Make experimentation ordinary
- Write and improve documents
- Summarize meetings and research
- Create presentation first drafts
- Assist with code and analysis
- Brainstorm, reframe and ideate
Organizational AI adoption
How Do Organizations Adopt AI Successfully?
AI adoption is not a technology project. It is an organizational transformation journey—one that starts by helping people find value in the work they already do.
Platforms, models and architecture matter later. First create the conditions for learning, then turn useful discoveries into reliable ways of working.
The AI adoption journey
What are the five stages of AI adoption?
Each stage has a different job. Applying enterprise controls too early suppresses discovery; scaling before a workflow is reliable multiplies inconsistency.
01CrawlCuriosity & experimentation
02WalkFirst proven business value
03EngineerRepeatable & governed workflow
04RunOrganization-wide scaling
05FlyContinuous innovation
The sequence
Discover value→prove it→make it repeatable→scale it→keep improving it
Curiosity before ROI
Stage 1: How do you start AI adoption safely?
The goal is not measurable business value yet. The goal is organizational learning.
Give employees a safe place to explore, test ideas and notice where AI removes friction from everyday work. Keep the cost of trying small—and the permission to learn wide.
What good looks like now
- Wide experimentation is encouraged
- Useful stories are collected and shared
- Individuals remove small points of friction
- Formal ROI is not the admission ticket
- Governance does not become a blanket blocker
Better, faster or cheaper
Stage 2: How do you prove one valuable AI workflow?
Experimentation becomes adoption when one workflow clearly outperforms the old way of working.
Find one outcome valuable enough that the team would not willingly return to the previous process. It does not need to save millions; it needs to matter to the people doing the work.
Choose a bounded workflow
Release notesSocial contentMeeting summariesSupport draftsKnowledge searchSales proposals
Is the new workflow meaningfully better, faster or cheaper after review and correction?
Compare it with a real manual baseline. Use the evidence claim boundary to keep a narrow result narrow, then choose a task-relevant benchmark or local test. A persuasive demo is not the same thing as a dependable result.
The scaling breakpoint
Stage 3: How do you make AI workflows repeatable?
A good pilot is not a process. This is where isolated success becomes transferable capability.
Most initiatives stall here.The pilot works because a few enthusiasts carry hidden knowledge. Remove that dependency before expanding.
Make the workflow explicit
- InputWhat enters the process?
- OutputWhat does “good” look like?
- OwnerWho is accountable?
- ControlWhat requires human review?
- FailureWhat happens when AI is wrong?
- MeasureHow will success be tracked?
Build the operating layer
- Workflow and exception design
- Prompt and input standardization
- Human review and escalation paths
- KPIs, acceptance tests and quality checks
- Security, compliance and data review
- Clear governance proportional to risk
Perfection is not the goal. Consistency across people is. For software delivery, the Thinking with AI workflow loop makes these controls reviewable.
Scale proven value
Stage 4: How do you scale AI across an organization?
Once the workflow is proven and repeatable, expand the pattern—not merely access to a tool.
Look for adjacent workflows across departments, standardize the reusable parts and measure realized benefits. AI now moves from individual productivity improvement to enterprise capability; multi-system operational work may need the interfaces and ownership model in the sovereign operational-intelligence stack.
SalesProposal generation
Customer serviceAssisted responses
HRRecruiting workflows
FinanceAssisted analysis
DevelopmentCoding assistants
Center of Excellence
Stage 5: How does an AI Center of Excellence help?
When AI becomes a core capability, give it dedicated leadership without creating a central bottleneck.
The operating principle
The CoE is an accelerator, not a gatekeeper.
It helps teams move faster and more safely by making scarce expertise, standards and reusable patterns available to everyone.
01
StrategySet direction and prioritize opportunity
02
GovernanceDefine policy proportional to impact
03
ArchitectureGuide platforms, patterns and vendors
04
EnablementTrain people and spread practice
05
MeasurementTrack value and improve the portfolio
06
InnovationEvaluate what becomes possible next
Adoption in numbers
How widely is AI actually being used?
Population use, organizational surveys, active users, web visits and revenue measure different layers of adoption. They are shown separately.
Working-age population estimated to have used generative AI in H2 2025.
Microsoft telemetry-based estimateSurveyed organizations reporting AI use in at least one business function in 2025.
McKinsey survey via Stanford AI IndexPeople aged 16–74 reporting generative-AI use in the previous three months.
Eurostat survey, 2025AI adoption by country
Microsoft estimates working-age use; Eurostat surveys recent use among people aged 16–74. The datasets remain separate.
The selected economy is emphasized with a ring and crosshair as well as a nearby text reading.
Loading the map values…
Access is broad; depth still varies
In 2025, 88% reported AI use in at least one function and 79% reported regular generative-AI use.
Loading the HTML chart…
Source note, checked 31 July 2026: the Stanford chapter PDF and figure report 79% for generative AI; the chapter landing-page summary says 70%. This chart retains the figure value while the publisher’s pages disagree.
Using AI is much more common than measuring value
McKinsey reports 88% regular use, 39% with any enterprise-level EBIT impact and 6% meeting its high-performer threshold. These are separate signals, not a funnel.
Loading the HTML chart…
19.8% of all U.S. businesses used AI
32.0% of firms with 100–249 employees
37.0% of firms with 250+ employees
39.7% in the information sector
33.9% in finance and insurance
≈14.0% in retail trade
Census BTOS, period ending May 3, 2026; use in any business function during the previous two weeks.
Codex grew fast—then its distribution changed
OpenAI disclosed 1.6M+ Codex weekly users in February and 5M+ in June. The broader post-rollout series rose from 8M in July to 15M+ in August; it combines Codex with ChatGPT Work.
Loading the HTML chart…
First-party agent telemetry
Codex and Claude Code behavioral samples
Vendor studies of task scope, work mix and human–agent control—not independent productivity audits.
Codex
May–June 2026Individual, organizational and OpenAI-user samples
- Users with ≥1 request >30 min
- 80.6%
- Users with ≥1 request >1 h
- 70.2%
- Users with ≥1 request >8 h
- 25.6%
Human-time equivalents estimate task scope, not time saved or business value.
Claude Code
Oct 2025–Apr 2026≈400,000 interactive sessions from ≈235,000 people
- Active runtime per user
- 20 h/week
- Sessions involving code work
- 56%
- Human planning / Claude execution
- 70% / 80%
Excludes third-party IDE, SDK and headless usage; work types and decisions are classified.
ChatGPT leads tracked generative-AI web visits
This Similarweb-based series excludes embedded, app, API and enterprise use. It measures web attention, not total product reach.
Highlight
Showing all five platform lines.
Exact chart data and definitions
AI diffusion by economy
| Economy | H1 2025 | H2 2025 | Change |
|---|---|---|---|
| United Arab Emirates | 59.4% | 64.0% | +4.6 pp |
| Singapore | 58.6% | 60.9% | +2.3 pp |
| Norway | 45.3% | 46.4% | +1.1 pp |
| Ireland | 41.7% | 44.6% | +2.9 pp |
| France | 40.9% | 44.0% | +3.1 pp |
| Spain | 39.7% | 41.8% | +2.1 pp |
| New Zealand | 37.6% | 40.5% | +2.9 pp |
| Netherlands | 36.3% | 38.9% | +2.6 pp |
| United Kingdom | 36.4% | 38.9% | +2.5 pp |
| Qatar | 35.7% | 38.3% | +2.6 pp |
| Australia | 34.5% | 36.9% | +2.4 pp |
| Israel | 33.9% | 36.1% | +2.2 pp |
| Canada | 33.5% | 35.0% | +1.5 pp |
| South Korea | 25.9% | 30.7% | +4.8 pp |
| Germany | 26.5% | 28.6% | +2.1 pp |
| United States | 26.3% | 28.3% | +2.0 pp |
| South Africa | 19.3% | 21.1% | +1.8 pp |
| Japan | 16.7% | 19.1% | +2.4 pp |
| Mexico | 16.7% | 17.8% | +1.1 pp |
| Brazil | 15.6% | 17.1% | +1.5 pp |
| China | 15.4% | 16.3% | +0.9 pp |
| India | 14.2% | 15.7% | +1.5 pp |
Europe individual use
| Geography | Used generative AI |
|---|---|
| European Union | 32.66% |
| Euro area | 34.20% |
| Belgium | 42.01% |
| Bulgaria | 22.50% |
| Czechia | 35.35% |
| Denmark | 48.44% |
| Germany | 32.25% |
| Estonia | 46.64% |
| Ireland | 44.93% |
| Greece | 44.09% |
| Spain | 37.88% |
| France | 37.46% |
| Croatia | 27.52% |
| Italy | 19.86% |
| Cyprus | 44.20% |
| Latvia | 33.40% |
| Lithuania | 36.89% |
| Luxembourg | 42.54% |
| Hungary | 29.56% |
| Malta | 46.46% |
| Netherlands | 44.70% |
| Austria | 39.42% |
| Poland | 22.68% |
| Portugal | 38.70% |
| Romania | 17.76% |
| Slovenia | 37.56% |
| Slovakia | 30.79% |
| Finland | 46.27% |
| Sweden | 42.01% |
| Norway | 56.32% |
| Switzerland | 47.02% |
| Bosnia and Herzegovina | 20.26% |
| North Macedonia | 22.03% |
| Albania | 27.28% |
| Serbia | 18.64% |
| Türkiye | 17.19% |
| Kosovo* | 44.85% |
* Eurostat’s geographic label; the designation is without prejudice to positions on status.
Organizational adoption
| Year | Any AI | Generative AI |
|---|---|---|
| 2023 | 55% | Not reported |
| 2024 | 78% | 71% |
| 2025 | 88% | 79% |
Generative-AI web traffic
| Platform | May 2025 | Mar 2026 | May 2026 |
|---|---|---|---|
| ChatGPT | 76.4% | 56.7% | 52.7% |
| Gemini | 8.9% | 25.5% | 27.3% |
| Claude | 1.6% | 6.0% | 8.9% |
| Grok | Not reported | 6.0% | 2.8% |
| Perplexity | Not reported | 2.0% | 1.3% |
Use versus enterprise value
| Signal | Share | Meaning |
|---|---|---|
| Regular AI use | 88% | Respondents saying their organization regularly uses AI in at least one business function. |
| Any enterprise EBIT impact | 39% | Respondents reporting any AI-attributable EBIT impact at enterprise level. |
| AI high performer | 6% | Respondents attributing at least 5% of EBIT to AI and reporting significant value from AI use. |
U.S. business use
| Business group | Using AI |
|---|---|
| All U.S. businesses | 19.8% |
| 100–249 employees | 32.0% |
| 250+ employees | 37.0% |
| Information sector | 39.7% |
| Finance and insurance | 33.9% |
| Retail trade | About 14.0% |
Codex product-adoption milestones
| Date | Population | Users | Evidence |
|---|---|---|---|
| Feb 27, 2026 | Codex weekly users | More than 1.6M | OpenAI company disclosure |
| Jun 2, 2026 | Codex weekly users | More than 5M | OpenAI company disclosure |
| Jul 14, 2026 | Codex + ChatGPT Work active users (post-rollout) | Reached 8M | OpenAI Codex engineering lead on X |
| Jul 21, 2026 | Codex + ChatGPT Work active users (post-rollout) | 10M milestone | OpenAI Codex engineering lead on X |
| Aug 13, 2026 | Codex + ChatGPT Work active users (post-rollout) | More than 15M | OpenAI Codex engineering lead on X |
The product scope changed on July 9. Later milestones combine Codex and ChatGPT Work and omit an activity window.
Public-market lens
What do we actually know about each major player?
SpaceX completed its IPO with xAI inside the group. OpenAI and Anthropic IPO filings and timing are still reported rather than public registration statements, so the status and financial numbers below remain dated signals—not investment advice.
| Player | Market status | Reach signal | Financial signal | How to read it |
|---|---|---|---|---|
| ChatGPTOpenAI | Private; confidential IPO filing reported | 900M+ weekly active usersFebruary 2026 | $30B 2026 revenue targetprojection reported by secondary sources | OpenAI also reported 50M+ consumer subscribers and 9M+ paying business users. WAU is a vendor-reported measure, not audited MAU. |
| GeminiAlphabet | Public parent (NASDAQ: GOOGL / GOOG) | 750M+ monthly active usersQ4 2025 | Not disclosed separatelyGemini is embedded across several Alphabet products | A later 900M+ MAU figure was reported from Google I/O 2026. The ledger keeps the 750M earnings-call disclosure as the primary-source baseline. |
| ClaudeAnthropic | Private; confidential IPO filing reported | No current company-wide MAU disclosedJuly 2026 check | $47B annualized run-rate revenuecompany disclosure, May 2026 | Revenue run rate is not realized annual revenue. Claude has unusually strong enterprise exposure, so consumer traffic understates its commercial footprint. |
| GrokSpaceX / xAI | Public parent following SpaceX IPO | 117M monthly active users of Grok featuresMarch 2026 | Not disclosed separatelyxAI is reported inside the combined SpaceX group | The SEC filing covers Grok features across X, web and apps; it is broader than standalone chatbot traffic. December 2025 was 89M MAU. |
| PerplexityPerplexity AI | Private | No comparable current MAU disclosedJuly 2026 check | $450M–$500M annualized revenuesecondary estimate, Q2 2026 | Queries, visits and MAU are often mixed in third-party summaries. The revenue range is useful as an order-of-magnitude estimate, not an audited result. |
Reach, not loyalty. Weekly and monthly active users use different windows, may count embedded features, and are usually vendor reported.
Momentum, not the whole market. Website visits miss APIs, mobile apps, workplace licences, search integration and other embedded distribution.
A pace, not booked annual revenue. It annualizes a recent period and can move quickly in either direction.
A plan, not an outcome. Keep targets visible because they matter to market expectations, but label them separately from realized revenue.
Sources, update policy and secondary watchlist
Primary evidence used in the charts
- Microsoft AI Economy Institute: Global AI Adoption in 2025Telemetry-based global and economy diffusion.
- Eurostat: use of AI by individualsSurveyed three-month use; data code
isoc_ai_iaiu. - Stanford HAI: 2026 AI Index — EconomyMcKinsey organizational adoption series and limitations.
- OpenAI: Scaling AI for everyoneChatGPT WAU, consumer subscribers and paying business users.
- OpenAI: Codex for knowledge workMore than 5M Codex weekly active users in June 2026.
- OpenAI: how agents are transforming workFirst-party Codex task-horizon, cross-role and internal-use telemetry.
- OpenAI: introducing ChatGPT WorkDesktop-app migration, ChatGPT Classic rename and bundle distribution context.
- OpenAI Help: ChatGPT Work and CodexCurrent product boundaries and availability.
- U.S. Census Bureau: AI use by businessesOperational use by company size and sector.
- McKinsey: The state of AI in 2025Enterprise EBIT impact, scaling and high-performer definitions.
- OpenAI Codex lead: 8M combined usersVendor-representative signal for Codex plus ChatGPT Work, 14 July 2026.
- OpenAI Codex lead: 10M milestoneLater combined milestone, 21 July 2026; activity window unstated.
- OpenAI Codex lead: crossed 15M active usersCombined Codex plus ChatGPT Work milestone disclosed 13 August 2026; activity window unstated.
- Alphabet Q4 2025 earnings callGemini app MAU.
- Anthropic Series H announcementAnnualized run-rate revenue.
- Anthropic: agentic coding and persistent returns to expertisePrivacy-preserving study of about 400,000 Claude Code sessions from about 235,000 people.
- SpaceX final IPO prospectusGrok-feature MAU and xAI group context.
Secondary sources monitored
These sources are useful for discovery, traffic estimates, projections and cross-checking. A number moves into a chart only when its definition, period and provenance survive review.
Chart data last refreshed . The automated updater accepts only bounded primary-source changes; blocked, ambiguous and secondary figures remain queued for manual review.
The leadership question
How do we get thousands of people to adopt AI successfully without turning it into chaos?