What happens when AI agents become the front door to the internet?

Global adoption needs a defined denominator

The Q2 2026 Microsoft report estimates generative AI use among people aged 15–64 at about 19% worldwide, rounded to the nearest percentage point. Selected economy values use the same report vintage.

Modeled telemetry with adjustments and residual coverage bias; no claim of precise national ranking or agent adoption.

Similar adoption can hide different breadth of use

EIBIS reports generative AI adoption of 37% in EU firms and 36% in US firms. Among AI-using firms, 55% and 81% respectively use it in at least two processes.

2025 survey; value-added weights. Different denominators in the two comparisons; no significance claim for the one-point adoption gap.

More active agents does not establish better outcomes

Microsoft reports 15 times as many active agents in the measured Microsoft 365 ecosystem between March 2025 and March 2026.

Active means at least one day of user-initiated usage or one autonomous run in a rolling 28-day period. The reported ratio does not count successful tasks or global users.

AI summaries and fewer outbound clicks were associated

Pew observed traditional-result clicks on 8% of Google search visits with an AI summary and 15% without one.

US browsing study, March 2025; search results reconstructed in April. Observational association, not a causal global traffic estimate.

AI's productivity effect depends on the work and measurement

A published customer-support study found an average 15% increase in resolved issues per hour. METR's later developer update highlights difficulty estimating current effects.

Different settings and designs, not opposing universal verdicts. The support study concerns human workers, not autonomous agents.

An agent combines a model with tools and feedback

An agent can attempt steps, observe tool results and revise its approach. A system still needs a goal, access boundaries and checks on its output.

The diagram is a conceptual synthesis, not an Atlas product test or a universal implementation.

Several competing product paths are emerging

Announcements and documentation span persistent personal assistance, enterprise workflows, coding agents and efficiency-focused models.

Provider descriptions identify directions; they do not establish a winner, complete availability or comparative reliability.

Mollick, LeCun and Li emphasize different capability questions

Mollick has revised his view of agent organization; LeCun argues for world models and planning; Li distinguishes convincing renderings from accurate simulations.

Attributed assessments, not consensus or a head-to-head experiment. Their roles and interests are disclosed in the source records.

Safety and responsibility need separate questions

Bengio calls for independent safety evidence and oversight; Schneier emphasizes responsibility for how agents are configured and deployed.

Different emphases, not a staged direct debate. No catastrophe probability or incident attribution is established by this article.

Who benefits is partly a design and institutional choice

Acemoglu, Autor and Johnson prioritize expertise and new tasks; Birhane asks for evidenced benefits. Willison explains the working feedback loop. Kamath sees customer-built interfaces changing brokerage distribution, while Nilekani emphasizes local-language access and useful applications.

Conceptual and normative positions, not measured universal effects. Institutional and financial interests are included in source records.

The three paths are conditional Atlas scenarios

User-controlled delegation, platform gatekeeping and uneven automation describe possible combinations of capability, access and control at three, five and ten years.

No probabilities, adoption forecasts or AGI deadlines are estimated. The sources motivate assumptions, not the dates or eventual outcomes.

From answering to acting

An assistant can explain a market. An agent preparing a weekly market brief may search, compare, calculate and revise across tools. The change is delegated action between your instruction and the final response.

A useful personal agent also needs relevant memory, service access, a budget, checks and permission boundaries. Memory must allow correction and deletion: a preference stored six months ago may no longer be yours.

A global transition, with uneven access

There is no single percentage for “how much the world uses AI.” Surveys and telemetry measure different populations. One consistent methodology can show variation without removing its coverage gaps.

Language, connectivity, mobile access, available services and payment systems shape which tasks are practical. High model capability in one market does not establish broad access elsewhere. General AI use also does not measure autonomous-agent adoption.

EU and US: adoption is only the first step

A common survey makes the EU–US comparison more useful. The European Investment Bank’s 2025 results show similar generative AI adoption, but a larger difference in breadth. The panels have different denominators: adopting AI and using it across processes are separate questions.

Count useful outcomes, not just agents

Useful effects are possible. A published study of 5,172 customer-support workers found an average 15% increase in issues resolved per hour after AI assistance was introduced, with different effects across workers. Here, “agents” means people in support roles. Conversely, METR's developer research update explains why selection and time-measurement problems limit conclusions about current coding tools.

Judge a workflow by acceptable results and total cost, including review, corrections, failed runs and integration—not the number of drafts it generates.

The internet becomes a place agents can act

An answer without a visit to its source changes discovery. Pew’s study of US Google browsing found fewer traditional-result clicks on visits with an AI summary.

Services may increasingly respond to agent requests. Cloudflare’s open-agent internet proposal and the Universal Commerce Protocol illustrate attempts to make them discoverable and callable; a specification does not establish adoption or dependable transactions.

Publishers still need credit and payment for original work. Brands need their offers understood accurately. The question is whose interests guide selection: user preferences, advertising, commissions and platform priorities can conflict.

Providers describe competing directions: persistent personal assistance at OpenAI and Google; work and usage patterns at Microsoft and Anthropic; Chinese model and tooling ecosystems around DeepSeek and Qwen; another work-and-coding route at Mistral. These are attributed strategies, not an independent ranking. Access, licenses and capabilities vary.

Three questions shape the debate

These are attributed perspectives, not a direct debate or a representative survey. They differ on what counts as capability, control and progress. Names link to the original sources.

Three possible paths: 2029, 2031 and 2036

These Atlas scenarios start in October 2026. Different regions and sectors could experience different combinations. They carry no probabilities or assumed AGI date. Watch completed tasks, intervention rates, total cost, provider switching and control of the customer relationship.

Global adoption needs a defined denominator

The Q2 2026 Microsoft report estimates generative AI use among people aged 15–64 at about 19% worldwide, rounded to the nearest percentage point. Selected economy values use the same report vintage.

Modeled telemetry with adjustments and residual coverage bias; no claim of precise national ranking or agent adoption.

Similar adoption can hide different breadth of use

EIBIS reports generative AI adoption of 37% in EU firms and 36% in US firms. Among AI-using firms, 55% and 81% respectively use it in at least two processes.

2025 survey; value-added weights. Different denominators in the two comparisons; no significance claim for the one-point adoption gap.

More active agents does not establish better outcomes

Microsoft reports 15 times as many active agents in the measured Microsoft 365 ecosystem between March 2025 and March 2026.

Active means at least one day of user-initiated usage or one autonomous run in a rolling 28-day period. The reported ratio does not count successful tasks or global users.

AI summaries and fewer outbound clicks were associated

Pew observed traditional-result clicks on 8% of Google search visits with an AI summary and 15% without one.

US browsing study, March 2025; search results reconstructed in April. Observational association, not a causal global traffic estimate.

AI's productivity effect depends on the work and measurement

A published customer-support study found an average 15% increase in resolved issues per hour. METR's later developer update highlights difficulty estimating current effects.

Different settings and designs, not opposing universal verdicts. The support study concerns human workers, not autonomous agents.

An agent combines a model with tools and feedback

An agent can attempt steps, observe tool results and revise its approach. A system still needs a goal, access boundaries and checks on its output.

The diagram is a conceptual synthesis, not an Atlas product test or a universal implementation.

Several competing product paths are emerging

Announcements and documentation span persistent personal assistance, enterprise workflows, coding agents and efficiency-focused models.

Provider descriptions identify directions; they do not establish a winner, complete availability or comparative reliability.

Mollick, LeCun and Li emphasize different capability questions

Mollick has revised his view of agent organization; LeCun argues for world models and planning; Li distinguishes convincing renderings from accurate simulations.

Attributed assessments, not consensus or a head-to-head experiment. Their roles and interests are disclosed in the source records.

Safety and responsibility need separate questions

Bengio calls for independent safety evidence and oversight; Schneier emphasizes responsibility for how agents are configured and deployed.

Different emphases, not a staged direct debate. No catastrophe probability or incident attribution is established by this article.

Who benefits is partly a design and institutional choice

Acemoglu, Autor and Johnson emphasize human expertise and new tasks. Birhane calls for evidence of promised benefits and attention to local communities.

Conceptual and normative positions, not measured universal effects. Institutional and financial interests are included in source records.

The three paths are conditional Atlas scenarios

User-controlled delegation, platform gatekeeping and uneven automation describe possible combinations of capability, access and control at three, five and ten years.

No probabilities, adoption forecasts or AGI deadlines are estimated. The sources motivate assumptions, not the dates or eventual outcomes.

Prepare reliable information and measure completed work

Atlas recommends accurate product information, original useful content, direct customer relationships and bounded workflow trials before wider delegation.

Editorial recommendations, not tested Playbooks or a promise of search visibility, ROI or transaction safety.

From answering to acting

Ask an assistant to explain a market and it returns information. Ask an agent to prepare a weekly market brief and it may search, read, compare, calculate and revise a draft across several tools. The important change is delegated action: the system can do work between your instructions and its final response.

A useful personal agent would need more than a capable model. It would need relevant memory, access to the right services, a budget, a way to check results and a clear account of what it may do without asking. Memory would need correction and deletion as well as recall. A preference stored six months ago is not automatically your preference today.

A global transition, with uneven access

There is no single trustworthy percentage for “how much the world uses AI.” Surveys ask different questions; telemetry sees different users; companies and individuals are different populations. One consistent methodology can reveal variation without resolving all those differences.

An agent future built around an English-speaking office worker is only one future. Language, mobile access, service availability, connectivity and local payment systems can change which tasks are practical. High model capability in one market does not establish broad access in another. Nor does a country's adoption share tell us how many people use agents autonomously.

EU and US: adoption is only the first step

A common survey is more informative than comparing unrelated national statistics. The European Investment Bank's 2025 results put reported generative AI adoption at similar levels in the EU and US, while revealing a larger difference in how broadly AI is used.

The panels have different denominators. They suggest why “we have adopted AI” is an incomplete management metric: it says little about what work has changed or whether the result improved.

Count useful outcomes, not just agents

Useful effects are possible. A published study of 5,172 customer-support workers found an average 15% increase in issues resolved per hour after AI assistance was introduced, with different effects across workers. Here, “agents” means people in support roles. Conversely, METR's developer research update explains why selection and time-measurement problems limit conclusions about current coding tools.

The practical question is whether a workflow produces more acceptable work at a sustainable total cost. Include review, corrections, failed runs and integration work. A system that generates ten drafts but requires extensive repair has not necessarily outperformed one careful human draft.

The internet becomes a place agents can act

If an agent answers a question without sending the person to the source, the economics of discovery change. Pew's study of US Google browsing found fewer traditional-result clicks on visits with an AI summary.

Sites can still supply the information, services and relationships behind an agent's answer. What may change is the route to them: from a person choosing links to a system selecting sources and requesting actions. Cloudflare's open-agent internet proposal and the Universal Commerce Protocol illustrate attempts to make services discoverable and callable. A specification's existence does not establish adoption or dependable transactions.

For publishers, the question becomes how original work is credited and paid for when fewer readers visit. For brands, it is whether the agent accurately understands the offer and whose interests guide the recommendation. User preferences, advertising, commissions and platform priorities can point in different directions.

Distribution is already a strategic battleground in providers' own materials. OpenAI and Google describe more persistent personal assistance; Microsoft and Anthropic emphasize work and usage patterns. DeepSeek and Qwen bring Chinese model and tooling ecosystems into the picture, while Mistral offers another work-and-coding route. These are strategic directions, not an independently established ranking. Product access, licenses and capabilities vary.

What prominent voices disagree about

Selected interpretations, not a representative survey or settled verdict on today’s systems. Names link to sources.

Three possible paths: 2029, 2031 and 2036

The following are Atlas scenarios, measured from October 2026. They are ways to test decisions, not predictions with assigned probabilities. Different countries, sectors and customers could experience different combinations.

None of these paths requires a particular AGI arrival date. Monitor completed tasks, intervention rates, total cost, repeated use, provider switching and who controls the customer relationship. A polished demonstration or benchmark improvement alone does not settle those questions.

What brands and companies should do

Atlas recommendations, not a tested Playbook. Start with dependable information; expand delegation when repeated results justify it.

Global adoption needs a defined denominator

The Q2 2026 Microsoft report estimates generative AI use among people aged 15–64 at about 19% worldwide, rounded to the nearest percentage point. Selected economy values use the same report vintage.

Modeled telemetry with adjustments and residual coverage bias; no claim of precise national ranking or agent adoption.

Similar adoption can hide different breadth of use

EIBIS reports generative AI adoption of 37% in EU firms and 36% in US firms. Among AI-using firms, 55% and 81% respectively use it in at least two processes.

2025 survey; value-added weights. Different denominators in the two comparisons; no significance claim for the one-point adoption gap.

More active agents does not establish better outcomes

Microsoft reports 15 times as many active agents in the measured Microsoft 365 ecosystem between March 2025 and March 2026.

Active means at least one day of user-initiated usage or one autonomous run in a rolling 28-day period. The reported ratio does not count successful tasks or global users.

AI summaries and fewer outbound clicks were associated

Pew observed traditional-result clicks on 8% of Google search visits with an AI summary and 15% without one.

US browsing study, March 2025; search results reconstructed in April. Observational association, not a causal global traffic estimate.

AI's productivity effect depends on the work and measurement

A published customer-support study found an average 15% increase in resolved issues per hour. METR's later developer update highlights difficulty estimating current effects.

Different settings and designs, not opposing universal verdicts. The support study concerns human workers, not autonomous agents.

An agent combines a model with tools and feedback

An agent can attempt steps, observe tool results and revise its approach. A system still needs a goal, access boundaries and checks on its output.

The diagram is a conceptual synthesis, not an Atlas product test or a universal implementation.

Several competing product paths are emerging

Announcements and documentation span persistent personal assistance, enterprise workflows, coding agents and efficiency-focused models.

Provider descriptions identify directions; they do not establish a winner, complete availability or comparative reliability.

Mollick, LeCun and Li emphasize different capability questions

Mollick has revised his view of agent organization; LeCun argues for world models and planning; Li distinguishes convincing renderings from accurate simulations.

Attributed assessments, not consensus or a head-to-head experiment. Their roles and interests are disclosed in the source records.

Safety and responsibility need separate questions

Bengio calls for independent safety evidence and oversight; Schneier emphasizes responsibility for how agents are configured and deployed.

Different emphases, not a staged direct debate. No catastrophe probability or incident attribution is established by this article.

Who benefits is partly a design and institutional choice

Acemoglu, Autor and Johnson emphasize human expertise and new tasks. Birhane calls for evidence of promised benefits and attention to local communities.

Conceptual and normative positions, not measured universal effects. Institutional and financial interests are included in source records.

The three paths are conditional Atlas scenarios

User-controlled delegation, platform gatekeeping and uneven automation describe possible combinations of capability, access and control at three, five and ten years.

No probabilities, adoption forecasts or AGI deadlines are estimated. The sources motivate assumptions, not the dates or eventual outcomes.

Prepare reliable information and measure completed work

Atlas recommends accurate product information, original useful content, direct customer relationships and bounded workflow trials before wider delegation.

Editorial recommendations, not tested Playbooks or a promise of search visibility, ROI or transaction safety.

From answering to acting

Ask an assistant to explain a market and it returns information. Ask an agent to prepare a weekly market brief and it may search, read, compare, calculate and revise a draft across several tools. The important change is delegated action: the system can do work between your instructions and its final response.

A useful personal agent would need more than a capable model. It would need relevant memory, access to the right services, a budget, a way to check results and a clear account of what it may do without asking. Memory would need correction and deletion as well as recall. A preference stored six months ago is not automatically your preference today.

Imagine a meeting brief assembled from approved documents. Simon Willison's account of agentic engineering explains why tools and feedback matter: a system that can check a result has a different working loop from one that only generates an answer.

A global transition, with uneven access

There is no single trustworthy percentage for “how much the world uses AI.” Surveys ask different questions; telemetry sees different users; companies and individuals are different populations. One consistent methodology can reveal variation without resolving all those differences.

An agent future built around an English-speaking office worker is only one future. Language, mobile access, service availability, connectivity and local payment systems can change which tasks are practical. High model capability in one market does not establish broad access in another. Nor does a country's adoption share tell us how many people use agents autonomously.

EU and US: adoption is only the first step

A common survey is more informative than comparing unrelated national statistics. The European Investment Bank's 2025 results put reported generative AI adoption at similar levels in the EU and US, while revealing a larger difference in how broadly AI is used.

The panels have different denominators. They suggest why “we have adopted AI” is an incomplete management metric: it says little about what work has changed or whether the result improved.

Count useful outcomes, not just agents

Useful effects are possible. A published study of 5,172 customer-support workers found an average 15% increase in issues resolved per hour after AI assistance was introduced, with different effects across workers. Here, “agents” means people in support roles. Conversely, METR's developer research update explains why selection and time-measurement problems limit conclusions about current coding tools.

The practical question is whether a workflow produces more acceptable work at a sustainable total cost. Include review, corrections, failed runs and integration work. A system that generates ten drafts but requires extensive repair has not necessarily outperformed one careful human draft.

The internet becomes a place agents can act

If an agent answers a question without sending the person to the source, the economics of discovery change. Pew's study of US Google browsing found fewer traditional-result clicks on visits with an AI summary.

Sites can still supply the information, services and relationships behind an agent's answer. What may change is the route to them: from a person choosing links to a system selecting sources and requesting actions. Cloudflare's open-agent internet proposal and the Universal Commerce Protocol illustrate attempts to make services discoverable and callable. A specification's existence does not establish adoption or dependable transactions.

For publishers, the question becomes how original work is credited and paid for when fewer readers visit. For brands, it is whether the agent accurately understands the offer and whose interests guide the recommendation. User preferences, advertising, commissions and platform priorities can point in different directions.

Distribution is already a strategic battleground in providers' own materials. OpenAI and Google describe more persistent personal assistance; Microsoft and Anthropic emphasize work and usage patterns. DeepSeek and Qwen bring Chinese model and tooling ecosystems into the picture, while Mistral offers another work-and-coding route. These are strategic directions, not an independently established ranking. Product access, licenses and capabilities vary.

What prominent voices disagree about

Ethan Mollick's October 2026 essay revises his earlier view of agent organization: he now thinks models can do more of the coordination that he previously expected humans to design. He still leaves open how well this works for long, ordinary organizational tasks. His argument supports experimentation, not a claim that human work has already become dispensable.

Yann LeCun puts more weight on the architecture needed for planning and understanding consequences, arguing for world models and abstraction. Fei-Fei Li stresses a related distinction: a convincing rendering is not necessarily an accurate simulation. These are research judgments, not a settled verdict on what today's systems can achieve.

Yoshua Bengio argues for independent safety evidence, accountability and international coordination as systems gain more power to act. Bruce Schneier draws attention to the humans and organizations configuring and deploying those systems. Responsibility should not disappear behind a story about a machine “going rogue.”

Daron Acemoglu, David Autor and Simon Johnson ask what kind of AI we choose to build: systems that expand human expertise and create new tasks have different implications from systems designed mainly to replace existing work. Abeba Birhane challenges a complementary assumption—that wider adoption automatically means social progress. She calls for evidence of benefits and attention to what communities actually need.

These voices are not disinterested or interchangeable. Mollick is an academic and author; LeCun, Li and Bengio lead or help lead AI organizations; Schneier works with Inrupt. Willison discloses sponsors, applied-AI work and past vendor compensation. The pro-worker paper includes funding disclosures. Birhane's accountability research has a critical policy perspective. Their arguments should be assessed on their merits, with those positions visible. This selection is not a representative survey of world opinion.

Three possible paths: 2029, 2031 and 2036

The following are Atlas scenarios, measured from October 2026. They are ways to test decisions, not predictions with assigned probabilities. Different countries, sectors and customers could experience different combinations.

None of these paths requires a particular AGI arrival date. Monitor completed tasks, intervention rates, total cost, repeated use, provider switching and who controls the customer relationship. A polished demonstration or benchmark improvement alone does not settle those questions.

What brands and companies should do

Now: make the offer dependable. Keep prices, availability, terms, product descriptions and supporting evidence accurate across the channels customers use. Publish material grounded in your own expertise and experience. Google's guidance supports useful original content and clear information; it does not guarantee inclusion in an AI answer.

Now: establish a baseline. Distinguish search traffic, identifiable AI referrals, direct visits and unknown attribution. Track qualified customers and completed purchases as well as visibility. Test one recurring workflow against the current process, measuring acceptable results, errors, review time and total cost. Give it an owner who can stop it and correct mistakes.

Over the next 6–24 months: prepare controlled service access where customers or partners need it. Let an agent check stock or order status before allowing it to change an order. Test local languages and customer journeys. Keep direct human contact available, and plan how junior staff continue to build the judgment needed to supervise the work.

Later: expand delegation when repeated results justify it. Agent-mediated purchasing needs verified identity, bounded authority, transaction records and a route to dispute or reverse mistakes. Consider protocols where actual partners support them. If one platform becomes a major source of customers, measure its effect on margins and preserve alternative routes.

Readable process files can help. A Markdown document can describe the goal, inputs, allowed actions, approval points, checks, stop conditions and responsible owner. But .md is an optional format for instructions, not an enforcement mechanism. The connected systems must apply the permissions. These are editorial recommendations, not an Atlas-tested Playbook.

Global adoption needs a defined denominator

The Q2 2026 Microsoft report estimates generative AI use among people aged 15–64 at 18.8% worldwide. Selected economy values use the same report vintage.

Modeled telemetry with adjustments and residual coverage bias; no claim of precise national ranking or agent adoption.

Similar adoption can hide different breadth of use

EIBIS reports generative AI adoption of 37% in EU firms and 36% in US firms. Among AI-using firms, 55% and 81% respectively use it in at least two processes.

2025 survey; value-added weights. Different denominators in the two comparisons; no significance claim for the one-point adoption gap.

More active agents does not establish better outcomes

Microsoft reports 15 times as many active agents in the measured Microsoft 365 ecosystem between March 2025 and March 2026.

Index 100 to 1,500 is an Atlas calculation from the reported ratio. Activity is not a count of successful tasks or global users.

AI summaries and fewer outbound clicks were associated

Pew observed traditional-result clicks on 8% of Google search visits with an AI summary and 15% without one.

US browsing study, March 2025; search results reconstructed in April. Observational association, not a causal global traffic estimate.

AI's productivity effect depends on the work and measurement

A published customer-support study found an average 15% increase in resolved issues per hour. METR's later developer update highlights difficulty estimating current effects.

Different settings and designs, not opposing universal verdicts. The support study concerns human workers, not autonomous agents.

An agent combines a model with tools and feedback

An agent can attempt steps, observe tool results and revise its approach. A system still needs a goal, access boundaries and checks on its output.

The diagram is a conceptual synthesis, not an Atlas product test or a universal implementation.

Several competing product paths are emerging

Announcements and documentation span persistent personal assistance, enterprise workflows, coding agents and efficiency-focused models.

Provider descriptions identify directions; they do not establish a winner, complete availability or comparative reliability.

Mollick, LeCun and Li emphasize different capability questions

Mollick has revised his view of agent organization; LeCun argues for world models and planning; Li distinguishes convincing renderings from accurate simulations.

Attributed assessments, not consensus or a head-to-head experiment. Their roles and interests are disclosed in the source records.

Safety and responsibility need separate questions

Bengio calls for independent safety evidence and oversight; Schneier emphasizes responsibility for how agents are configured and deployed.

Different emphases, not a staged direct debate. No catastrophe probability or incident attribution is established by this article.

Who benefits is partly a design and institutional choice

Acemoglu, Autor and Johnson emphasize human expertise and new tasks. Birhane calls for evidence of promised benefits and attention to local communities.

Conceptual and normative positions, not measured universal effects. Institutional and financial interests are included in source records.

The three paths are conditional Atlas scenarios

User-controlled delegation, platform gatekeeping and uneven automation describe possible combinations of capability, access and control at three, five and ten years.

No probabilities, adoption forecasts or AGI deadlines are estimated. The sources motivate assumptions, not the dates or eventual outcomes.

Prepare reliable information and measure completed work

Atlas recommends accurate product information, original useful content, direct customer relationships and bounded workflow trials before wider delegation.

Editorial recommendations, not tested Playbooks or a promise of search visibility, ROI or transaction safety.

From answering to acting

Ask an assistant to explain a market and it returns information. Ask an agent to prepare a weekly market brief and it may search, read, compare, calculate and revise a draft across several tools. The important change is delegated action: the system can do work between your instructions and its final response.

A useful personal agent would need more than a capable model. It would need relevant memory, access to the right services, a budget, a way to check results and a clear account of what it may do without asking. Memory would need correction and deletion as well as recall. A preference stored six months ago is not automatically your preference today.

Imagine a meeting brief assembled from approved documents. The agent can draft the analysis and identify uncertainty; sending it to a client is a separate permission. This is an illustrative workflow, not an Atlas test. Simon Willison's account of agentic engineering explains why tools and feedback matter: a system that can check a result has a different working loop from one that only generates an answer.

A global transition, with uneven access

There is no single trustworthy percentage for “how much the world uses AI.” Surveys ask different questions; telemetry sees different users; companies and individuals are different populations. The chart below uses one consistent Microsoft methodology for selected economies. Its estimated world share rose from 16.3% in H2 2025 to 18.8% in Q2 2026.

These are adjusted estimates for people aged 15–64, not a census. The method remains sensitive to desktop access and the people Microsoft can observe. Its 147-economy coverage is useful, but the decimal places should not imply perfect precision. The selected countries illustrate geographic variation; they do not stand for entire continents. Microsoft notes that several widely used domestic Chinese tools are missing from its coverage; including them would be expected to raise the China estimate.

An agent future built around an English-speaking office worker is only one future. Language, mobile access, service availability, connectivity and local payment systems can change which tasks are practical. High model capability in one market does not establish broad access in another. Nor does a country's adoption share tell us how many people use agents autonomously.

EU and US: adoption is only the first step

A common survey is more informative than comparing unrelated national statistics. The European Investment Bank's 2025 survey reports generative AI use at 37% for EU firms and 36% for US firms. Those results use value-added weights; they are not a simple count of all businesses.

The larger distinction is breadth. Among AI-using firms, 55% in the EU and 81% in the US report use in at least two processes. The panels have different denominators. They suggest why “we have adopted AI” is an incomplete management metric: it says little about what work has changed or whether the result improved.

Count useful outcomes, not just agents

Microsoft reports a fifteenfold rise in active agents within its measured Microsoft 365 ecosystem from March 2025 to March 2026. An agent can qualify as active through limited user-initiated activity or a single autonomous run in a 28-day window. That is evidence of activity inside one ecosystem, not fifteenfold growth in global productivity.

Useful effects are possible. A published study of 5,172 customer-support workers found an average 15% increase in issues resolved per hour after AI assistance was introduced, with different effects across workers. Here, “agents” means people in support roles. Conversely, METR's developer research update explains why selection and time-measurement problems limit conclusions about current coding tools.

The practical question is whether a workflow produces more acceptable work at a sustainable total cost. Include review, corrections, failed runs and integration work. A system that generates ten drafts but requires extensive repair has not necessarily outperformed one careful human draft.

The internet becomes a place agents can act

If an agent answers a question without sending the person to the source, the economics of discovery change. In Pew's study of US Google browsing, traditional-result links received clicks on 8% of visits with an AI summary and 15% without. This is an association from March 2025, not proof that AI caused a particular traffic loss everywhere.

Sites can still supply the information, services and relationships behind an agent's answer. What may change is the route to them: from a person choosing links to a system selecting sources and requesting actions. Cloudflare's open-agent internet proposal and the Universal Commerce Protocol illustrate attempts to make services discoverable and callable. A specification's existence does not establish adoption or dependable transactions.

For publishers, the question becomes how original work is credited and paid for when fewer readers visit. For brands, it is whether the agent accurately understands the offer and whose interests guide the recommendation. User preferences, advertising, commissions and platform priorities can point in different directions.

Distribution is already a strategic battleground in providers' own materials. OpenAI and Google describe more persistent personal assistance; Microsoft and Anthropic emphasize work and usage patterns. DeepSeek and Qwen bring Chinese model and tooling ecosystems into the picture, while Mistral offers another work-and-coding route. These are strategic directions, not an independently established ranking. Product access, licenses and capabilities vary.

What prominent voices disagree about

Ethan Mollick's October 2026 essay revises his earlier view of agent organization: he now thinks models can do more of the coordination that he previously expected humans to design. He still leaves open how well this works for long, ordinary organizational tasks. His argument supports experimentation, not a claim that human work has already become dispensable.

Yann LeCun puts more weight on the architecture needed for planning and understanding consequences, arguing for world models and abstraction. Fei-Fei Li stresses a related distinction: a convincing rendering is not necessarily an accurate simulation. These are research judgments, not a settled verdict on what today's systems can achieve.

Yoshua Bengio argues for independent safety evidence, accountability and international coordination as systems gain more power to act. Bruce Schneier draws attention to the humans and organizations configuring and deploying those systems. Responsibility should not disappear behind a story about a machine “going rogue.”

Daron Acemoglu, David Autor and Simon Johnson ask what kind of AI we choose to build: systems that expand human expertise and create new tasks have different implications from systems designed mainly to replace existing work. Abeba Birhane challenges a complementary assumption—that wider adoption automatically means social progress. She calls for evidence of benefits and attention to what communities actually need.

These voices are not disinterested or interchangeable. Mollick is an academic and author; LeCun, Li and Bengio lead or help lead AI organizations; Schneier works with Inrupt. Willison discloses sponsors, applied-AI work and past vendor compensation. The pro-worker paper includes funding disclosures. Birhane's accountability research has a critical policy perspective. Their arguments should be assessed on their merits, with those positions visible. This selection is not a representative survey of world opinion.

Three possible paths: 2029, 2031 and 2036

The following are Atlas scenarios, measured from October 2026. They are ways to test decisions, not predictions with assigned probabilities. Different countries, sectors and customers could experience different combinations.

In user-controlled delegation, useful cross-service tasks spread by 2029. By 2031, agents coordinate recurring work with portable context. By 2036, a personal agent could become a common front door to digital services. This path depends on dependable results, usable permissions and the ability to change providers. Persistent lock-in or heavy rework would weaken it.

In platform gatekeeping, integrated ecosystems deliver the easiest early experience by 2029. By 2031, more discovery and transactions pass through their agents. By 2036, a small set of intermediaries could exert substantial influence over access to customers and creators' income. Easy switching and durable direct relationships would weaken this path.

In uneven automation, short and bounded tasks improve by 2029 while longer chains still need intervention. By 2031, controlled environments move further than open-ended work across the web. By 2036, agents coexist with apps and human services. Repeated, economical success on long tasks across organizations and languages would weaken this scenario.

None of these paths requires a particular AGI arrival date. Monitor completed tasks, intervention rates, total cost, repeated use, provider switching and who controls the customer relationship. A polished demonstration or benchmark improvement alone does not settle those questions.

What brands and companies should do

Now: make the offer dependable. Keep prices, availability, terms, product descriptions and supporting evidence accurate across the channels customers use. Publish material grounded in your own expertise and experience. Google's guidance supports useful original content and clear information; it does not guarantee inclusion in an AI answer.

Now: establish a baseline. Distinguish search traffic, identifiable AI referrals, direct visits and unknown attribution. Track qualified customers and completed purchases as well as visibility. Test one recurring workflow against the current process, measuring acceptable results, errors, review time and total cost. Give it an owner who can stop it and correct mistakes.

Over the next 6–24 months: prepare controlled service access where customers or partners need it. Let an agent check stock or order status before allowing it to change an order. Test local languages and customer journeys. Keep direct human contact available, and plan how junior staff continue to build the judgment needed to supervise the work.

Later: expand delegation when repeated results justify it. Agent-mediated purchasing needs verified identity, bounded authority, transaction records and a route to dispute or reverse mistakes. Consider protocols where actual partners support them. If one platform becomes a major source of customers, measure its effect on margins and preserve alternative routes.

Readable process files can help. A Markdown document can describe the goal, inputs, allowed actions, approval points, checks, stop conditions and responsible owner. But .md is an optional format for instructions, not an enforcement mechanism. The connected systems must apply the permissions. These are editorial recommendations, not an Atlas-tested Playbook.