Business
Last reviewed
May 13, 2026
Sources
40 citations
Review status
Source-backed
Revision
v2 ยท 5,607 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 13, 2026
Sources
40 citations
Review status
Source-backed
Revision
v2 ยท 5,607 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: Business ChatGPT Plugins
Business uses of artificial intelligence cover the application of machine learning, generative AI, and AI agents inside companies and other organizations to produce goods and services, manage operations, and serve customers. The current wave, which began in late 2022 after the release of ChatGPT, shifted enterprise spending from earlier predictive analytics and robotic process automation toward large language models embedded in office software, customer service, sales, finance, legal, and software development. By 2024 most large companies had at least begun trials, and McKinsey's annual survey found 65 percent of respondents reporting regular use of generative AI inside their organizations, up from 33 percent ten months earlier (McKinsey).
The vendor landscape consolidated quickly. Microsoft packaged its Microsoft Copilot into Microsoft 365 at $30 per user per month for enterprise customers. Salesforce launched Agentforce in October 2024. ServiceNow set a target of $1 billion in annual contract value from its Now Assist generative AI by 2026. Cursor and GitHub Copilot raced to dominate coding. Anthropic, OpenAI, and Google signed enterprise deals with banks, consultancies, law firms, and drug makers. Productivity studies from MIT, Harvard, BCG, and Stanford reported double-digit gains for many knowledge tasks, though high-profile failures such as Air Canada's chatbot ruling and Klarna's reversal of its all-AI customer service plan tempered expectations.
The term covers a wide range of uses. On one end are mature, narrowly scoped applications: fraud detection in card networks, demand forecasting in retail supply chains, propensity models in marketing, and routing in logistics. These were built mostly with classical machine learning and have been in production for years. On the other end is generative AI, which writes drafts, summarizes documents, generates code, answers customer questions, and increasingly takes multi-step actions through agentic AI. Most large enterprises now run both kinds of system, often layered. A bank may keep credit underwriting models built on gradient boosting from the 2010s while using Claude or GPT-4o to summarize disclosures, draft client emails, and answer internal policy questions.
A third strand is workflow automation. Robotic process automation tools from UiPath, Automation Anywhere, and Blue Prism boomed in the late 2010s by recording and replaying clicks across legacy back office systems. After 2023 these vendors repositioned around AI agents, blending deterministic scripts with language models that can read unstructured documents, draft responses, and decide when to escalate. The same blurring is happening at Microsoft, Salesforce, Workday, and SAP. The category of "enterprise software" is being redrawn around AI, and the question for most companies in 2024 and 2025 was less whether to adopt it than how to redesign work so the investment paid back.
The business application of AI predates the recent generative wave by decades. In the 1980s a class of programs called expert systems briefly looked like a turning point. The most famous early system, MYCIN, was built at Stanford in the 1970s to recommend antibiotics for blood infections; it scored about 69 percent accuracy on test cases, comparable to specialists, but never reached patients because hospitals had no way to integrate it into workflows. The first commercial hit was XCON, a rule-based system Digital Equipment Corporation deployed in 1980 to configure VAX minicomputer orders. By 1986 XCON had processed roughly 80,000 orders with reported accuracy of 95 to 98 percent and was credited with saving DEC tens of millions of dollars a year. The brief expert systems boom ended badly. Maintenance costs grew faster than benefits, specialized Lisp machines lost out to general purpose hardware, and the broader AI winter from the late 1980s pushed the field out of corporate IT budgets (Wikipedia).
The next wave was statistical machine learning. From the late 1990s, credit card fraud detection, search ranking, recommendation systems, demand forecasting, and dynamic pricing became routine. By the mid 2010s, deep learning revived image and speech tasks at scale, powering Amazon's recommendations, Google's ad systems, Netflix's content suggestions, and a generation of computer vision deployments in manufacturing and logistics. None of this carried the label "AI" in many boardrooms; it was just analytics.
A parallel track was robotic process automation. UiPath, founded in Bucharest in 2005 as DeskOver, released its first desktop automation product in 2013 and grew into a public company by 2021 (UiPath). RPA worked by recording a human's keystrokes inside legacy enterprise software and replaying them at scale. It was brittle, but it gave finance and shared services teams a way to automate work without rewriting old systems.
The current era opened when OpenAI released ChatGPT on November 30, 2022. The product hit 100 million users within two months, and within weeks corporate IT teams started both blocking and piloting it. Microsoft's $10 billion investment in OpenAI in January 2023 was followed by the announcement of Microsoft 365 Copilot in March 2023. Google launched what would become Gemini for Workspace. Amazon poured up to $8 billion into Anthropic and made Claude available through Amazon Bedrock. By 2024 the dominant theme had shifted from chat assistants to agents, software that not only answers questions but books a meeting, files a ticket, or completes a return.
Measuring adoption is harder than it sounds because most surveys conflate trials, point use cases, and production deployments. The two most-cited recurring sources are McKinsey's State of AI survey and the Stanford AI Index. They roughly agree on the direction even when the numbers differ.
McKinsey's May 2024 survey of 1,363 respondents across 100 countries reported that 65 percent of respondents said their organizations regularly used generative AI in at least one business function, nearly double the 33 percent figure from late 2023. Overall AI adoption (including older predictive AI) rose to 72 percent (McKinsey 2024). The most common functions cited were marketing and sales, product and service development, and IT. Within those, marketing and sales saw the biggest year over year jump.
The Stanford AI Index 2025 reported that 78 percent of organizations used AI in 2024, up from 55 percent the year before. Total corporate AI investment hit $252.3 billion in 2024, with U.S. private AI investment growing to $109.1 billion, more than twelve times China's $9.3 billion. Generative AI specifically attracted $33.9 billion in private investment, up 18.7 percent on the year (Stanford HAI).
McKinsey's 2025 follow up shifted the question from adoption to value capture. Only 39 percent of respondents reported any enterprise level EBIT impact from generative AI, suggesting that the gap between piloting and scaling remained the dominant story. High performers were nearly three times as likely to have redesigned individual workflows rather than bolting AI onto existing ones, which McKinsey identified as the single biggest factor in seeing financial impact (McKinsey 2025).
Gartner placed generative AI in the "trough of disillusionment" in its 2024 hype cycle, noting that the average company had spent $1.9 million on generative AI initiatives while fewer than 30 percent of AI leaders reported their CEOs were happy with the return. The firm still expected over 95 percent of enterprises to have deployed generative AI in production by 2028 (Gartner).
Most companies do not build foundation models. They buy access through a handful of platforms that bundle the models with security, identity, data connectors, and tools for building workflows on top.
| Vendor | Product | Foundation models | Launch / Notable date |
|---|---|---|---|
| Microsoft | Microsoft Copilot for Microsoft 365 | OpenAI GPT family via Azure OpenAI; in-house Phi models | General availability November 1, 2023 at $30 per user per month |
| Microsoft | Azure OpenAI Service | OpenAI GPT, o-series, embeddings | January 2023 GA |
| Gemini for Google Workspace | Gemini | Workspace bundling 2024 | |
| Vertex AI | Gemini, third-party models | 2021 launch, generative additions 2023 to 2024 | |
| Amazon Web Services | Amazon Bedrock | Anthropic Claude, Meta Llama, Cohere, AI21, Amazon Nova, Mistral | GA September 2023 |
| Salesforce | Agentforce | Atlas Reasoning Engine, mix of LLMs | Announced September 12, 2024; GA October 25, 2024 |
| ServiceNow | Now Assist | Mix of LLMs plus in-house models | GA late 2023, target $1B ACV by 2026 |
| SAP | Joule | Multiple LLMs orchestrated | Announced 2023; agents expanded 2024 to 2025 |
| Workday | Illuminate | Workday-trained models on 800B annual transactions | Announced September 17, 2024 at Workday Rising |
| IBM | watsonx | IBM Granite, Llama, Mistral | 2023 |
| Oracle | Oracle AI Apps | Cohere, in-house | 2024 |
| OpenAI | ChatGPT Enterprise / Team | OpenAI GPT and o-series | Launched August 2023 |
| Anthropic | Claude for Work / Enterprise | Claude | Enterprise tier launched September 2024 |
| Glean | Glean Work AI | Multi-model search and assistant | Founded 2019; $200M ARR by December 2025 |
Microsoft's run rate for AI products reached an annual revenue pace of $13 billion by January 2025 and $37 billion by April 2026, up 123 percent year over year, with Azure cloud services growing 40 percent that quarter (Microsoft Q3 FY26). OpenAI's API and ChatGPT business has driven much of that, and OpenAI itself reported more than three million paying business users by mid 2025.
Anthropic's enterprise business expanded along similar lines. By 2024 Pfizer was using Claude on Amazon Bedrock to accelerate drug research while reporting tens of millions of dollars in operational savings, and Bridgewater Associates, ADP, Amdocs, Broadridge, Delta Air Lines, Intuit, LexisNexis, Pfizer, Siemens, and the PGA Tour all appeared on AWS's published customer list (AWS). Over 100,000 customers ran Claude through Bedrock by the end of 2024.
Customer service was the first function where companies tried to fully automate work with generative AI, and it was the first to show the limits. Vendors include Intercom Fin, Zendesk AI Agents, Sierra, Ada, Cresta, and Decagon. Resolution rates vary widely with implementation quality. Intercom reports Fin resolves around 50 percent of tickets out of the box, with well-tuned deployments hitting 80 percent or more; a 2025 Forrester benchmark put Fin at roughly 50 percent and Zendesk AI Agents at about 38 percent on standard configurations (Fini Labs).
Sierra, the agent company co-founded by former Salesforce co-CEO Bret Taylor and Clay Bavor in early 2024, hit a $100 million annual revenue run rate within 21 months and counts Deliveroo, Discord, Ramp, Rivian, SoFi, ADT, Bissell, Vans, Cigna, and SiriusXM as customers (TechCrunch). Sierra raised funding at a $15.8 billion valuation in 2026.
The most-discussed cautionary tale was Klarna. The Swedish payments company announced in February 2024 that its OpenAI-powered assistant had taken on 75 percent of customer chats, handling about 2.3 million conversations in 35 languages, and credited the system with the work of 700 full-time agents. CEO Sebastian Siemiatkowski talked about a workforce cut in half. By 2025 quality scores had fallen, complaints rose, and Klarna began rehiring human agents under a flexible "Uber-style" model. Siemiatkowski conceded that cost had been an overly dominant evaluation factor (Tech.co).
Cresta, spun out of the Stanford AI Lab and backed by Greylock, Sequoia, and Andreessen Horowitz, takes a different angle. Instead of replacing agents it coaches them in real time, scoring 100 percent of calls and surfacing the patterns of top performers as guidance for newer hires (Cresta). That second design pattern, AI as a coach for humans rather than a replacement, is closer to the productivity studies described below than the full-replacement model Klarna tried.
Gong and Outreach dominate revenue intelligence and sales execution respectively. Gong records and transcribes every customer call, applying language models to surface objections, buying signals, and coaching opportunities. Outreach focuses on outbound sequence automation. Salesforce's Agentforce ships with a sales development representative (SDR) agent that qualifies leads, books meetings, and updates the CRM. Microsoft Copilot for Sales pulls CRM data into Outlook and Teams. McKinsey estimates that sales and marketing alone could absorb 28 percent of the total economic value generative AI is projected to create, the largest single share (McKinsey).
Jasper, Copy.ai, Writer, and Adobe Firefly each carved out a niche in enterprise content generation. Jasper's customer list includes Intuit, UnitedHealthcare, UiPath, Spotify, L'Oreal, Uber, and Accenture; its pitch centers on brand voice modeling and marketing-specific workflows. Adobe added Firefly generative imagery to Photoshop and Express in 2023. Coca-Cola, WPP, and Mondelez built dedicated generative AI marketing labs through 2024.
Code generation produced some of the clearest productivity numbers. A controlled experiment by GitHub Next with Microsoft's Office of the Chief Economist had developers implement an HTTP server in JavaScript; those using GitHub Copilot finished in 71 minutes against 161 minutes for the control group, a 55 percent speedup with a 95 percent confidence interval of 21 to 89 percent (GitHub blog). Cursor, an AI-first code editor built on a fork of VS Code by Anysphere, hit $100 million ARR by January 2025, $500 million by June, $1 billion by November, and $2 billion by February 2026, with corporate buyers accounting for about 60 percent of revenue and over half of the Fortune 500 reported as customers (Anysphere). Cognition launched Devin in March 2024 as a more autonomous coding agent and raised at a reported $25 billion valuation by April 2026.
A Microsoft and GitHub study reported Microsoft 365 Copilot users were 29 percent faster across search, writing, and summarization tasks, with 70 percent reporting higher productivity and 68 percent reporting better work quality. Copilot in Word cut editing time by 26 percent; in Outlook it cut email composition by 45 percent (Microsoft). Forrester's commissioned Total Economic Impact study modeled a 112 to 457 percent three-year ROI for a composite organization of 25,000 employees.
The picture is not uniformly rosy. A 2024 Uplevel study of engineering teams found no statistically significant difference in pull request throughput between teams using GitHub Copilot and those without; a separate Visual Studio Magazine report cited similar findings. Acceptance is not the same as correctness, and correctness is not the same as customer value.
The legal industry moved faster than many expected. Harvey, founded in 2022 by Winston Weinberg (a former O'Melveny associate) and Gabriel Pereyra (a former DeepMind researcher), launched its exclusive partnership with the magic circle firm Allen & Overy in February 2023. During the initial trial, 3,500 A&O lawyers ran roughly 40,000 queries. By 2024 the partnership reported more than 4,000 lawyers across 43 jurisdictions using Harvey, with an average 2 to 3 hours saved per week per lawyer, a 30 percent reduction in contract review time, and around 7 hours saved on complex document analysis (A&O Shearman). Harvey raised $160 million in late 2025 to expand into multi-step agentic workflows for antitrust filings, cybersecurity reviews, fund formation, and loan review.
Thomson Reuters acquired Casetext, the maker of CoCounsel, for $650 million in June 2023, and the tool became embedded in many large U.S. firms. Hebbia, used by 30 percent of the top 50 asset managers by assets under management as of mid 2024, took a different angle, focusing on multi-document search and structured extraction across PDFs, filings, and emails (Contrary Research).
Within the CFO's organization the visible wins so far have been concentrated in accounts payable, contract management, and analyst workflows. Vic.ai trained on more than a billion invoices to automate accounts payable; its customers include KPMG, PwC, BDO, Armanino, HSB, and Intercom (Vic.ai). Klarity automates contract and disclosure review.
In investment banking, Morgan Stanley deployed an internal assistant called AskResearchGPT across its institutional securities group in 2024; the firm reported that close to half of its 80,000 employees were using OpenAI-powered tools by year end (Morgan Stanley). Bain & Company gave all 13,000 consultants access to ChatGPT Enterprise in August 2024 and assigned around 50 staff to build industry specific tools with OpenAI. Goldman Sachs CEO David Solomon said in January 2025 that AI could draft 95 percent of an IPO prospectus, with humans handling judgment-intensive sections; the firm later partnered with Anthropic to apply Claude to accounting and compliance tasks (Fortune).
HR is the most contested function. Workday's Illuminate, launched at Workday Rising in September 2024, draws on the 800 billion business transactions Workday processes every year. Workday simultaneously announced agents for recruiting, expenses, succession planning, and workforce optimization, plus the acquisition of contract intelligence vendor Evisort (Workday). Vendors such as Eightfold, HireVue, and Paradox apply AI to sourcing, screening, and interviewing. The legal and reputational risks have been substantial. In January 2025 a class action alleged that Eightfold scraped data on more than a billion workers and ranked candidates on a zero to five scale without complying with the Fair Credit Reporting Act. The ACLU of Colorado filed a discrimination complaint against HireVue's video assessment platform in March 2025 (Fisher Phillips).
Blue Yonder (acquired by Panasonic) and C3.ai dominate the analytics side of enterprise supply chain AI. Blue Yonder reports a 12 percent improvement in forecast accuracy from its machine learning demand planning system and added generative AI orchestration with the launch of Blue Yonder Orchestrator. C3.ai offers a suite of agents for sourcing, granular forecasting, production scheduling, and inventory optimization. SAP and Oracle have layered their own AI agents into core ERP. Logistics companies including Maersk and DHL run their own optimization stacks.
Glean, founded in 2019, became the most prominent independent vendor for AI-powered enterprise search and assistant capabilities. Glean's hooks into Slack, Google Drive, Notion, Salesforce, GitHub, Jira, and over a hundred other systems made it a default choice for companies that wanted a single answer surface across all internal knowledge. The company crossed $100 million ARR in early 2025 and doubled to $200 million by December 2025, at a $7.2 billion valuation in June 2025 (Sacra). Notion AI and Slack AI offer overlapping features inside their respective products, and Microsoft Copilot wraps the M365 graph in a similar way.
The most talked-about shift in 2024 and 2025 was the move from copilots to agents, software that does not just suggest the next line but takes multi-step actions across tools. Salesforce's launch of Agentforce in October 2024 set the marketing tone. By Dreamforce 2024 customers had built more than 10,000 autonomous agents on the platform. The platform underwent rapid iteration: Agentforce 2 in December 2024 added the Atlas Reasoning Engine, Agentforce 2dx in March 2025 embedded agents in triggered cross-functional workflows, and Agentforce 3 in June 2025 added observability and governance for scale (Salesforce).
A second cluster of startups attacked particular high value workflows. Sierra focused on customer service. Cognition's Devin tackled software engineering, reporting a SWE-bench Verified score of 13.86 percent on launch, far above the 1.96 percent prior state of the art (Cognition). Decagon and Cresta worked the contact center. Imbue, Adept (acquired by Amazon in 2024), and Lindy chased general personal and business agents. Crew AI and LangChain's LangGraph provided open-source frameworks for orchestrating multi-agent systems.
Research from Anthropic's Economic Index, which analyzes how users interact with Claude across 1 million conversations on Claude.ai and 1 million API transcripts, shows directive task delegation rising from 27 percent to 39 percent over an eight month period, with roughly 49 percent of jobs seeing at least a quarter of their tasks performed with Claude (Anthropic).
The gap between launch demos and production reality remains wide. Gartner forecasts that 40 percent of agentic AI projects will be canceled by the end of 2027, mostly because companies underestimated the engineering cost of moving from prototype to scaled deployment.
The most reliable evidence on workplace effect comes from controlled experiments and large-scale field deployments, not from vendor claims. The pattern across studies is consistent: generative AI helps less on the hardest creative tasks and more on routine, language-heavy work, with the largest gains accruing to less experienced workers.
| Study | Year | Population | Task | Headline effect |
|---|---|---|---|---|
| Brynjolfsson, Li, Raymond (NBER) | 2023 | 5,179 customer support agents | Live chat at a Fortune 500 software firm | 14% more issues resolved per hour on average; 34% gain for novice agents; near zero for top performers |
| Noy and Zhang (Science) | 2023 | 453 college-educated professionals | Writing tasks (memos, plans, emails) | Time down 40%, output quality up 18%; ChatGPT compressed productivity distribution |
| Dell'Acqua et al. (Harvard / BCG) | 2023 | 758 BCG consultants | 18 realistic management tasks | Within AI's capability: 12.2% more tasks, 25.1% faster, 40% higher quality. Outside: accuracy dropped from 84% to 60-70% |
| Peng, Kannan et al. (GitHub Next) | 2022 | 95 professional developers | Build an HTTP server in JavaScript | 55% faster with Copilot (71 vs 161 min); statistically significant |
| Microsoft Work Trend Index (Microsoft) | 2023-2024 | Early Copilot customers | M365 office work | 29% faster on search, writing, and summarization; 70% report productivity gain |
| Forrester TEI (Forrester for Microsoft) | 2024 | Composite 25,000-employee org | M365 Copilot rollout | Projected 112-457% three-year ROI |
| PwC Global AI Jobs Barometer (PwC) | 2024 | Half a billion job ads in 15 countries | Labor market exposure | AI-exposed sectors saw 4.8x faster productivity growth; AI skills earned a 25% wage premium |
| PwC Global AI Jobs Barometer (PwC 2025) | 2025 | Updated dataset | Labor market exposure | Wage premium rose to 56%; revenue per employee grew 27% in most exposed industries vs 9% in least exposed |
| Uplevel study, reported by Visual Studio Magazine | 2024 | 800 developers | Pull request throughput | No statistically significant gain from Copilot use |
A few patterns repeat. First, less experienced workers benefit most; the Brynjolfsson study found novice agents performed after two months as well as untreated agents with six months of tenure. Second, the gains are larger when AI is asked to do something close to what it can already do well (drafting, summarizing, generating boilerplate code) and shrink or invert when the task is genuinely hard. Dell'Acqua's experiment found that consultants given GPT-4 on a deliberately difficult task did worse than consultants without AI: 60 to 70 percent accuracy with AI compared to 84 percent without. This is the "jagged frontier" the authors describe: the boundary between what AI does well and what it does badly is not always visible until you cross it.
Third, vendor demos overstate scale-up gains. Most field studies that look beyond individual task completion (such as the 2024 Uplevel work on developer throughput) struggle to find significant effects at the team level. The most plausible reading is that AI helps some tasks a lot, those tasks are a fraction of total work, and the remaining bottlenecks (meetings, decisions, integration) absorb the gains.
McKinsey estimates that generative AI could add $2.6 to $4.4 trillion in annual economic value across 63 use cases, with about 75 percent of that concentrated in customer operations, marketing and sales, software engineering, and R&D (McKinsey). These projections sit on top of base-rate AI productivity gains of a few percentage points of GDP per year over a decade or two.
Microeconomic returns inside individual companies are easier to verify in pockets. Microsoft's 2026 Q3 figures show $37 billion in annual AI revenue run rate, up 123 percent year over year, with Azure growing 40 percent. ServiceNow grew Now Assist deals more than 150 percent quarter over quarter in Q4 2024 and saw "Pro Plus" deals (which include Now Assist) more than quadruple year over year. Customers reportedly raised their ServiceNow spend by up to 60 percent after adding AI tier features (io-fund).
Vendor-funded ROI studies are routinely picked apart by procurement teams. CNBC's October 2024 Technology Executive Council survey found CIOs evenly split on whether Copilot was worth $30 per user per month, with the largest group, 50 percent, saying it was too soon to know (CNBC TEC). UBS bought 50,000 Copilot licenses; Accenture committed to 200,000. By mid 2025 OpenAI reported more than three million paying business users for ChatGPT Enterprise, Team, and EDU, and Anthropic's annualized revenue had crossed $4 billion in part on the strength of Claude Code adoption.
The risks fall into three rough categories: factual errors and hallucination, data leakage and security, and bias or discrimination.
The Air Canada chatbot case in 2024 became the standard reference for liability. A British Columbia tribunal ruled that the airline owed Jake Moffatt a partial refund after its chatbot invented a bereavement fare policy that did not exist. Air Canada argued that the chatbot was a separate legal entity responsible for its own actions; the tribunal rejected the argument. The case established that enterprises are accountable for what their AI systems say, which has shaped governance and indemnification clauses ever since (Computerworld). Gartner analyst Avivah Litan estimated roughly 30 percent of generative AI outputs as hallucinations in early deployments.
Data leakage was the other early scare. In March and April 2023 Samsung Semiconductor engineers pasted source code and internal meeting transcripts into ChatGPT on at least three separate occasions while debugging chip equipment and summarizing meetings. Samsung banned generative AI on company devices in May 2023. Several large banks including JPMorgan, Bank of America, Citi, Wells Fargo, Goldman Sachs, and Deutsche Bank imposed similar restrictions before later approving enterprise tier deployments with stronger data controls (TechCrunch).
Discrimination claims arrived most heavily in HR. The Eightfold class action filed in January 2025 alleged the company built "hidden credit reports" on more than a billion workers without complying with the Fair Credit Reporting Act. CVS settled a class action in mid 2024 over HireVue's video interview product, which used facial expression analysis. These cases have pushed many companies to require human review of any AI-driven hiring or firing decision.
The governance response has been a wave of internal policies, role creation, and tooling. McKinsey's 2024 survey found only 18 percent of organizations had an enterprise-wide council or board with authority over responsible AI decisions, and only a third required AI risk awareness as a technical skill set. Companies including Boeing, Booz Allen, Mastercard, and Dell appointed Chief AI Officers. KPMG reported that nearly half of organizations were planning to add a CAIO within a year, with packages at large firms reaching $350,000 to $650,000 or more (KPMG).
The two reference frameworks for enterprise AI governance are the EU AI Act and the NIST AI Risk Management Framework.
The EU AI Act entered into force on August 1, 2024, with phased application. Prohibitions and AI literacy obligations applied from February 2, 2025; rules for general-purpose AI models took effect on August 2, 2025; obligations for Annex III high-risk systems become enforceable on August 2, 2026. High-risk categories explicitly include AI used in recruitment, credit scoring, education access, law enforcement, critical infrastructure, and medical devices. Providers of high-risk systems must register them on a public database, complete conformity assessments, document data governance and risk management, provide instructions to deployers, appoint an authorized EU representative if based outside the EU, and report serious incidents. Penalties for the worst violations reach 7 percent of global turnover. A Cloud Security Alliance survey found that more than half of organizations had not yet built systematic inventories of their AI systems, a prerequisite for any conformity assessment (CSA).
The NIST AI Risk Management Framework, released in January 2023, is voluntary in the United States but has become the default reference for internal governance globally. It defines four functions: GOVERN, MAP, MEASURE, and MANAGE. GOVERN sits at the top, MAP identifies risks in context, MEASURE assesses and tracks them, and MANAGE prioritizes treatments. NIST recommends folding AI risk into broader enterprise risk management alongside cybersecurity and privacy (NIST).
In the United States, state and federal action has been patchier. California passed several disclosure laws on AI training data and deepfakes. New York City's Local Law 144 requires bias audits on automated employment decision tools. The federal Executive Order on AI from October 2023 was rescinded by the Trump administration in January 2025, replaced with a deregulatory order in April 2025. Most large U.S. companies kept aligning to NIST regardless.
The relationship between enterprise AI and jobs has been the most politically charged part of this story, and the evidence so far is mixed. PwC's 2025 Global AI Jobs Barometer, which analyzed roughly half a billion job postings across 15 countries, found a 38 percent rise in postings for roles most exposed to AI, an average 56 percent wage premium for AI skills, and revenue per employee growing 27 percent in AI-exposed industries compared to 9 percent in the least exposed (PwC). Skill requirements were changing 66 percent faster in exposed occupations than in the rest of the economy.
Layoff narratives have been louder than the data sometimes supports. IBM CEO Arvind Krishna said in 2023 that he expected to pause hiring for roles that AI could do, mentioning back-office HR functions. Klarna's reversal in 2025 punctured part of the narrative that AI would eliminate large numbers of customer service jobs quickly. The flip side is the kind of internal redesign Goldman Sachs and Morgan Stanley have pursued, where AI absorbs portions of work rather than full roles.
The Brynjolfsson, Li, and Raymond study suggests one of the more durable predictions: AI tools tend to compress the productivity distribution, lifting beginners more than experts. That logic, applied at the firm level, means companies that adopt AI well may need fewer experienced workers per unit of output but a similar or larger number of less experienced ones, who are now more productive than before. Whether this story holds across knowledge work generally is one of the open questions of the next several years.