How one can Measure Generative AI Adoption within the Office

Introduction

Studying methods to measure generative AI adoption within the office has change into a board-level query as a result of utilization is climbing quicker than most corporations can rely it. The St. Louis Fed discovered that 39.2% of employed adults used generative AI for work up to now week through the second quarter of 2026. That share was 28.2% within the third quarter of 2024, so weekly use rose by about eleven factors in beneath two years. A license rely or a login whole says virtually nothing about whether or not these folks use the instruments nicely, how typically they return, or whether or not their work improves. Leaders who report adoption from seat counts alone find yourself defending a quantity that can’t survive a single follow-up query from a finance companion. This information builds a extra trustworthy image in layers, protecting attain, frequency, depth, and outcomes, and it ends with an eight-step tutorial that features working SQL and Python. Alongside the best way it covers survey design, shadow AI, worker privateness, and the cohort-level diagnostics that separate wholesome adoption from a shallow plateau. Readers who need background on the know-how itself can begin with our overview of what generative AI is earlier than returning right here.

Fast Solutions on How one can Measure Generative AI Adoption at Work

How do you measure generative AI adoption at work?

To learn to measure generative AI adoption within the office, observe 4 layers: attain, frequency, depth, and outcomes. Pull utilization telemetry from every AI software’s admin console, then verify what the logs can’t present with a brief survey.

Which generative AI adoption metrics matter most?

Weekly lively price amongst licensed customers, return frequency, the share of multi-step classes, and value per lively consumer matter most. License totals and one-time logins overstate adoption as a result of they ignore repeat conduct and depth of use.

What’s a traditional generative AI adoption price for workers?

Benchmarks range by function and survey. Gallup discovered 15% of U.S. staff use AI each day and 30% use it a couple of occasions every week or extra in Could 2026, so set targets by job household.

Key Takeaways

• Adoption is a conduct change, so measure attain, frequency, depth, and outcomes collectively as a substitute of counting licenses or logins.
• Mix admin telemetry with a brief pulse survey; telemetry reveals what folks do whereas surveys clarify why they do it.
• Report adoption by staff and cohort slightly than by particular person, and respect minimal group sizes to guard worker belief.
• Deal with shadow AI as a visibility hole to shut, as a result of unmeasured instruments cover each danger and essentially the most motivated customers.

What Is Generative AI Adoption within the Office?

How one can measure generative AI adoption within the office begins with defining adoption itself: the share of staff who use generative AI instruments frequently, deeply, and productively of their actual work, slightly than merely holding a license or logging in as soon as.

Why Seat Counts and Login Totals Mislead Leaders

Most organizations purchase generative AI seats earlier than they agree on what success appears to be like like, which leaves the primary dashboard dominated by procurement information. A seat rely tells you what number of licenses finance authorised, and a login whole tells you the way many individuals opened a software at the least as soon as. Neither quantity describes conduct change, and conduct change is what adoption really means. A evaluation of adoption analysis by IntuitionLabs cites SHRM information exhibiting that 56% of HR professionals don’t formally measure the success of their AI investments. That hole explains why so many govt updates depend on anecdotes, vendor slides, and license utilization figures that no person has audited.

Breadth of entry additionally hides a steep drop in depth, and the very best latest proof comes from a paper on adoption telemetry. The writer studies that ActivTrak information present 82% of AI customers maintain utilization quarter over quarter, but solely about 2% attain constantly embedded use. Learn that sentence twice, as a result of it means a dashboard that celebrates retention can look wholesome whereas virtually no person has modified how they work. The identical paper reveals an artificial cohort by which 99.4% of customers attain recurring use and the cohort nonetheless stalls, as a result of depth by no means develops. A single headline share can’t reveal that sample, so any trustworthy measurement program wants at the least two or three complementary views.

Login counts fail for a 3rd purpose, which is that they deal with each session as equal. An individual who opens a chatbot as soon as 1 / 4 and an individual who runs a dozen multi-step analysis classes each week each register as lively customers in a naive report. Staff additionally unfold their utilization throughout a number of instruments, so a single-vendor dashboard undercounts the individuals who want a distinct assistant. The denominator issues simply as a lot because the numerator, since reporting towards licensed seats, whole headcount, and eligible roles produces three very completely different percentages. Determine which denominator you’ll use earlier than the primary report goes out, and preserve it mounted in order that traits stay comparable from one quarter to the subsequent.

A 4-Layer Mannequin for Studying Adoption

A layered mannequin retains a measurement program from collapsing right into a single vainness quantity, and it provides every stakeholder a query they’ll reply. The 4 layers used all through this information are attain, frequency, depth, and outcomes. Attain asks who has entry and who has really tried the software, whereas frequency asks how typically these folks return. Depth asks how refined their use is, and outcomes asks whether or not the work improved. Every layer builds on the one beneath it, so a weak point at a decrease layer normally explains a weak point above it. Whenever you learn the layers collectively, a dip in outcomes could be traced again to the precise layer the place folks stopped progressing.

The mannequin maps onto three questions that Worklytics makes use of to arrange its AI adoption KPIs: is AI getting used, is it used nicely, and is it value it. Attain and frequency reply the primary query, depth solutions the second, and outcomes reply the third. Putting your metrics beneath these headings stops groups from arguing about whether or not a given quantity belongs on the dashboard. It additionally reveals the place you don’t have any information in any respect, which is widespread for depth and virtually common for outcomes. Most corporations begin with wonderful attain information and virtually nothing else.

Attain deserves a exact definition as a result of each later metric inherits its denominator. Activation price is the share of licensed or eligible staff who’ve run at the least one qualifying session within the final twenty-eight days. Time to first use measures the times between license project and that first session, and lengthy delays level to an entry or consciousness downside. Attain is a precondition for every little thing else, however additionally it is the layer the place leaders most frequently declare victory too early. Frequency then measures return conduct by weekly lively price, the ratio of weekly to month-to-month customers, and early return inside a fortnight.

Depth and outcomes spherical out the image and are more durable to instrument, which is precisely why they’re definitely worth the effort. Depth captures what number of distinct duties somebody brings to the software, how lengthy and multi-step their classes run, and whether or not they use superior options comparable to brokers or file evaluation. Outcomes translate that utilization into cycle time, high quality, error charges, and value, normally by evaluating groups or intervals. For steerage on scaling packages as soon as the measurements are in place, see these 4 methods for scaling generative AI throughout a corporation. Every layer wants its personal proprietor, its personal information supply, and its personal evaluation cadence, in any other case the dashboard turns into ornament.

What the Newest Workforce Surveys Present About Utilization

Constructing on that four-layer mannequin, it helps to anchor your inner numbers towards what nationwide surveys report, as a result of executives will ask the way you examine. The St. Louis Fed tracker places weekly use amongst employed adults at 39.2% within the second quarter of 2026, with generative AI helping 6.3% of all work hours. The identical tracker estimates that AI saved about 2.2% of labor hours, up from 1.6% within the third quarter of 2024. These time financial savings are self-reported, so the Fed treats them as approximate, and it is best to apply the identical warning to any inner survey. The development nonetheless issues as a result of it reveals utilization rising quicker than measured time financial savings.

Frequency tells a extra sober story than the weekly headline suggests. Gallup’s office indicator reveals that 15% of U.S. staff used AI each day as of Could 2026, whereas 30% used it a couple of occasions every week or extra. Solely 52% reported utilizing it at the least a couple of occasions a yr, which implies roughly half of staff stay occasional or non-users even now. The hole between weekly use and each day use is the only most helpful benchmark for judging whether or not your personal program is shallow or deep. If your organization studies 70% weekly use however 5% each day use, you most likely have a attain story and never but a behavior story.

Variation inside comparable jobs is the opposite lesson, and it ought to form the way you section your personal information. A abstract of an NBER working paper by the Data Know-how and Innovation Basis studies that greater than 20% of staff use generative AI in 4 of 5 detailed occupations. But just one in six occupations exceeds a 70% adoption price, so the everyday job sits nicely under saturation. The paper additionally finds that staff doing very comparable duties differ considerably in how a lot they use the instruments. For measurement, which means averages by division cover the individuals who matter most, specifically the heavy customers who can coach friends and the non-users who want entry or coaching. Phase by function and activity, not simply by org chart.

Constructing on Telemetry: The Information Sources Price Connecting

Constructing on these survey benchmarks, the sensible query turns into the place your personal utilization information lives and the way reliable every supply is. Most enterprise AI merchandise ship an admin console with mixture utilization studies, and the higher ones additionally expose an API or export. Microsoft 365 Copilot presents utilization studies over rolling home windows of seven, 28, 90, and 180 days. Its Graph report endpoint returns per-user element however remains to be in beta. The Viva Insights Copilot Dashboard provides adoption traits by group and hides any group smaller than 5 folks. Our information to utilizing Microsoft Copilot explains the product floor if you’re new to it.

Developer instruments expose richer telemetry, they usually make a helpful mannequin for what good instrumentation appears to be like like. GitHub paperwork its Copilot utilization metrics as IDE lively customers, each day and weekly lively customers, requests per chat mode, and agent adoption. The dashboard refreshes each day with information as much as three days behind, and the API could be filtered by day or by consumer login to seek out newly activated customers. No matter your toolset, insist on a user-level export so that you could compute your personal metrics as a substitute of trusting a vendor’s definition of lively. Complement vendor consoles with id supplier logs, community proxy information, and expense studies, as a result of these sources reveal instruments your official dashboards by no means see.

Attain and Frequency Metrics That Maintain Up

Shifting from sources to calculations, begin with the handful of attain and frequency metrics which might be low cost to compute and laborious to sport. Activation price counts licensed customers with at the least one qualifying session within the final twenty-eight days, divided by all licensed customers. Weekly lively price counts distinct customers lively in a seven-day window, divided by the identical denominator. Stickiness divides weekly lively customers by month-to-month lively customers, so a worth close to one means the identical folks return continually and a low worth means informal, sporadic use. Retention tracks the share of a first-week cohort that’s nonetheless lively 4 and eight weeks later.

Benchmarks for these metrics exist, however most of them are weaker than they appear. A vendor information from Larridin proposes that weekly lively utilization above 70% is robust and under 40% wants intervention, but it presents no supply or methodology for these thresholds. Deal with such numbers as hypotheses to check towards your personal historical past slightly than as targets to chase. A development that rises steadily for eight weeks is extra persuasive proof than a snapshot that occurs to clear another person’s threshold. Report your personal baseline within the first month, then choose each later studying towards it.

Firm context adjustments what quantity appears to be like like, so keep away from copying a price from a agency of a distinct dimension or sector. Adoption patterns differ sharply between small corporations and enormous enterprises, a spot explored in our evaluation of the small enterprise versus enterprise adoption hole. Position combine issues too, as a result of analysts and writers use assistants extra naturally than area technicians. At all times publish the denominator beside the proportion, for instance 61% of licensed data staff slightly than a naked 61% adoption. That one behavior prevents a lot of the arguments that erupt when a quantity travels from a dashboard right into a slide.

Depth, Breadth, and the Shallow Plateau

Past attain and frequency lies depth, the layer that separates enthusiasm from actual workflow change. The adoption telemetry paper proposes 5 levels: discover, try, navigate, rework, and embed. A consumer navigates when exercise seems in at the least three distinct weeks. A consumer transforms when sustained multi-step work spans six or extra lively weeks, and embeds when exercise continues in at the least 90% of weeks since first use. The paper evaluates cohorts slightly than people, and that design selection is value borrowing as a result of it avoids scoring any single worker. Stage counts provide you with a funnel, and the funnel reveals precisely the place your folks stall.

The commonest failure sample is the shallow plateau, by which retention appears to be like wholesome however depth by no means grows. Within the paper’s artificial information, a shallow-plateau cohort locations 99.4% of customers on the navigate stage and carries a composite rating of 49.9. A wholesome cohort scores 60.5 with 25.1% of customers on the rework or embed levels. Breadth alone can’t inform these two cohorts aside, since their weekly lively charges look comparable. Solely a depth measure, such because the share of classes that run twelve or extra turns, exposes the distinction. The prescribed response additionally differs, as a result of a shallow plateau requires workflow redesign and never for extra licenses or extra generic coaching.

Different failure signatures deserve a spot in your diagnostic guidelines. Champion dependency seems when exercise concentrates in a couple of customers, which the paper flags when the Gini coefficient of utilization exceeds 0.5. Utilization regression seems when late-window exercise falls under 40% of early-window exercise, an indication that novelty pale with out reinforcement. Low activity success seems when customers try complicated work and fail at it, which normally signifies a capability hole slightly than an consciousness hole. Worklytics provides manager-facing views of breadth, that means distinct use circumstances per staff, and of time to proficiency, that means days from first use to regular use.

Deal with these depth metrics as diagnostics, not targets, as a result of each rating could be gamed as soon as folks know it’s watched. The paper itself warns about Goodhart danger, noting that changing scores into targets might degrade their that means. A staff instructed to hit twelve-turn classes will fortunately pad its conversations with filler. Pair every depth metric with an final result test, comparable to activity success or reviewer scores, in order that inflated exercise can’t go for progress. The analysis on why pilots stall in massive organizations, summarized in our piece on why AI pilots fail to scale, factors to the identical lesson.

Measuring Outcomes With out Fooling Your self

Transferring on to outcomes, the outcomes layer is the place most measurement packages both earn govt belief or lose it completely. The core downside is attribution, as a result of a staff that adopts generative AI not often adjustments just one variable. Headcount, seasonality, product launches, and tooling upgrades all transfer the identical cycle-time metrics that you just hope to credit score to the assistant. With out a baseline and a comparability group, any enchancment you report is a narrative, not a measurement. Seize at the least eight weeks of pre-rollout information for every metric, and the place attainable stagger the rollout in order that late adopters function a pure management group. The information to measuring ROI on AI investments on this website walks by the monetary framing in additional element.

Select a small set of final result measures that match how every operate already defines good work. Help groups would possibly observe decision time and reopen price, engineering groups would possibly observe pull request cycle time and defect escapes, and gross sales groups would possibly observe proposal turnaround and win price. Pair each pace metric with a top quality metric, as a result of quicker output that wants heavy rework just isn’t a acquire. Value issues as nicely, and value per lively consumer, which divides whole AI spend by individuals who really used the instruments, ought to fall as adoption deepens. For a broader view of translating utilization into enterprise outcomes, examine delivering actual worth with generative AI throughout capabilities.

Surveys and Interviews as a Test on Telemetry

Stepping again from the logs, do not forget that telemetry reveals what folks do however by no means why they do it. A pulse survey of 5 to seven questions, despatched each quarter, fills that hole at very low price. Helpful objects ask how typically folks use AI, which duties they use it for, how assured they really feel, what blocks them, and the way a lot time it saves. Preserve the wording behavioral, comparable to “Within the final week, what number of days did you utilize AI for work?” as a result of attitudes drift away from habits. Nameless responses are usually extra candid, however it is best to nonetheless tag the division and tenure band so that you could section outcomes.

Self-reported numbers want cautious dealing with as a result of folks misremember and since some want to please the sponsor. The St. Louis Fed notes that its time-savings estimates are self-reported and due to this fact approximate, which is the correct perspective for any inner survey as nicely. Use surveys to rank issues and generate hypotheses, then confirm the necessary ones towards telemetry or work product. Perceived utilization typically runs forward of logged utilization in groups that take a look at a number of instruments, and behind it in groups utilizing embedded options unknowingly. Massive gaps between the 2 views are themselves a discovering value investigating.

Brief interviews add the feel that dashboards lack, and 5 conversations typically reveal greater than 5 hundred survey responses. Ask essentially the most lively customers to show an actual activity from begin to end, and ask non-users what must change for them to attempt it. Hear for course of obstacles, comparable to unclear information guidelines, awkward approvals, or lacking integrations, as a result of these not often seem in a log. The identical train exposes inequity, since entry, language, and confidence are usually not evenly distributed throughout a workforce, a theme explored in our take a look at the rising AI office divide. Feed what you be taught again into coaching and tooling selections shortly, as a result of folks cease answering surveys that change nothing.

Shadow AI Dangers and the Blind Spots in Your Information

Turning to the instruments you can’t see, each adoption dashboard has a blind spot known as shadow AI, that means AI instruments that staff use with out approval. A 2025 survey by ManageEngine and Censuswide discovered that 93% of staff admit coming into info into AI instruments with out approval. The identical examine studies that 85% of IT resolution makers say staff undertake instruments quicker than IT can assess them. Though 91% of these resolution makers have some AI coverage, solely 54% have clear, enforced insurance policies with lively monitoring for unauthorized use. Shadow AI is greatest understood as a measurement hole first and a coverage downside second, since you can’t govern exercise you can’t see.

You possibly can shrink the blind spot with strategies that don’t require studying anybody’s prompts. Overview safe net gateway or DNS logs for domains in recognized AI classes, audit single sign-on grants and OAuth consents for AI functions, and scan expense studies for subscriptions. Then compute a shadow AI share, outlined as exercise on unsanctioned AI domains divided by all AI exercise, and observe it by staff over time. A falling share means that authorised instruments are assembly wants, whereas a rising share suggests the other. Our overview of the 2025 shadow AI risk covers the safety aspect of the identical downside.

Privateness, Belief, and the Ethics of Utilization Monitoring

Given the sensitivity of utilization information, the ethics of measurement deserve as a lot design effort because the metrics themselves. Staff moderately fear that adoption information will change into a efficiency rating, a layoff record, or a software for monitoring what they kind. In the event that they imagine that, they’ll route round official instruments, and your numbers will quietly degrade. Belief is due to this fact a precondition for correct measurement and never a pleasant additional. State the aim of this system in plain language earlier than gathering a single information level, and write down what you’ll by no means do with the information.

Concrete safeguards make that promise credible to skeptical staff and works councils. Worklytics’ information to Microsoft Copilot reporting notes that admin heart studies cover usernames and group names by default and that the Copilot Dashboard suppresses teams smaller than 5 folks. Undertake the identical guidelines in your personal analytics: report on cohorts, apply a minimal group dimension, and like anonymized identifiers in exports. By no means retailer verbatim prompts for adoption analytics when immediate size and phrase counts reply the identical query. Test native labor guidelines and works council agreements earlier than you reveal identities, since some jurisdictions require session or consent.

Equity is the second moral thread, as a result of adoption is formed by entry and confidence as a lot as by motivation. Low utilization in a single group might mirror a lacking language mannequin, an inaccessible interface, or a supervisor who discourages experimentation. Measuring these teams as laggards, and rewarding or penalizing them accordingly, would deepen the very divide you hope to shut. A greater strategy is to deal with low adoption as a sign to research limitations and provide assist. Pair each dashboard with an express assertion that adoption information isn’t an enter to particular person efficiency critiques.

Governance ties these commitments collectively and retains them from eroding as leaders change. Publish a brief measurement constitution that names the information collected, the retention interval, who can see it, and the way staff can ask questions. Overview it everytime you add a software or a knowledge supply, since every addition adjustments the privateness floor. Our information to constructing an AI utilization coverage for workers reveals methods to write the acceptable-use guidelines that sit alongside the constitution. Insurance policies that staff helped form are way more more likely to be adopted than insurance policies handed down with out rationalization.

Placing Adoption Information to Work in Opinions and Budgets

On high of the dashboards, adoption information solely creates worth when it adjustments selections, so construct an everyday working rhythm round it. A month-to-month staff evaluation ought to take a look at attain, frequency, and depth for every cohort and identify one motion per cohort. A quarterly govt evaluation ought to present traits, the most important stall level, and the result proof collected up to now. Deal with the entire apply of methods to measure generative AI adoption within the office as an working course of with homeowners and deadlines, not as a one-off analysis challenge. With out that cadence, the very best dashboard within the firm turns into a bookmark no person opens.

Let the identified stall level resolve the intervention, as a result of the improper repair wastes finances. The telemetry paper maps every stall to a distinct response. A stall at discover requires visibility and entry as a substitute of coaching, whereas a shallow plateau requires workflow redesign as a substitute of extra licenses. Its stage mannequin treats every stall as a prognosis, so the correct repair at all times will depend on the place the cohort really stalled. Use price per lively consumer and license utilization to reclaim idle seats and redirect that cash towards the cohorts which might be able to go deeper. The broader price levers are lined in our information to enterprise AI price optimization methods.

The Way forward for Adoption Measurement: Brokers and Activity-Stage Information

Wanting forward, the unit of measurement is shifting from the chat session to the duty, and from the human immediate to the agent run. The NBER paper behind the ITIF abstract builds the primary task-level generative AI adoption indexes from a consultant survey. It finds that adoption is vast however shallow, with fewer than half of staff adopting in most occupations and duties. It additionally reveals that indexes constructed from chat logs differ conceptually from survey measures and have a tendency to over-classify conversations into generic actions. Anticipate distributors so as to add task-type fields, final result flags, and acceptance indicators to their telemetry in response. Till then, complement uncooked utilization with a brief activity stock for every function.

Brokers change the arithmetic of adoption as a result of one individual can now set off many automated actions with out typing a single immediate. Metrics such because the ratio of agentic to interactive utilization, spend velocity, and activity completion price will be part of weekly lively price on the dashboard. When software program acts on behalf of staff, the related query shifts from whether or not folks use AI as to whether the work it does is appropriate, protected, and value its price. Groups are already constructing scorecards for that query, and our walkthrough on methods to measure AI agent efficiency is an effective place to begin. Disposition indicators, comparable to how typically a human accepts, edits, or overrides an agent output, are more likely to change into essentially the most informative depth metric of all.

Measurement itself may even face regulatory and market stress as disclosure expectations develop. Boards, auditors, and prospects more and more ask how AI is ruled and what proof helps claims of productiveness positive factors. Organizations with clear cohort-level telemetry, documented definitions, and privateness safeguards will reply these questions in days, whereas others will scramble for months. Construct the inspiration for methods to measure generative AI adoption within the office now, with easy, well-defined metrics you can clarify on one slide. The eight-step methodology that follows is designed to supply precisely that basis, and it really works with any mixture of instruments.

How one can Measure Generative AI Adoption within the Office in Eight Steps

Subsequent, the eight steps under flip the four-layer mannequin right into a working measurement pipeline {that a} small analytics staff can construct in a few month. The pipeline assumes you’ll be able to export utilization information out of your AI instruments as CSV recordsdata, and it makes use of SQLite and Python in order that no particular platform is required. You possibly can swap in your warehouse of selection later with out altering the logic. Every step produces an artifact that the subsequent step consumes, which retains the challenge auditable and simple handy over. The code examples are intentionally small, so adapt the column names to no matter your distributors really export.

Step 1 – Outline adoption outcomes and the denominator

Begin with a one-page definition that names the result you need, the individuals who rely, and the interval you’ll measure. An final result may be that at the least 50% of licensed data staff use an authorised AI software weekly, and that 25% of them run multi-step classes. Write down whether or not your denominator is licensed customers, eligible roles, or whole headcount, and by no means change it mid-year. Agree on a qualifying session, comparable to any chat or agent run of at the least one flip, in order that opening the appliance doesn’t rely. Identify an govt proprietor and a knowledge proprietor, and think about using the method in our information to establishing an AI governance board to ratify the definition.

Subsequent, translate the result into the 4 layers so that each metric has a house. Record attain metrics comparable to activation price, frequency metrics comparable to weekly lively price, depth metrics such because the stage distribution, and final result metrics comparable to cycle time. For every metric report the components, the information supply, the proprietor, the refresh cadence, and the minimal group dimension for reporting. This desk turns into the measurement constitution that you’ll publish to staff within the remaining step. A brief, express constitution prevents the most typical failure, which is 2 groups reporting the identical metric identify with completely different math.

Step 2 – Stock each AI software and information supply

Record each generative AI software that staff can attain, together with authorised assistants, embedded options in current software program, and developer instruments. For every one, notice whether or not it has an admin console, a utilization report, an API, and a user-level export. Microsoft 365 Copilot, for instance, presents admin heart studies over mounted home windows and a Graph endpoint, as described in this Worklytics tutorial on measuring Copilot utilization. Add id supplier logs, proxy logs, and expense information as secondary sources for instruments you don’t formally handle. Mark every supply as authoritative, partial, or inferred, as a result of that label tells readers how a lot to belief any metric constructed on it.

Step 3 – Export uncooked utilization information and pseudonymize customers

Pull user-level exports from every console on a set schedule, ideally each 7 days, and retailer the uncooked recordsdata in a restricted folder. GitHub, for instance, lets directors obtain user-level information from the Copilot utilization dashboard as NDJSON for evaluation in different instruments. Then run a small normalizer that maps every vendor’s columns to at least one shared schema and replaces each id with a salted hash. The hash helps you to rely distinct folks and comply with their conduct over time with out exposing names to analysts. Professional tip: preserve the salt in a secrets and techniques supervisor, rotate it every quarter, and by no means commit it to a repository.

import hashlib
import os

import pandas as pd

SALT = os.environ["ADOPTION_SALT"]


def pseudonymize(id: str) -> str:
    digest = hashlib.sha256((SALT + id.strip().decrease()).encode("utf-8"))
    return digest.hexdigest()[:16]


uncooked = pd.read_csv("vendor_export.csv", parse_dates=["activity_date"])
occasions = pd.DataFrame({
    "user_id": uncooked["user_email"].map(pseudonymize),
    "software": "copilot",
    "event_date": uncooked["activity_date"].dt.strftime("%Y-%m-%d"),
    "session_id": uncooked["session_id"],
    "turns": uncooked["turn_count"],
})
occasions.to_csv("occasions.csv", index=False)

Step 4 – Load a clear roster and occasion desk

Create 2 core tables in your evaluation database: a roster of everybody who must be measured and an occasions desk of classes. The roster comes out of your HR system and carries the division, function, license flag, and license date for every pseudonymized individual. The roster is what makes non-adopters countable, which is the only function that separates actual adoption analytics from an inventory of lively customers. Load each tables and test that each occasion maps to a roster row. Examine any orphan occasions, as a result of they normally sign unmanaged accounts or contractors lacking from the HR feed.

Information hygiene determines whether or not your numbers will survive scrutiny from finance and audit groups. Deduplicate classes by identifier, convert timestamps to a single time zone, and drop inner take a look at accounts earlier than any metric is computed. Report row counts at every stage so that you could clarify any change from final week’s report. Many groups skip this self-discipline after which spend conferences debating whether or not a dip is actual or an artifact of a failed export. Retailer the load script in model management in order that reviewers can see precisely how the information was formed.

CREATE TABLE roster (
  user_id TEXT PRIMARY KEY,
  division TEXT NOT NULL,
  function TEXT,
  licensed INTEGER NOT NULL DEFAULT 1,
  first_license_date TEXT
);

CREATE TABLE occasions (
  user_id TEXT NOT NULL,
  software TEXT NOT NULL,
  event_date TEXT NOT NULL,
  session_id TEXT NOT NULL,
  turns INTEGER NOT NULL DEFAULT 1
);

CREATE INDEX idx_events_user ON occasions (user_id, event_date);

Step 5 – Compute attain and frequency metrics

With clear tables in place, the attain and frequency layers cut back to a couple queries. The question under computes weekly lively price by division towards licensed customers, and it suppresses any division with fewer than 5 licensed folks. Run it weekly and retailer the output so that you could chart traits with out recomputing historical past. Add a second question for activation price, utilizing the primary session date per individual, and a 3rd for stickiness, dividing weekly lively customers by month-to-month lively customers. Examine every outcome to your personal baseline from the primary month slightly than to an exterior benchmark.

WITH weekly AS (
  SELECT strftime('%Y-%W', e.event_date) AS week,
         r.division,
         COUNT(DISTINCT e.user_id) AS active_users
  FROM occasions e
  JOIN roster r ON r.user_id = e.user_id
  WHERE r.licensed = 1
  GROUP BY week, r.division
),
base AS (
  SELECT division, COUNT(*) AS licensed_users
  FROM roster
  WHERE licensed = 1
  GROUP BY division
)
SELECT w.week,
       w.division,
       w.active_users,
       b.licensed_users,
       ROUND(100.0 * w.active_users / b.licensed_users, 1) AS weekly_active_rate_pct
FROM weekly w
JOIN base b USING (division)
WHERE b.licensed_users >= 5
ORDER BY w.week, w.division;

Step 6 – Measure depth and classify cohorts by stage

Depth is the place your evaluation begins to vary from a vendor dashboard, so give it essentially the most care. The script under assigns every licensed individual to a simplified model of the 5 levels from the adoption telemetry paper: discover, try, navigate, rework, and embed. It omits the task-success situation as a result of most exports don’t carry final result flags, so deal with the outcome as a primary approximation. Combination the levels by division, apply the minimal group dimension, and by no means print particular person identifiers within the output. Look first for cohorts the place most individuals sit at navigate, as a result of that sample is the shallow plateau.

Interpret the stage desk as a prognosis and never as a rating. A division with many individuals at discover has a visibility or entry downside, one with many at try has an onboarding downside, and one caught at navigate wants workflow redesign. Re-run the script month-to-month and examine the stage distribution with the earlier month, as a result of motion between levels is the clearest proof that interventions work. Share the output with staff leads in a gaggle view solely. Resist the temptation to construct a leaderboard, since a leaderboard invitations gaming and erodes the belief you want for trustworthy information.

import sqlite3

import pandas as pd

con = sqlite3.join("adoption.db")
occasions = pd.read_sql_query(
    "SELECT user_id, event_date, session_id, turns FROM occasions",
    con, parse_dates=["event_date"])
roster = pd.read_sql_query(
    "SELECT user_id, division FROM roster WHERE licensed = 1", con)

occasions["week"] = occasions["event_date"].dt.to_period("W")
classes = occasions.drop_duplicates("session_id")
per_user = classes.groupby("user_id").agg(
    active_weeks=("week", "nunique"),
    first_week=("week", "min"),
    multi_step_share=("turns", lambda s: (s >= 12).imply()),
)
last_week = occasions["week"].max()
per_user["weeks_since_first"] = per_user["first_week"].map(
    lambda w: (last_week - w).n + 1)
per_user["active_ratio"] = per_user["active_weeks"] / per_user["weeks_since_first"]


def assign_stage(row):
    if row["active_weeks"] >= 6 and row["multi_step_share"] >= 0.6:
        return "embed" if row["active_ratio"] >= 0.9 else "rework"
    if row["active_weeks"] >= 3:
        return "navigate"
    return "try"


per_user["stage"] = per_user.apply(assign_stage, axis=1)
desk = roster.merge(per_user["stage"], left_on="user_id",
                     right_index=True, how="left")
desk["stage"] = desk["stage"].fillna("discover")

sizes = desk.groupby("division")["user_id"].rely()
abstract = (desk.groupby("division")["stage"]
           .value_counts(normalize=True).unstack(fill_value=0))
print((abstract[sizes >= 5] * 100).spherical(1))

Step 7 – Add a pulse survey and a shadow AI test

Telemetry can’t see intent or unofficial instruments, so add two light-weight devices. Ship a quarterly pulse survey of 5 to 7 behavioral questions on frequency, duties, confidence, limitations, and perceived time saved. Then measure shadow AI by classifying proxy or DNS visitors into sanctioned and unsanctioned AI domains and computing the unsanctioned share by division. Examine the survey’s reported utilization with logged utilization for every group, since massive gaps level to unseen instruments or to embedded options that individuals don’t acknowledge as AI. Pair the findings with an amnesty message that invitations folks to reveal instruments they depend on, and level them to the acceptable-use guidelines in your accountable AI coverage template.

SELECT division,
       ROUND(100.0 * SUM(CASE WHEN sanctioned = 0 THEN requests ELSE 0 END)
             / SUM(requests), 1) AS shadow_share_pct
FROM proxy_ai_traffic
GROUP BY division
HAVING COUNT(DISTINCT user_id) >= 5;

Step 8 – Publish the constitution, report by cohort, and iterate

Shut the loop by publishing the measurement constitution, scheduling the critiques, and committing to a small variety of actions per quarter. The constitution lists every metric, its definition, its information supply, its minimal group dimension, and the dedication that adoption information by no means feeds particular person efficiency critiques. Report back to executives on traits, the most important stall level, and the proof for outcomes, and report back to staff leads on their very own cohort solely. Revisit definitions each quarter as instruments and utilization patterns change. Brokers that act with out prompts are a rising purpose to take action, as our roundup of agentic AI adoption statistics for 2026 reveals. Preserve the pipeline easy sufficient {that a} new analyst can rerun it in a day.

metrics:
  - identify: weekly_active_rate
    layer: frequency
    denominator: licensed_users
    window_days: 7
    supply: occasions
    min_group_size: 5
    proprietor: analytics
  - identify: stage_distribution
    layer: depth
    denominator: licensed_users
    supply: occasions
    min_group_size: 5
    proprietor: analytics
privateness:
  individual_scoring: false
  store_verbatim_prompts: false
  retention_days: 400

Really useful by AIplusInfo

Books to go deeper on measurement and adoption

Three well-reviewed titles that map to the measurement and training work described above.

As an Amazon Affiliate, AIplusInfo earns from qualifying purchases.

Ebook

How one can Measure Something: Discovering the Worth of Intangibles in Enterprise

Hubbard’s methodology for estimating hard-to-measure outcomes helps you construct baselines and calibrated estimates for AI time financial savings and high quality positive factors.

Purchase on Amazon

Ebook

Co-Intelligence: Dwelling and Working with AI

Mollick’s sensible information to working alongside AI helps managers coach groups towards deeper and extra assured on a regular basis use of the instruments.

Purchase on Amazon

Ebook

Measure What Issues: How Google, Bono, and the Gates Basis Rock the World with OKRs

Doerr reveals methods to flip targets into measurable key outcomes, a helpful sample for setting adoption targets that groups can really personal.

Purchase on Amazon

Key Insights

• Weekly generative AI use amongst employed U.S. adults reached 39.2% within the second quarter of 2026, so inner adoption charges must be judged towards a rising exterior baseline.
• Solely 15% of U.S. staff used AI each day as of Could 2026, in line with Gallup, which reveals that each day behavior lags behind weekly attain and wishes its personal metric.
• Simply one in six occupations exceeds 70% generative AI adoption, so most roles nonetheless sit nicely under saturation and wish role-specific targets as a substitute of 1 company-wide quantity.
• ActivTrak information cited in an adoption telemetry paper present that 82% of AI customers maintain utilization whereas solely about 2% attain embedded use, so retention alone overstates change.
• A 2025 survey discovered that 93% of staff admit coming into info into AI instruments with out approval, making shadow AI a measurement hole to shut earlier than totals could be trusted.
• A examine of 5,179 assist brokers discovered that AI help raised productiveness by 14% on common and by 34% for novices, so final result positive factors focus amongst much less skilled staff.
• Analysis summarized by IntuitionLabs studies that 56% of HR professionals don’t formally measure AI funding success, which implies disciplined measurement is itself a aggressive benefit right now.
• Microsoft’s Copilot dashboard hides teams smaller than 5 folks, in line with a Worklytics information, a privateness default that your personal cohort reporting ought to copy.

The proof factors in a single path: adoption is broad however uneven, and the headline numbers cover most of what leaders must know. Weekly attain retains climbing, but each day use, depth of use, and embedded workflow change stay far smaller than license counts indicate. Measurement due to this fact has to separate the layers, report them by cohort, and pair utilization information with final result proof from comparability teams. Shadow AI and privateness considerations add two extra constraints, as a result of unseen instruments distort the numbers and heavy-handed monitoring erodes the belief that trustworthy information will depend on. Organizations that outline their denominators, publish a constitution, and diagnose stall factors will make higher selections than those who chase a single adoption share. The comparability under reveals how the principle metric households differ in effort, danger, and usefulness.

Adoption Measurement in Observe at Three Organizations

The St. Louis Fed Generative AI Adoption Tracker

In apply, the clearest public instance of ongoing adoption measurement is the Generative AI Adoption Tracker constructed by economists Alexander Bick, Adam Blandin, and David Deming. The staff ran the Actual-Time Inhabitants Survey every quarter with a nationally consultant pattern of adults aged 18 to 64. Employed respondents report how typically and the way intensively they use generative AI. Between late 2024 and mid 2026, weekly use amongst employed adults rose from 28.2% to 39.2%, and AI-assisted work hours climbed from 4.1% to six.3%. Estimated hours saved elevated from 1.6% to 2.2%, a acquire of roughly 37% relative to the beginning stage. The design separates attain, depth, and time financial savings into distinct sequence as a substitute of mixing them right into a single rating. Its essential restrict is that respondents self-report their time financial savings, so the Fed calls the measurements approximate and corporations ought to pair survey solutions with logged utilization earlier than reporting financial savings.

Cohort Stage Classification within the NANTE Telemetry Mannequin

A second instance comes from the reference implementation of the NANTE adoption telemetry mannequin, which classifies cohorts into 5 levels utilizing utilization logs alone. The writer constructed a deterministic classifier and ran it on six artificial profiles protecting 167 customers over 149 days. Within the shallow-plateau profile, 99.4% of customers reached the navigate stage but the composite rating was solely 49.9, in contrast with 60.5 for the wholesome cohort. The classifier additionally flagged champion dependency, with a Gini coefficient of 0.65, in a cohort the place a couple of customers generated a lot of the exercise. These outcomes present how stage counts expose depth issues {that a} weekly lively price utterly hides. The primary limitation is that the information are artificial and the thresholds are proposed slightly than validated, so actual cohorts might present a number of pathologies without delay. Groups ought to deal with the levels as a diagnostic place to begin and calibrate the cutoffs towards their very own historical past.

A Staggered Rollout for Buyer Help Brokers

A 3rd instance reveals how utilization information can anchor a rigorous final result examine. Researchers studied 5,179 buyer assist brokers at an organization that carried out a generative AI conversational assistant in levels, which allow them to examine brokers with and with out entry over time. Productiveness measured as points resolved per hour rose by about 14% on common, and by about 34% amongst novice and low-skilled staff. Skilled and extremely expert brokers noticed minimal affect, so the common hid very completely different outcomes by talent stage. The design labored as a result of the rollout was staggered and the result metric was mounted earlier than the evaluation started. Its restrict is scope, for the reason that findings cowl one occupation and will not switch to open-ended data work. Readers ought to nonetheless borrow the self-discipline of a comparability group, which is the most affordable protection towards crediting AI with positive factors that different adjustments produced.

Classes From Three Case Research in Enterprise AI Adoption

Case Research: Sanofi’s Concierge AI Companion

Turning to massive enterprise deployments, Sanofi confronted a basic scaling problem with its inner AI companion known as Concierge. The corporate wanted to increase the assistant throughout its world workforce whereas holding the pliability to evolve and assembly strict information sensitivity necessities. Its answer was to construct Concierge in-house on Amazon Bedrock and preserve it model-agnostic to keep away from supplier lock-in. Engineers deployed it throughout a number of areas to scale back latency and started constructing an agent hub for information retrieval and workflow automation. In keeping with AWS, Concierge reached 50,000 weekly customers and 72,000 month-to-month customers as of the primary quarter of 2026, with 11 million conversations. That weekly-to-monthly ratio of about 0.69 is a decent stickiness sign for an enterprise assistant. Amongst surveyed staff, 90.5% reported saving time on routine duties.

The bounds matter simply as a lot because the headline adoption numbers do. The time financial savings are self-reported survey outcomes, and the 90% positive-feedback score is a satisfaction measure slightly than a productiveness measure. AWS printed the case examine and the identical vendor provides the platform, which is a gentle concern about independence. The information additionally describe attain and frequency much better than they describe outcomes, since no task-level time or high quality measurements are reported. Readers ought to deal with the numbers as sturdy proof of adoption and as suggestive proof just for productiveness. Anybody borrowing the strategy ought to add a baseline and a comparability group earlier than claiming financial savings.

Case Research: Novo Nordisk’s Self-Service Generative AI Platform

Novo Nordisk needed staff to innovate with generative AI in nonregulated processes whereas staying safe, compliant, and cost-effective. The issue was that many staff had options in thoughts however couldn’t construct the functions or preserve the supporting infrastructure. The corporate additionally discovered it laborious to foretell which use circumstances would succeed. Its answer was a self-service platform on AWS, constructed round chatbots and brokers that use basis fashions in Amazon Bedrock. Staff can construct and customise chatbots and share them with their groups, whereas serverless companies preserve base prices low. Greater than 25,000 staff have used the platform to create over 2,500 chatbots, at roughly $10 monthly per use case. That ratio implies about one chatbot for each ten staff who used the platform.

These figures measure breadth of adoption unusually nicely, since distinct use circumstances per staff is precisely the breadth metric described earlier. They’re nonetheless exercise and value measures, and they don’t present how a lot every chatbot is value or what number of sit idle after a pilot. The associated fee determine covers working bills per use case, so it excludes improvement effort and staffing, a restrict value stating in any govt abstract. Monitoring the month-over-month enhance in lively use circumstances would flip this snapshot right into a development. The platform additionally lacks a reported enterprise final result, which is a good criticism of activity-based success tales. The lesson is to pair breadth counts with an exercise threshold that separates working instruments from deserted experiments.

Case Research: Pfizer’s PACT Generative AI Search

Pfizer groups have been stretched skinny, so the Pfizer-Amazon Collaboration Staff, often known as PACT, was created to deal with the bandwidth wanted for prototyping. Scientists confronted a knowledge discovery downside as a result of growing one drug can generate roughly 20,000 paperwork that they’d searched manually. PACT paired Pfizer groups with AWS prototyping specialists and options architects, and it developed a generative AI search software that scientists attain by an inner platform known as Vox. The staff additionally constructed an anomaly detection prototype for manufacturing, and it pursued 14 tasks in whole. AWS studies that scientists might save as much as 16,000 hours of search time yearly and that infrastructure prices fell by 55%. Information discovery time dropped by as much as 80% for 1,500 scientists. A PACT prototype took not more than six weeks, in contrast with at the least three months for an inner effort.

Slim, well-defined use circumstances like these make final result measurement far simpler than broad assistant rollouts. 5 of the 14 tasks had moved into manufacturing, which is a reminder that pilot success and sustained adoption are various things. The primary criticism is that an up-to determine describes a ceiling and never a mean, and that the numbers come from a vendor case examine with out unbiased auditing. The outcomes additionally say little about what number of scientists use the search software every week or how deeply they use it. A whole measurement program would add attain, frequency, and depth information beside these financial savings estimates. The takeaway is that attributable final result metrics are invaluable however ought to sit beside adoption information and never change it.

Widespread Questions About Measuring Generative AI Adoption within the Office