to hide their work.
The four layers are the plan. In practice, the sequence is straightforward:
Because the same Worklytics platform already measures employee productivity, engagement, employee well-being, meeting effectiveness, and manager effectiveness, the AI efficiency map does not sit in a silo. It plugs into the fuller picture of how work happens, which is what keeps AI investment accountable quarter after quarter.
Combine two measurements. First, track who uses AI, how often and on which kinds of work, from the tools' own usage data, and apply a time-saved assumption per task category. Then check the estimate by comparing output for high-use and low-use groups doing comparable work over the same period.
On average, less than headlines suggest. The St. Louis Fed's tracker estimates that generative AI saved about 2.1% of all US work hours in Q3 2026, and its late-2024 survey found users saved 5.4% of their hours, about 2.2 hours a week. Field experiments show much larger gains on tasks AI handles well, such as a 14% rise in issues resolved per hour for support agents.
Track coding assistant usage by team, such as active days, sessions and the share of pull requests that AI contributed to, and compare merged pull requests and review times for high-use and low-use groups. Pair output with quality signals such as review time, because a 2025 METR trial found experienced developers took 19% longer with AI while believing they were faster.
Use observed usage data, not surveys. Connect your AI tools to one view, break usage down by team, role and task category, and focus on the groups with the most usage and the clearest output measures. Worklytics does this from usage metadata and shows results only for groups above a minimum size.
Worklytics analyzes usage metadata, such as how often each tool is used and by which team, not the content of prompts, messages or files. Personal identifiers are pseudonymized before analysis, and results are only shown for groups above a minimum size, typically 5 to 8 people.
Ready to see where AI is saving time in your organization? Explore Worklytics to map AI adoption, engagement, productivity, and impact from the tools you already use.
