Applications

What it is about

A model on its own is an engine without a car. It takes text in and gives text out, and it knows nothing about your customers, your files or your rules. An application builds everything else around it: an interface, instructions, the right context, access to your own data, tools that let it act, checks that catch mistakes and a way to bill for the work. The EU calls the whole package an AI system.

This layer decides whether AI creates value. The same model can sit inside a product that saves a support team hours every day or inside one that nobody opens twice. It is where most founders build, where prices are set and where the law attaches duties: whoever offers an application, or uses it at work, is responsible for it.

What an AI application is made of

Every serious AI product follows the same basic pattern. A request arrives through an interface, the application assembles what the model should see, the model answers or calls a tool, the answer is checked and the cost is recorded. Anthropic's engineers call the core skill context engineering: choosing, for each model call, the smallest set of information that carries the most signal (vendor guidance).

More context can make answers worse. As the context window fills up, models recall what is in it less reliably, an effect called context rot. Good applications therefore load data just in time through tools and summarise long histories. Most builders also combine several models: in ICONIQ's 2026 survey of about 300 software companies building AI products, the average company used 3.3 models (investor survey).

Security belongs to this layer too. The OWASP Top 10 for LLM applications, maintained by an open security community, ranks prompt injection first: hidden instructions in a document or web page that take over the model. Further risks on the list include leaking sensitive data, giving the model more autonomy than the job needs and runaway costs.

1. 01Interface. where people type, speak or click: a chat window, a button inside existing software, or an API for other programs.
2. 02Instructions. the system prompt sets role, tone and rules before the user writes a word.
3. 03Context. the conversation so far, a few worked examples and the descriptions of the tools the model may use.
4. 04Own data. retrieval looks up relevant passages in your documents and hands them to the model, so it answers from your sources. Connectors, often built on MCP, reach other systems.
5. 05Tools. through tool calls the model can search, calculate, write to a database or send an email.
6. 06Guardrails. guardrails and permissions filter inputs and outputs and limit what the app may do. OWASP's warning against excessive agency translates into one rule: grant the fewest rights that get the job done.
7. 07Tests. evals are the application's exam: fixed tasks with automatic grading, run before every release.
8. 08Monitoring. observability: logging real conversations, tool calls, errors and costs in production so problems show up early. In ICONIQ's survey 63% monitor their models, but only 21% test proactively with simulated attacks (investor survey).
9. 09Billing. metering every call. ICONIQ describes a workflow budgeted at $0.10 per run that climbed past $1.50 through retries.

This nine-part order is an Aulory synthesis of guides by Anthropic and OpenAI and the OWASP list. No single source uses exactly these steps.

Which kinds of applications exist

No standard taxonomy exists. The categories below come from investors, mainly Menlo Ventures, Bessemer and Andreessen Horowitz (a16z), and they overlap: one product can be vertical and AI-native at the same time. Read them as lenses for who buys and how broad the use is.

Menlo estimates that US companies spent $19 billion on AI applications in 2025: $8.4 billion on horizontal tools, $7.3 billion on departmental ones, of which coding alone took $4.0 billion (up from $550 million in 2024), and $3.5 billion on vertical ones. Menlo leaves out AI features inside existing software, so the full amount is higher (investor estimate; Menlo holds stakes in AI companies).

The lines are blurring. a16z widened its 2026 ranking of consumer apps to products such as Canva, Notion and Grammarly, and its authors write that the gap between AI-first companies and everyone else no longer holds. The model labs also run the largest horizontal applications themselves, so one company can sit in the model layer and the application layer at once.

• Horizontal: general assistants for many tasks across industries, such as ChatGPT, Gemini, Claude and Microsoft Copilot. Companies often roll them out as a custom assistant with their own instructions.
• Departmental: one business function in any industry: coding agents such as Cursor, Claude Code or Codex, customer service such as Intercom Fin.
• Vertical: vertical AI for one industry with its data, rules and language: Abridge and OpenEvidence in health, Harvey, Legora and EvenUp in law.
• Embedded: embedded AI as a feature in software you already use, such as Intuit Assist or the AI functions in Canva and Notion.
• AI-native: AI-native products that would not exist without AI: Suno for music, ElevenLabs for voice, Lovable for building apps.

An agent is an application, or part of one, in which the model plans its own steps. The label is popular: Gartner uses the term agent washing for vendors who rebrand assistants and chatbots as agents.

The wrapper debate: what makes an application hard to copy

Wrapper is a term of criticism. It describes a product that adds little more than an input box to someone else's model. In January 2023 a16z found many generative AI apps hard to tell apart because they relied on similar models, and saw no systemic moats in the market yet, meaning lasting advantages that keep competitors out (investor analysis).

Two forces keep the worry alive. Raw model capability gets cheap fast: Epoch AI measured that the price of reaching GPT-4-level performance on a set of PhD-level science questions fell about 40-fold per year, and by 9 to 900 times a year across other milestones. And the labs ship features themselves. In a16z's 2026 consumer ranking Midjourney fell from the top 10 to rank 46 after the big labs' own image generation improved, while music and voice, which the labs had left alone, gave independents room.

Defensible is what the model provider lacks. Trust is part of it: in IBM's 2025 breach study, 13% of breached organisations reported attacks on AI models or applications, and 97% of those lacked proper access controls for AI (study by the Ponemon Institute for IBM, which sells security).

• Own data and context: information a lab cannot get, such as customer history, domain documents and the feedback from every correction.
• Workflow: the product sits where the work happens and becomes its record, so switching hurts. For your own customers this is vendor lock-in, which they will weigh.
• Distribution: existing customers, channels and brand. Software gets easier to build; reaching buyers stays hard.
• Trust: security, compliance and reliability that you can show with evals, logs and audits.

These four sources of defensibility are an Aulory synthesis of the cited investor, research and security sources. No study ranks them.

Build, buy or customise

Companies increasingly buy. In Menlo's survey of 495 US enterprise decision-makers, 53% of AI use cases were bought in 2024 and 76% in 2025 (investor survey). MIT's NANDA project reported in 2025 that bought tools and partnerships succeed more often than internal builds. Press reports of the study disagree on the size of the gap (67% against 33%, or against about 22%), and we could not check the primary report. The decision is known as build vs buy.

Building is expensive. Gartner put the cost of generative AI deployment approaches at $5 million to $20 million in July 2024 and forecast that at least 30% of generative AI projects would be abandoned after the proof of concept by the end of 2025, citing poor data, weak risk controls, rising costs and unclear value (forecast; outcome not verified).

Customising has a legal catch in the EU. A deployer who puts its own name on a high-risk AI system, changes one substantially, or turns a general tool into a high-risk use is treated as its provider (Art. 25 of the AI Act). Rebranding a chatbot as a tool for screening job applicants is the textbook case.

• Buy: standard functions such as writing help, meeting notes or general chat. Fast and cheap, and the vendor carries maintenance.
• Customise: a bought tool or a model via API, plus your instructions, your data via retrieval and your connectors. Fine-tuning is rarely needed for this.
• Build: where the result is your differentiation, and only with evals, monitoring and a team that stays on it.

Rule of thumb, Aulory synthesis of Menlo, NANDA and the AI Act: buy the standard, customise with your context, build your differentiation, then check your role under the AI Act.

Pricing, margins and cost per task

Classic software costs almost nothing for each extra user. An AI application pays for every answer, because each one needs inference on rented chips. That shapes both prices and margins.

Three pricing models appear in public price lists. Microsoft charges a list price of $21 per user and month for the Copilot Business add-on in the US and bills advanced agent features by usage through credits. Salesforce charges $2 per conversation for external Agentforce agents, or $500 per 100,000 Flex Credits. Intercom bills its support agent Fin from $0.99 per outcome; used with another helpdesk it needs no seats (vendor price pages, read on 25 October 2026).

The vendor defines what an outcome is. Intercom's billable outcomes include resolutions and also certain handoffs, routings and qualifications, so the price covers more cases than solved problems alone.

ICONIQ's 2026 survey of about 300 AI builders shows the shift: consumption-based pricing rose from 35% to 42% of companies within six months, outcome-based pricing from 18% to 23%, and companies blend 1.7 pricing models on average. Their gross margin on AI products averaged 45% in 2025, projected to reach 53% in 2026 (investor survey, mostly North America). In 2023 a16z estimated that many AI apps ran at 50 to 60%, with 20 to 40% of revenue going to inference and customer-specific model tuning.

Costs drift. Retries, tool loops and longer contexts add up, which is why ICONIQ calls token counts a poor success metric. About 20% of respondents in McKinsey's 2026 survey say AI operating costs, including tokens, limit their use (self-reported). Language matters as well: the same text can need up to 15 times more tokens in some languages, so speakers of those languages pay more, wait longer and fit less into the context (Petrov and co-authors, NeurIPS 2023).

• Per seat: a fixed price per user and month. Easy to budget, and the vendor carries the risk of heavy users.
• By usage: per token, credit, call or conversation. Close to the real cost structure, harder for buyers to budget.
• By outcome: outcome-based pricing per resolved case or finished task. It ties price to value when the outcome is clearly defined.

Cost per task, an Aulory formula built from the cited components: input and output tokens times their prices, multiplied by all model calls including retries and tool loops, plus retrieval and tool fees, plus the time people spend reviewing. Measure it per successful task.

Measuring value, and what randomised studies show

Value shows up on three levels: does the system do the task, do people use it, and does it change results. For the first level, evals give your product a fixed set of tasks with automatic grading. Anthropic's engineers recommend starting with 20 to 50 tasks taken from real failures. For customer-facing flows they track pass^k, the chance that all of k attempts succeed: a system that is right 75% of the time passes three runs in a row only about 42% of the time (vendor guidance). The gap comes from non-determinism, the same input giving different outputs.

Self-reports are weak evidence. In a randomised study by METR, 16 experienced open-source developers expected AI tools to speed them up by 24% and afterwards believed they had been 20% faster. Measured, they were 19% slower with tools from early 2025 (small sample; tools have changed since). In Microsoft's 2024 survey, 59% of leaders worried about quantifying AI's productivity gains (vendor survey).

Randomised studies, in which chance decides who gets the tool, give the cleanest answers. They show real gains on specific tasks, largest for less experienced people, and small or no effects on whole jobs so far. Fabrizio Dell'Acqua, Ethan Mollick and co-authors call the uneven shape of AI ability the jagged frontier, a case of jagged capability: inside it AI helps a lot, outside it people do worse with AI than without.

• Writing, USA: 453 college-educated professionals in an online experiment finished writing tasks about 40% faster and with 18% higher quality using ChatGPT (Noy and Zhang, Science 2023).
• Consulting, global firm: 758 BCG consultants finished 12.2% more tasks, 25.1% faster and with higher quality inside the frontier. On a task outside it they were 19 percentage points less likely to be correct (Dell'Acqua et al., working paper 2023).
• Customer support: 5,179 agents resolved 14% more cases per hour, newcomers 34% more, experienced agents barely more (Brynjolfsson, Li and Raymond, Quarterly Journal of Economics 2025; staggered rollout).
• Office AI, 66 firms: 7,137 knowledge workers given Microsoft 365 Copilot spent about 1.4 hours less per week on email, with no detectable change in the amount or mix of their tasks (Dillon et al., 2025; co-authors include Microsoft researchers).
• Entrepreneurs, Kenya: small business owners given a GPT-4 business assistant on WhatsApp showed no average effect. Strong performers gained about 15%, weak ones did about 8% worse (Otis et al., Harvard working paper 2023).
• Whole jobs, Denmark: linked survey and register data show precise null effects of chatbot use on earnings and recorded hours; effects above 2% are ruled out two years after adoption (Humlum and Vestergaard, working paper 2025).

Almost all rigorous studies come from the US, Europe or multinational firms. The Kenya experiment is the main field study from the Global South in our research.

Who uses AI, and whose work it touches

Adoption figures differ widely, mostly because they count different people. Microsoft's diffusion index uses its own anonymised telemetry, adjusted for device shares and population, and counts people aged 15 to 64 who used a generative AI product: 18.8% worldwide in the second quarter of 2026, 28.8% in the Global North and 16.2% in the Global South. The United Arab Emirates (73.3%) and Singapore (64.3%) lead (vendor data; Microsoft sells Copilot).

National surveys often show higher levels because they ask online panels or internet users; the table below puts both side by side. China shows the largest gap: Microsoft counts 16.4%, while the internet agency CNNIC reports more than 700 million users of generative AI in mid-2026 and a penetration above 50% on an unclear base. Microsoft itself expects upward revisions for China, where its telemetry misses domestic apps.

Firms show the same pattern. Official statistics that count every small company find about one in five or fewer: 19.95% of EU firms with 10 or more employees in early 2025 (Eurostat), 19.8% of US employer businesses in May 2026 (US Census Bureau) and 12.2% of Canadian businesses in mid-2025 (Statistics Canada). An OECD survey of more than 5,000 SMEs in seven countries in Europe, North America and East Asia found generative AI in 31% of them (2024). Surveys of managers, often in larger firms, report 85 to 98%, for example McKinsey's nearly nine in ten. The denominator drives the gap.

Size matters everywhere. In the EU, 55.03% of large firms used AI against 17% of small ones, and lack of expertise was the top barrier (70.89% of firms that had considered AI). In the US, 37% of firms with 250 or more employees used AI, against under 20% of firms with one to four. The World Bank's 2026 enterprise survey of formal firms reports about 40% in Nigeria and 23% in India (press reports of the survey; the two reports give 40% or 43% for the US).

For single assistants, only vendors publish user numbers, and they use different yardsticks. OpenAI reported 700 million weekly users of ChatGPT in September 2025; higher figures for 2026 circulate only as press reports of internal data. Google reports more than 1 billion monthly users of the Gemini app (August 2026, relayed by a secondary source). Weekly and monthly counts cannot be compared: for the same product a monthly count is always higher. About 70% of consumer ChatGPT conversations concern life outside work (OpenAI with economist David Deming, 2025).

Whose work AI touches is measured as exposure: the share of tasks in a job that AI could do or support. Exposure says nothing yet about job losses, and two international organisations reach different results because they measure differently.

• ILO, 2025: about one in four workers worldwide is in an occupation with some generative AI exposure: 34% of employment in high-income countries, 11% in low-income ones. The ILO sees transformation of jobs as the most likely effect, with clerical work most exposed.
• IMF, 2024: almost 40% of global employment is exposed to AI: about 60% in advanced economies, 40% in emerging markets and 26% in low-income countries. In advanced economies roughly half of the exposed jobs may benefit from AI.
• Why they differ: the ILO grades exposure to generative AI by degree; the IMF counts exposure to AI in general, including jobs where AI complements people.
• Gender gap: in the highest exposure group, the ILO counts 4.7% of women's employment against 2.4% of men's worldwide, and 9.6% against 3.5% in high-income countries.
• Young workers, USA: in payroll data, employment of 22 to 25 year olds in the most AI-exposed occupations is 19% below the path of less exposed peers, mainly through less hiring. An earlier version of the study reported 13%; the authors call the findings early and descriptive (Brynjolfsson, Chandar, Chen, August 2026).

Why AI projects fail, and the problem of shadow AI

The best qualitative evidence comes from RAND. Its researchers interviewed 65 experienced data scientists and engineers in 2023 and found five root causes, listed below. A better model fixes none of them.

Popular failure rates need care. RAND's often quoted figure that more than 80% of AI projects fail is an estimate it cites from a 2022 press article. MIT NANDA's claim that 95% of generative AI pilots show no measurable profit effect comes from one report that critics call methodologically weak. Gartner's figure of at least 30% of generative AI projects abandoned after the proof of concept by the end of 2025 is a forecast. A more solid signal comes from McKinsey's 2026 survey: 37% of respondents see any profit effect from AI, and about 6% get at least 5% of their operating profit from it (self-reported).

Shadow AI means staff using AI tools the company has not approved. In Microsoft's 2024 survey of 31,000 knowledge workers in 31 markets, 78% of AI users brought their own tools, 80% in small and medium companies, and only 39% had received AI training from their employer (vendor survey). BCG found in 2025 that more than half of employees would use other tools anyway if the official ones fall short. In IBM's 2025 breach study, one in five organisations had a breach caused by shadow AI; high shadow AI use added about $670,000 to the average breach cost, and 63% had no AI policy or were still writing one (vendor study). Each such breach can mean data leakage of customer or company information.

Adoption is a question of people. In BCG's 2025 survey of more than 10,600 employees in 11 countries and regions, regular use among frontline staff stalled at 51%. Positive sentiment rose from 15% to 55% with strong leadership support, yet only about a quarter of frontline staff reported such support. Regular use was much higher among people with at least five hours of training plus in-person coaching.

• Wrong problem: leaders misunderstand or miscommunicate which problem AI should solve.
• Missing data: the data the system needs is missing or unusable.
• Technology first: teams chase the newest technology and lose sight of the user's problem.
• Thin infrastructure: too little investment in data pipelines and in running the system in production.
• Too hard: the problem lies beyond what current AI can do.

RAND's interviews took place before most companies used generative AI and cover machine learning broadly. ICONIQ's 2026 builders name similar hurdles for agents: reliability across many steps, missing context and memory, cost drift and testing that only reacts to problems.

Responsible deployment: the rules by region

Under the EU AI Act, your role decides your duties. Whoever develops a system and offers it under their own name is its provider; whoever uses it under their own authority at work is a deployer. Most small companies are deployers. This is a neutral overview for orientation; it does not replace legal advice.

Two duties apply already. Art. 4 asks providers and deployers to take measures to support the AI literacy of their staff; since the 2026 amendment it requires no guaranteed level for individuals and carries no fine of its own. Art. 50 has applied since 2 August 2026: providers must design chatbots so people know they are talking to an AI and must mark synthetic content in a machine-readable way (labelling); deployers must disclose deepfakes and tell people when emotion recognition is used. Systems already on the market have until 2 December 2026 for the marking.

High-risk uses are listed in Annex III, among them recruitment and screening of applicants, performance monitoring at work, admission and grading in education, creditworthiness, and pricing of life and health insurance. Deployers of such systems must follow the provider's instructions, assign human oversight to people with training and authority, keep logs for at least six months, inform workers before use at work and inform the people affected (Art. 26). After the AI Omnibus, Regulation (EU) 2026/1744 in force since 27 July 2026, these duties apply from 2 December 2027, and from 2 August 2028 for products under Annex I. Fines reach up to EUR 15 million or 3% of worldwide turnover for most duties and EUR 35 million or 7% for banned practices; for SMEs and start-ups the lower amount applies.

Elsewhere, labelling and transparency rules spread fastest, while duties for consequential decisions were softened or delayed in 2026. That is our reading of the sources below.

• China: labels for AI-generated content mandatory since 1 September 2025, visible and in the file's metadata. App stores check the labelling during review, and users must declare synthetic content they publish.
• USA, Colorado: the original AI Act was challenged in court and replaced by SB 26-189, effective 1 January 2027: notice to consumers for consequential decisions, an explanation within 30 days after a negative outcome, data correction and human review. The duty to assess algorithmic discrimination was dropped.
• USA, California: since 2 August 2026 the AI Transparency Act requires providers of generative AI with 1 million or more monthly users in California to offer a free detection tool and disclosures. Companion chatbots have had duties since 1 January 2026, and rules on automated decision-making take effect in steps up to 2028.
• India: no AI statute. Amended IT Rules in force since 20 February 2026 require labels on synthetic content and removal within 3 hours on official order, within 2 hours for deepfakes and non-consensual intimate images. National AI governance guidelines are voluntary.
• South Korea: AI Basic Act in force since 22 January 2026: labelling of generative output, advance notice to users, duties for high-impact AI and a local representative for large foreign firms. Fines reach KRW 30 million, with a one-year grace period announced.
• Japan: AI Promotion Act of May 2025 without penalties. Businesses are asked to make reasonable efforts and to cooperate with government policy.
• Brazil: bill PL 2338/2023 passed the Senate in December 2024 and was still awaiting a committee report in the Chamber in September 2026 (secondary source).

Responsibility also means reach. An AI app is software, so accessibility standards such as WCAG 2.2 apply: test chat interfaces with screen readers and with the keyboard alone. Evaluate every language you serve with its own test set, since token costs and quality differ by language (Petrov et al.).

A founder's guide: from idea to responsible launch

The following order is an Aulory synthesis. Each step rests on the evidence on this page; no study has tested the sequence as a whole.

1. 01One problem, one metric. Name one problem and one number that shows success, and keep a team on it for at least a year (RAND).
2. 02Buy, customise, build. Buy the standard, customise with your context, build only your differentiation, and check whether customising makes you a provider (build vs buy; AI Act Art. 25).
3. 03Design the context. Write the system prompt, add a few worked examples and load your data just in time through retrieval (Anthropic).
4. 04Least privilege. Give tools only the rights they need and put a person in front of irreversible actions (human in the loop; OWASP).
5. 05Evals before launch. Write 20 to 50 evals from real failures, track pass^k for customer-facing flows and defend against prompt injection. Only 44% of ICONIQ's builders detect it (Anthropic; OWASP; ICONIQ).
6. 06Price with the cost in view. Know your cost per successful task before you choose seat, usage or outcome pricing, and check it every month (ICONIQ; Intercom).
7. 07Measure value step by step. Evals first, then task success in production, active use, time and quality against a baseline, and finally the business metric. Treat self-reported gains as the weakest evidence (METR).
8. 08Clarify your role. Provider or deployer? Disclose AI chats and deepfakes, and check Annex III before you touch hiring, credit or education (AI Act Art. 3, 6, 25, 26, 50).
9. 09Answer shadow AI with a better offer. An approved tool with good data terms, a one-page AI policy and at least five hours of hands-on training (Microsoft; BCG; IBM; AI Act Art. 4).
10. 10Reach everyone. Meet WCAG 2.2 at level AA and test every language you serve (W3C; Petrov et al.).

What the layer is made of

Key figures, with source and date

• 18.8%of the world's working-age population used generative AI in Q2 2026, by Microsoft's telemetry. Global North 28.8%, Global South 16.2%.Microsoft AI Economy Institute · 21 Sep 2026 · vendor
• 19.95%of EU firms with 10 or more employees used AI in early 2025, from 5.21% in Romania to 42.03% in Denmark. Large firms 55.03%, small ones 17%.Eurostat · Dec 2025 · government
• 19.8%of US employer businesses used AI in some business function in May 2026. Among firms with 250 or more employees it was 37%.US Census Bureau · 26 May 2026 · government
• 700 millionweekly users of ChatGPT, as OpenAI reported in September 2025. Newer figures exist only as press reports of internal data.OpenAI · 15 Sep 2025 · vendor
• $19 bnspent by US companies on AI applications in 2025, coding alone $4.0 bn. AI features inside existing software are not counted.Menlo Ventures · 9 Dec 2025 · investor
• 45%average gross margin on AI products in 2025, projected 53% for 2026. Consumption-based pricing rose from 35% to 42% of companies in six months.ICONIQ, State of AI 2026 · Jul 2026 · investor
• +14%more cases resolved per hour by 5,179 support agents with AI, +34% for newcomers, barely any gain for experienced agents.Brynjolfsson, Li, Raymond · 2025 · research
• 1 in 4workers worldwide is in an occupation with some generative AI exposure. In the highest group: 4.7% of women's jobs, 2.4% of men's.ILO, Working Paper 140 · 20 May 2025 · policy research
• about 6%of companies earn at least 5% of their operating profit from AI; 37% report any profit effect.McKinsey, State of AI 2026 · 25 Aug 2026 · consultancy, self-reported

Regions compared

The table puts two ways of counting people side by side. Microsoft's index applies one method to every country but sees only what its telemetry sees. National surveys ask people directly, often online, and count different populations. Both are useful once you name the population.

Use and building can sit far apart. The Latin American AI Index finds that 19 countries in the region receive 14% of global visits to AI web tools but 1.12% of global AI investment, against 6.6% of world GDP (CENIA and ECLAC, 2025).

Sources: Microsoft AI Diffusion Report Q1 2026 and blog (21 Sep 2026), Pew (Feb 2026), CNNIC via ECNS (29 Sep 2026), MIC Information and Communications White Paper 2026, KISDI via Korea Times (28 May 2026), CETIC.br (2026), Google/Ipsos via Techpoint Africa (21 Jan 2026), Eurostat (Dec 2025), US Census Bureau (26 May 2026), Statistics Canada (11 Sep 2025), World Bank survey via Nairametrics (4 Aug 2026) and TechEnclave (10 Oct 2026), law-firm summaries for the rules. Online panels overstate use in the whole population; surveys and telemetry are not directly comparable.

Examples

• Intercom bills its service agent Fin at $0.99 per resolved case, not per licence.
• Midjourney fell out of the a16z top 10 to rank 46, because the big labs now make good images themselves.
• Companies buy more often than they build: 76% of use cases in 2025 were bought (Menlo Ventures).

Why it matters

The application layer is where AI turns into money, or fails to. Surveys of managers find AI in almost every company, official statistics count about one in five firms, and few report a large profit effect. The gap between what models can do and how people use them is the founders' opportunity. It is also their risk: a product that sits too close to what the model labs ship can lose its reason to exist with one release.

It is also where responsibility lands. Whoever offers or uses an application decides what data goes in, what the system may do and who checks it. Rules on labelling, transparency and high-risk uses now exist in the EU, China, South Korea, India and several US states, and they all attach to the application.

Common misconceptions

• A better model will fix our AI project.

  The five root causes RAND found are mostly organisational: the wrong problem, missing data, thin infrastructure (RAND, 2024).

• More context always gives better answers.

  Models recall information less reliably as the context window fills, a context rot effect. Good applications hand the model the smallest set of high-signal information (Anthropic, vendor guidance).

• AI makes everyone faster at everything.

  Inside the jagged frontier of AI capability BCG consultants were 25% faster; on a task outside it they were 19 percentage points less often correct (Dell'Acqua et al., 2023).

• We banned ChatGPT, so we have no AI risk.

  78% of AI users bring their own tools, and one in five organisations had a breach caused by shadow AI (Microsoft 2024; IBM 2025).

• AI software earns the margins of classic software.

  Every answer costs inference. ICONIQ's builders report an average gross margin of 45% for 2025 (investor survey).

• Outcome pricing means you pay only when the problem is solved.

  The vendor defines the outcome. Intercom's billable outcomes include certain handoffs and routings (price page, October 2026).

• 80% or 95% of AI projects fail.

  RAND's 80% is cited from a press article, MIT NANDA's 95% is contested, and Gartner's abandonment figures are forecasts. No representative measurement exists.

Perspectives from leading experts

Ten perspectives from investors, vendors, economics, labour policy and critical scholarship, from the USA, China, India, South Africa and Geneva. Each summary paraphrases a published work, linked below it. Investors and vendors are marked as such.

• Sonya Huang and Pat Grady

  Partners at Sequoia Capital, USA · investor

  Pre-trained knowledge, reasoning at answer time and long-running agents now combine into systems that can work things out. Citing METR, they note that the length of tasks agents can complete doubles roughly every seven months. They advise founders to build agents that do work, move interfaces from chat to delegation, invest in reliability and price around outcomes.

  2026: This is AGI (Sequoia Capital) · 14 Jan 2026

• Ethan Mollick

  Professor at the Wharton School, University of Pennsylvania, USA · academic

  Today's models can already do far more than almost anyone uses them for. He calls this gap a capability overhang and sees it as an opportunity. Deep knowledge, broad knowledge, taste and agency make human work with AI more valuable, and he asks for ways of working that strengthen people.

  The Overhang (One Useful Thing) · 18 Sep 2026

• Erik Brynjolfsson and co-authors

  Stanford Digital Economy Lab, USA · academic

  US payroll data show no economy-wide job losses from AI so far. Young workers aged 22 to 25 in exposed occupations, however, are employed 19% below the path of less exposed peers, mostly because fewer are hired. Declines appear where AI substitutes tasks; where it complements workers, employment holds or grows.

  Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence · 12 Aug 2026

• Daron Acemoglu

  Economist at MIT, USA · academic

  In a task-based model, the gain for the whole economy depends on how many tasks AI affects and how much it saves per task. With current estimates, total factor productivity rises by at most 0.66% over ten years, and by less once harder, context-dependent tasks are counted. He expects AI to widen the gap between income from capital and from labour.

  The Simple Macroeconomics of AI (Economic Policy) · 2025

• Arvind Narayanan and Sayash Kapoor

  Computer scientists at Princeton University, authors of AI Snake Oil, USA · academic

  Better AI methods turn into economic impact only through applications and their spread through organisations, and that takes decades. Benchmark scores say little about usefulness in practice. They cite estimates that 40% of US adults used generative AI in August 2024, yet this amounted to only 0.5% to 3.5% of work hours.

  AI as Normal Technology (Knight First Amendment Institute) · 15 Apr 2025

• Joy Buolamwini and Timnit Gebru

  Computer scientists; Buolamwini founded the Algorithmic Justice League, USA · critical research

  Two common face datasets consisted of 79.6% and 86.2% lighter-skinned people. On a balanced test set, three commercial gender classifiers misclassified darker-skinned women up to 34.7% of the time and lighter-skinned men at most 0.8%.

  Gender Shades (Proceedings of Machine Learning Research) · 2018

• Kai-Fu Lee

  Founder and CEO of 01.AI, Beijing, China · vendor

  01.AI stopped the most cash-hungry work, pre-training its own foundation models, and moved to enterprise applications for top management. Many CEOs want AI transformation but end up with demos that change neither productivity nor financials. Demand is there; the bottleneck is people who can deliver.

  Interview with Bloomberg, reported by 36Kr · 27 Jul 2026

• Nandan Nilekani

  Co-founder of Infosys, architect of India's Aadhaar digital ID, India · policy research

  India's chance lies in practical AI applications, because infrastructure becomes a commodity and value moves up to the application layer. To reach a billion people, AI must work in Indian languages and by voice, built on digital public infrastructure for farming, health and education. He expects work to spread across a million one-person companies.

  Speech at the NCAER India Policy Forum, reported by Business Today · 7 Aug 2026

• Pawel Gmyrek, Janine Berg and co-authors

  Researchers at the International Labour Organization, Geneva · policy research

  About one in four workers worldwide is in a job with some generative AI exposure, 34% in high-income and 11% in low-income countries. In the most exposed group women are about twice as affected as men, largely through clerical work. Transformation of jobs is the most likely effect.

  Generative AI and Jobs: A Refined Global Index of Occupational Exposure (ILO Working Paper 140) · 20 May 2025

• Rachel Adams and co-authors

  Global Center on AI Governance, South Africa · policy research

  Across 135 countries the average score for responsible AI is about 35 out of 100. Almost all governments state commitments, yet evidence of implementation exists in only 55% of cases, 45% in the Global South. 55 countries have frameworks on gender, and only 24 show implementation.

  Global Index on Responsible AI 2026 · Jul 2026

Terms for this layer

• ModelThe trained file of numbers that turns input into output. On its own it can't browse, remember or act.
• ContextThe information the model can actually see right now. Better context beats clever wording.
• ToolA function an agent is allowed to use, such as search, read a file or send an email.
• AI SystemThe EU AI Act's core legal term: a machine-based system that infers from its input how to produce outputs such as content, recommendations or decisions.
• Context EngineeringThe craft of choosing what goes into the context, and what stays out, so the model can do good work.
• Context WindowThe maximum amount of text a model can take in at once. When it is full, something has to go.
• Context RotThe gradual drop in answer quality as a conversation or context grows long and cluttered.
• Prompt InjectionHidden instructions in content the AI reads that hijack what it does.
• APIThe programmable connection to a model provider. You pay per token and keep more control than in a chat app.
• System PromptStanding instructions that shape every answer, set by the app or by you, before your message arrives.
• Retrieval-Augmented GenerationLooking up relevant documents first and handing them to the model, so it answers from your sources.
• Model Context ProtocolAn open standard for connecting AI apps to tools and data sources, so integrations are built once and reused.
• Tool CallThe model's structured request to use a tool. Software around the model decides whether to run it.
• GuardrailsChecks and limits built around an AI system to block unwanted inputs, outputs or actions before they cause harm.
• PermissionsThe rules that decide what an AI agent may do on its own, what it must ask about, and what it may never do.
• EvalsRepeatable tests that show whether your AI setup actually works, and whether a change made it better or worse.
• Custom AssistantA reusable AI assistant set up with your own instructions, files and tools, such as GPTs, Projects or Gems.
• Coding AgentAn AI agent that reads, writes and runs code in a project, working through programming tasks step by step.
• Vertical AIAI applications built for one industry, with its data, rules and language, such as law, health or real estate.
• Embedded AIAI built as a feature into software you already use, such as a writing assistant in your word processor or email.
• AI-NativeWorking, building or organising with AI built in from the start, instead of adding it on top of old processes.
• AI AgentA model connected to tools and given a goal, so it can take steps on its own instead of only answering.
• AI WrapperA product that adds little more than an interface to someone else's AI model, and is easy to copy for that reason.
• Vendor Lock-inWhen switching to another AI provider becomes so costly or painful that you effectively can't.
• Build vs BuyThe decision whether to develop an AI solution yourself, buy a ready product, or customise one with your own data.
• Fine-TuningTraining an existing model further on your own examples to change its behaviour. Rarely the first step.
• InferenceRunning a finished model to get an answer. What you pay for every time you send a message.
• TokenThe small chunk of text a model reads and writes. Price, speed and memory limits are all counted in tokens.
• Gross MarginThe share of revenue left after paying the direct costs of delivering a product. For AI products, inference is one of those costs.
• Outcome-Based PricingPaying for AI per result, such as a resolved support case or a finished task.
• Cost per TaskWhat one finished piece of work really costs with AI: tokens, tools, retries and the human time to check it.
• Non-DeterminismThe same question can get a different answer each time.
• Jagged CapabilityThe uneven skill profile of AI models: excellent at some hard tasks, unreliable at some easy ones.
• Excessive AgencyWhen an AI system has more functions, rights or autonomy than its task needs, so a wrong or manipulated output can cause real damage.
• ObservabilitySystematically recording what an AI system does, every model call, tool call and result, so you can see why it acted as it did.
• Generative AIAI that creates new content, such as text, images, audio, video or code, instead of only sorting or predicting.
• Shadow AIAI tools employees use for work without the company knowing or approving them.
• AI PolicyYour organisation's short, practical rules for which AI tools may be used, with which data, and who checks what.
• Data LeakageConfidential information ending up where it shouldn't: in a prompt, a log, or another user's answer.
• EU AI ActThe EU's law for AI, sorting uses by risk and setting duties for those who build and deploy AI.
• ProviderUnder the EU AI Act, whoever develops an AI system or model, or has it developed, and offers it under their own name. Providers carry the most duties.
• DeployerUnder the EU AI Act, anyone who uses an AI system under their own authority at work. If your company uses AI tools, this is your role.
• AI LiteracyThe knowledge and skills to understand, use and judge AI. In the EU, organisations must actively support it.
• AI Labelling & WatermarksMarking content and interactions as AI-generated, visibly for people and invisibly in the file for machines.
• DeepfakeRealistic fake images, audio or video of real people or events, generated or manipulated with AI.
• High-Risk AI SystemAI used in sensitive areas like hiring, credit or education, which faces the strictest EU rules.
• Human OversightThe legal and practical duty that people can monitor, understand and override an AI system.
• Human in the LoopA person reviews or approves what the AI does before it counts.

Sources

1. The continued state of global AI diffusion in 2026, Microsoft AI Economy Institute (vendor), 21 Sep 2026
2. AI Diffusion Report 2026 Q1, Microsoft (vendor), May 2026
3. Americans and AI 2026: chatbot use, Pew Research Center, Jun 2026
4. More than 700 million people in China use generative AI (CNNIC), ECNS / China News Service, 29 Sep 2026
5. Information and Communications White Paper 2026, Ministry of Internal Affairs and Communications, Japan, Jul 2026
6. Four out of 10 Koreans used generative AI in 2025, Korea Times, citing KMCC and KISDI, 28 May 2026
7. TIC Domicílios 2025, executive summary, CETIC.br, May 2026
8. Google report: Nigeria leads AI learning (Google/Ipsos survey), Techpoint Africa, 21 Jan 2026
9. Use of artificial intelligence in enterprises, Eurostat, Dec 2025
10. AI use among businesses, US Census Bureau, 26 May 2026
11. Analysis in Brief: Canadian Survey on Business Conditions, AI use, Statistics Canada, 11 Sep 2025
12. Generative AI and the SME Workforce, OECD, Nov 2025
13. World Bank: AI adoption among Nigerian firms jumps to 40%, Nairametrics, 4 Aug 2026
14. World Bank survey finds AI use in 23% of formal Indian firms, TechEnclave, 10 Oct 2026
15. Latin American Artificial Intelligence Index (ILIA) 2025, CENIA and ECLAC, 18 Dec 2025
16. How people are using ChatGPT, OpenAI (vendor), 15 Sep 2025
17. Google puts Gemini at 1 billion monthly users, implicator.ai, 12 Aug 2026
18. 2025: The State of Generative AI in the Enterprise, Menlo Ventures (investor), 9 Dec 2025
19. 2026 State of AI: The Builder's Economy, ICONIQ (investor), Jul 2026
20. Who Owns the Generative AI Platform?, Bornstein, Appenzeller, Casado, a16z (investor), 19 Jan 2023
21. Top 100 Gen AI Consumer Apps, 6th edition, Olivia Moore, a16z (investor), 9 Mar 2026
22. LLM inference price trends, Epoch AI, 12 Mar 2025
23. Cost of a Data Breach 2025, IBM and Ponemon Institute (vendor), 30 Jul 2025
24. Effective context engineering for AI agents, Anthropic (vendor), 29 Sep 2025
25. Demystifying evals for AI agents, Anthropic (vendor), 9 Jan 2026
26. OWASP Top 10 for LLM Applications 2025, OWASP GenAI Security Project, 2025
27. Gartner predicts 30% of generative AI projects will be abandoned after proof of concept, Gartner, 29 Jul 2024
28. Microsoft 365 Copilot for business, pricing, Microsoft (vendor), 25 Oct 2026
29. Agentforce pricing, Salesforce (vendor), 25 Oct 2026
30. Fin pricing, Intercom (vendor), 25 Oct 2026
31. Language Model Tokenizers Introduce Unfairness Between Languages, Petrov, La Malfa, Torr, Bibi, NeurIPS, 2023
32. Experimental evidence on the productivity effects of generative artificial intelligence, Noy and Zhang, Science, Jul 2023
33. Navigating the Jagged Technological Frontier, Dell'Acqua, Mollick et al., HBS Working Paper, Sep 2023
34. Generative AI at Work, Brynjolfsson, Li, Raymond, Quarterly Journal of Economics, 2025
35. Shifting Work Patterns with Generative AI, Dillon, Jaffe, Immorlica, Stanton, 13 Nov 2025
36. Randomised controlled trial with experienced open-source developers (arXiv 2507.09089), METR, 12 Jul 2025
37. The Uneven Impact of Generative AI on Entrepreneurial Performance, Otis, Clarke, Delecourt, Holtz, Koning, Harvard, Dec 2023
38. Large Language Models, Small Labor Market Effects, Humlum and Vestergaard, BFI Working Paper, 15 Apr 2025
39. Generative AI and Jobs: A Refined Global Index of Occupational Exposure, ILO, 20 May 2025
40. AI will transform the global economy, Kristalina Georgieva, IMF, 14 Jan 2024
41. The Root Causes of Failure for Artificial Intelligence Projects, Ryseff, De Bruhl, Newberry, RAND, 13 Aug 2024
42. MIT report: 95% of generative AI pilots failing, Fortune, 18 Aug 2025
43. The state of AI in 2026, McKinsey, 25 Aug 2026
44. Work Trend Index 2024: AI at work is here, Microsoft and LinkedIn (vendor), 8 May 2024
45. AI at Work 2025: Momentum Builds, but Gaps Remain, BCG, Jun 2025
46. AI Act, consolidated text incl. AI Omnibus (Art. 3, 4, 25, 26, 50, 99, Annex III), European Commission, AI Act Service Desk, 27 Jul 2026
47. EU AI Act: AI Omnibus formally adopted, Addleshaw Goddard, 1 Jul 2026
48. Colorado AI law in flux, McDermott Will & Schulte via JD Supra, 28 May 2026
49. California AI laws, terms.law, 19 Jul 2026
50. China's AI labeling measures take effect September 1, Loeb & Loeb, Mar 2025
51. India AI Regulatory Tracker, 2nd edition, AZB & Partners, 8 Sep 2026
52. South Korea's AI Basic Act: overview, Cooley, 27 Jan 2026
53. Japan's first AI legislation becomes law, White & Case, 14 Apr 2026
54. Brazil AI bill PL 2338 status, CASRAI (secondary), 25 Sep 2026
55. Web Content Accessibility Guidelines (WCAG) 2.2, W3C, 12 Dec 2024