How large organisations are buying, building, and deploying AI: Fortune 500 case studies, vendor partnerships, platform launches, and the workflow changes shipping inside companies right now.
Topic brief
Brief updated Aug 1, 2026
Enterprise AI is the adoption of AI models, agents and automation inside businesses, spanning productivity copilots, customer-experience tools, back-office automation and custom agentic applications. It is where frontier-model capability meets the constraints of real organizations: legacy systems, governed data, identity and permissions, security review, procurement and measurable return on investment. For practitioners, the defining challenge is not model quality alone but turning capability into governed, reliable production workflows.
The stakeholders span the whole organization. Engineering and platform teams integrate models and manage governance, security and cost. Line-of-business leaders chase productivity and customer-experience gains. Executives weigh build versus buy and vendor risk. A dense vendor ecosystem of frontier labs, cloud providers, systems integrators, agent-platform startups and governance-tooling makers competes to supply the stack.
The topic matters because enterprise AI is where the technology's economic impact is realized or stalls, and the evidence about that arrives in two incompatible forms. Aggregate surveys repeatedly report weak or unmeasurable return and low organizational readiness. Firm-level accounts report large gains. Both can be true at once, and reconciling them is the analytical task this hub exists to support. Underneath sit the recurring structural themes: multi-vendor sprawl, identity-aware access, security drift in connectors that change after approval, durable agent state and memory, token cost and pricing models, and the organizational change management that decides whether a pilot becomes a durable system.
Agent governance stopped being an argument and became a product line, with a single week producing a control plane, a permissions model, an acquisition and a funding round aimed at the same problem. Snowflake introduced Cortex AI Gateway on July 28 as a centralized layer for controlling how first- and third-party agents reach models, tools, MCP servers, data and enterprise systems, with policy, activity and cost controls and support for more than 100 MCP servers; it is planned for public preview soon and several identity integrations remain planned for private preview, so production enforcement is not yet demonstrated. The same day, 1Password launched Privileged Access, applying just-in-time, task-scoped permissions to people, service accounts and AI agents by creating permissions in the target system on request and removing them when the session ends; Cyera announced a $1 billion agreement to acquire non-human identity company Oasis Security, currently documented as a letter of intent still subject to a binding agreement; and Hush Security raised a $30 million Series A for centralized agent registration, just-in-time permissions and revocation. MinIO's AIStor Memory, announced July 29, addresses the adjacent gap by separating durable organizational knowledge, active workspace state and credential custody into distinct services, though the announcement carries no pricing, retrieval benchmarks or independent production results. Microsoft's contribution is narrower and slower: a Purview DLP rule in preview that excludes externally received email from Microsoft 365 Copilot grounding by checking sender-domain metadata, with general availability listed for January 2027. Read together, these are answers to the same question, which is what an agent is allowed to touch and who can prove it afterwards, and the honest status on nearly all of them is preview rather than production.
The economics moved in the opposite direction from the governance spend, which is the tension worth planning around. OpenAI said its models reach more than 1 billion active users and more than 2 million businesses, company-reported figures with no stated measurement window, and cut GPT-5.6 Luna to $0.20 per million input tokens and $1.20 output; Amazon Bedrock made the same Sol, Terra and Luna family generally available through the OpenAI Responses API with a 272K-token context window and prompt caching whose reported 90% cached-input discount matters most for exactly the long-context agent traces this hub cares about. Moonshot AI published the weights of its 2.8-trillion-parameter Kimi K3 on July 27, and ABC documented individual Australian companies routing simpler work to cheaper models. But provenance now has a procurement cost: on July 31 US House committee chairmen requested documents from DoorDash about its evaluation and deployment of Chinese AI systems, following disclosures that it used Moonshot's Kimi K2.6 for lower-level work, a request that alleges no wrongdoing but extends congressional scrutiny from model developers to the enterprises that adopt them. Cheaper tokens are also not translating into controlled spend, with Harness estimating 26% of enterprise AI spending wasted and 52% of its 700 respondents lacking a clear AI-cost owner. The consolidation around all this continued, with Nscale signing a definitive agreement for Anyscale expected to close in the second half of 2026 on undisclosed terms, Bloomberg agreeing to acquire Canoe Intelligence, and Sierra acquiring Takeoff to build a long-horizon agent platform called Horizon. The starkest signal was monday.com's July 22 decision to cut about 20% of its workforce, roughly 620 people, while reorganizing around its AI Work Platform and raising its non-GAAP operating-margin outlook to about 15%, taking $45 million to $55 million in net restructuring charges for a bet whose product and customer effects are still unproven.
Most of this month's governance news carries an explicit not-yet date, so the checkable signals are availability milestones: whether Snowflake's Cortex AI Gateway reaches the public preview it says is coming soon and whether its planned private-preview identity integrations land, and whether Microsoft's Purview DLP control for Copilot grounding hits the January 2027 general-availability rollout it has listed. On regulation, BaFin's initial monitoring covers transparency duties and prohibited practices, with oversight of high-risk systems including certain creditworthiness uses scheduled to begin in December 2027, and the House committee request to DoorDash is an open document demand rather than a finding. Several deals are agreements rather than closings: Nscale's acquisition of Anyscale is expected to close in the second half of 2026 subject to regulatory approvals with terms undisclosed, so the $1.65 billion figure in secondary coverage remains unconfirmed; Cyera's $1 billion agreement for Oasis Security is documented as a letter of intent needing a binding agreement; Bloomberg's Canoe Intelligence deal has no disclosed terms or closing date; and Vusion's In-Store Media purchase is subject to regulatory approval. Product availability to track includes Amex GBT's Egencia connector, with access for Egencia customers planned for Q3 2026, and IHG's US conversational search beta expanding to additional markets after testing. Finally, watch whether the vendor-reported performance numbers get methodology attached, specifically Torq's move from 85% to a claimed more than 91% agreement with analyst-corrected verdicts, Brown and Brown's 2x to 8x pilot gains now to be tracked by a new Value Management Office, and monday.com's $45 million to $55 million restructuring charge against its raised roughly 15% non-GAAP operating-margin outlook.
Four moves in roughly a week converged on the same control problem. Snowflake introduced Cortex AI Gateway as a central policy, activity and cost layer for first- and third-party agents reaching models, tools, MCP servers and data, with support for more than 100 MCP servers and public preview planned. 1Password launched Privileged Access, provisioning just-in-time task-scoped permissions for people, service accounts and AI agents and removing them when the session ends. Cyera agreed to acquire non-human identity company Oasis Security for $1 billion, and Hush Security raised $30 million for agent registration, just-in-time permissions and revocation. All of these are announcements or previews; none come with independent production evidence.
Because the savings and the waste live at different layers. OpenAI cut GPT-5.6 Luna to $0.20 per million input tokens and $1.20 output, and Bedrock's prompt caching carries a reported 90% cached-input discount that matters for agents resending long context and tool traces. In the same week Harness's survey of 700 respondents across five countries estimated 26% of enterprise AI spending is wasted, with 52% lacking a clear AI-cost owner, 72% hit by a surprise bill in the past year and only 20% able to diagnose a doubled bill within hours. Those are self-reported figures from a vendor-sponsored survey, but they point at attribution rather than unit price.
It is now a documentation requirement at minimum. Moonshot AI published weights for the 2.8-trillion-parameter Kimi K3 on July 27, and reporting describes Australian companies routing simpler work to cheaper models. On July 31 US House committee chairmen requested documents from DoorDash about its evaluation and deployment of Chinese AI systems, citing its use of Moonshot's Kimi K2.6 for lower-level work as part of a joint congressional investigation. The request alleges no breach or wrongdoing, but it makes model-provenance, data-flow and supply-chain records a practical deliverable rather than an internal note.
MinIO's framing is at least a concrete architecture to evaluate against. AIStor Memory separates durable organizational knowledge, active workspace state and credential custody into distinct services with different access boundaries and lifecycles, and MinIO says its Agent Biography can preserve evidence, decisions, corrections, provenance, outcomes and unfinished work so authorized agents do not rebuild context from full transcripts. The company itself cautions that the product does not change the underlying model or guarantee perfect recall, and the announcement includes no pricing, retrieval benchmarks, latency data or independent production results.
There is one large restructuring in this evidence and a lot of softer signal around it. monday.com announced on July 22 that it will cut about 20% of its global workforce, roughly 620 people, while reorganizing around its AI Work Platform, expecting $45 million to $55 million in net restructuring charges, continuing to hire in selected strategic areas and raising its non-GAAP operating-margin outlook to about 15%. The company said the reduction was not designed to replace people with AI or primarily improve margins. Bessemer's survey of 173 leaders at 113 portfolio companies found 49% delivering more without adding headcount and 13% slowing or pausing hiring, which is a portfolio sample rather than a labor-market measurement.
Separate the ones with a published evaluation from the ones without. Brown and Brown reported pilot gains of roughly 2x to 8x and 80% to 90% faster troubleshooting in some use cases without disclosing sample size or methodology, and Torq reported 85% agreement with analyst-corrected verdicts before its CTO said recent deployments exceeded 91%, again without published methodology, sample sizes or false-positive and false-negative rates. Contrast that with the Myriad Genetics pipeline reported with AWS, which states a 1,200-document evaluation set, accuracy moving from 94% to 98%, per-page cost falling 77% from 3.1 to 0.7 cents and time falling from 8.5 to 1.5 minutes per document.
Latest coverage
Showing newest first with full-topic pagination
Uber CTO Praveen Neppalli Naga said the company paired about 30 AI-proficient engineers with domain experts and ran 16 two-week “Agentic Pods” across 16 business functions. In a July 7 post, he reported cutting several workflows from hours or days to minutes and said Uber is forming a dedicated team to expand the model; the figures are company-reported and not independently audited.
Okta's Enterprise AI Index, published July 16, found that most customers in a sample of more than 20,000 organizations used multiple AI platforms by June 2026. The study tracked sanctioned applications accessed through Okta single sign-on, so it measures deployment within Okta's customer base rather than the entire market, and it points to growing governance demands as companies combine AI tools.
OpenAI said on July 31 that its models reach more than 1 billion active users and more than 2 million businesses. The company also detailed price cuts announced a day earlier: GPT-5.6 Luna now costs $0.20 per million input tokens and $1.20 per million output tokens, while Terra costs $2 and $12. The usage figures are company-reported and not independently audited.
Virgin Atlantic's VP of digital engineering described at an April 2026 Adobe Summit session how its AI Concierge uses seven contextual traveler preferences, including family size and weather tolerance, to guide its holidays agent, according to a July 31 CMSWire account. The airline launched the concierge in December 2025, so the new development is the disclosed design approach, not a fresh product launch.
Amazon Quick has introduced the Agentic Catalog Experience in preview, an AI-powered workflow that lets data curators discover catalog assets through natural-language prompts and create dataset representations in bulk. AWS documentation states that the workflow uses upstream metadata, including descriptions, data-quality scores, and lineage, while retaining the connected catalog as the source of truth.
Plutus announced on July 30 a set of six AI-assisted workflow tools for life and annuity insurers, deployed on ServiceNow Financial Services Operations and grounded in the company's Atlas semantic layer. The tools cover payout transactions, not-in-good-order review, quality checks, and orphaned-policy remediation, but the launch materials do not provide customer results, benchmark accuracy, or measured automation rates.
ABC reported on August 1 Australian time that cheaper Chinese-developed models are changing how some businesses allocate AI workloads. Moonshot AI released its 2.8-trillion-parameter Kimi K3 on July 16 and published the weights on July 27; Moonshot says the model trails the leading proprietary systems overall, while independent reporting describes growing interest in lower-cost, self-hosted alternatives.
India's Ministry of Electronics and Information Technology said on July 13 that six of seven cyber threats forecast for the banking and payments sector a year earlier had reached full-scale realization. Economic Times BFSI reported that banks are increasing defenses as attacks combine AI, stolen identities, and manipulated payment workflows, while Punjab National Bank allocated roughly Rs 7-8 billion to cybersecurity for FY27.
Onapsis published The State of AI, Security and ERP on July 30, 2026, reporting that 86% of 204 surveyed U.S. cybersecurity leaders said their organizations had integrated or would soon integrate AI into ERP code. Only 30% were fully confident they could detect an AI-based attack, underscoring a readiness gap within this specific large-enterprise sample.
U.S. House committee chairmen requested documents from DoorDash on July 31 regarding its evaluation and deployment of Chinese AI systems, CNBC reported. The inquiry follows public disclosures that DoorDash has used Moonshot AI's Kimi K2.6 for lower-level AI work. The request is part of a joint congressional investigation into security implications of U.S. companies using China-developed models.
Walmart's July 16, 2026 Jobs Spotlight report said the retailer is helping its 2.1 million associates build skills and confidence as AI changes work. The report follows April remarks from chief people officer Donna Morris that Walmart aims over several years to give every associate some level of AI skills; role-specific certifications were then available to 1.7 million workers in the United States and Canada.
AttackIQ announced AVA Agentic OS on July 30, introducing an agentic operating system for operationalizing Continuous Threat Exposure Management (CTEM). According to the company's press release, AVA coordinates specialized AI agents and tools across workflows including threat intelligence, security validation, detection engineering, control optimization, and AI security validation.
Zocks launched Zocks Scheduling on July 30, adding automated meeting booking from client conversations and emails to its financial-services AI platform. According to the company's announcement, the feature links bookings with meeting preparation, notes, and follow-up actions, using information from prior conversations and connected systems such as CRM and planning tools.
American Express Global Business Travel announced the Egencia AI connector for Anthropic's Claude on July 27, 2026, enabling policy-controlled air and hotel search, booking, and management from the assistant. The company describes it as the first deployment of its agent-to-agent architecture, connected through Model Context Protocol. Business Travel News reported the launch independently, while ET TravelWorld said access for Egencia customers is planned for Q3 2026.
Encore AI announced a $30 million Series A on July 29 to expand its enterprise customer-interaction platform. The company says Team8, Planven, and The Garage led the round, with Lukatz and several bank and insurer customers also investing. Encore analyzes calls, messages, and CRM records to turn behaviors associated with successful outcomes into playbooks for voice and text agents.
Google began a gradual rollout of Gemini visual tools in Google Docs on July 28, enabling eligible users to create and edit images, diagrams, and infographics within a document. The web-only features can use document context and apply one natural-language request to multiple visuals. Access is limited to specified Workspace, education, and Google AI subscription tiers.
Harness released its 2026 State of AI in FinOps survey on July 29, estimating that enterprises waste 26% of AI spending. The 700-person survey across five countries found that 52% of respondents lacked a clear AI-cost owner, 72% had encountered a surprise AI bill in the past year, and only 20% could diagnose a doubled bill within hours.
Nscale said on July 30 that it signed a definitive agreement to acquire Anyscale, the company behind the Ray distributed-computing framework. The deal is expected to close in the second half of 2026, subject to closing conditions and regulatory approvals. Nscale did not disclose financial terms, so a $1.65 billion figure cited in secondary coverage is not confirmed by the companies.
DataBahn announced a $40 million Series B on July 30 led by Insight Partners, with Forgepoint, GTM Capital and S3 Ventures participating. The company said the round brings total funding to $59 million and will expand its agentic data control plane; SiliconANGLE reports that the platform ingests, normalizes and routes enterprise telemetry from more than 600 sources.
MinIO introduced AIStor Memory on July 29, a customer-controlled persistence layer for AI agents that keeps durable knowledge, active workspace state, and credentials as distinct services on the AIStor platform. The company says it can preserve evidence, decisions, corrections, provenance, and unfinished work across agent runs, while cautioning that the product does not change the underlying model or guarantee perfect recall.
Bloomberg agreed on July 29 to acquire Canoe Intelligence, an AI-assisted platform that collects and structures private-markets data. Canoe says its system processes more than 1.5 million documents a month across over 44,000 funds for more than 500 institutional clients representing $11 trillion in assets under service. The public announcement did not disclose financial terms.
Freight Hero said on July 27 that it raised a $5 million seed round led by Field Ventures, with Flybridge Capital, Tip Top VC and Front Porch Venture Partners participating. The company plans to expand sales, engineering and operations while extending its AI-and-human back-office service into billing, accounting and carrier sales; it says total funding now exceeds $6 million.
Germany's BaFin said on July 29 that it had begun monitoring AI use by banks, insurers and other financial firms after new national legislation assigned it responsibilities under the EU AI Act. The regulator's initial work covers transparency duties and prohibited practices, while oversight of high-risk systems—including certain creditworthiness uses—is scheduled to begin in December 2027.
Torq introduced SOC Brain on July 28, a private, per-customer learning layer for its AI SOC Platform that combines historical-case retrieval, organization-specific model training, and imports of resolved incidents. Torq initially reported 85% agreement with analyst-corrected verdicts; CTO Leonid Belkind later told MSSP Alert that recent deployments exceeded 91%, without publishing methodology, sample sizes, or false-positive and false-negative rates. Low-confidence cases are escalated to analysts.
Microsoft made a Purview Data Loss Prevention control available in preview that can stop Microsoft 365 Copilot and Copilot Chat from using email received from external domains. The policy checks sender-domain metadata and excludes matching email from grounding, summarization, and citation; it does not inspect the message body. Microsoft lists the general-availability rollout for January 2027.
Hong Kong's Housing Bureau, Housing Authority, Cyberport, and University of Hong Kong launched the Smart QS Hackathon 2026 on July 29. The program focuses on AI prototypes for contract advisory and payment valuation, with applications due August 14. Organizers say selected solutions may be tested at public-housing construction sites, but no production deployment has been confirmed.
Munich-based 5U AI said on July 28 that it raised $3.2 million in pre-seed funding led by Emerge Capital. The startup plans to expand AI workers that handle quoting, bookings, tracking and invoice reconciliation inside freight forwarders' existing systems, while recording the reasoning behind operational decisions.
IncQuery's survey of consulting and investment professionals found 70% expect their primary research needs to increase over the next year, at exactly the moment AI can answer almost anything instantly. In an email interview with Lets Data Science, Nishank Singhal, who led the research, set out the AI failure modes that catch experienced professionals off guard, including synthetic respondents, false consensus that erases outliers, and output that sounds equally confident whether it is right or inventing.
Siemens reported on June 24 that Jack Technology ordered 2,000 humanoid robots tailored for apparel manufacturing, describing it as the sector's first large-scale deployment. The supplier, factory tasks and rollout timetable were not disclosed; a related Siemens collaboration will add low-code development, digital engineering and industrial AI software while Jack targets up to 30% higher manufacturing efficiency.
Indeed research published July 28 found that 43% of surveyed managers felt poorly equipped or not equipped to lead AI-native workers, while 45% felt ready. The U.S. survey of 1,301 people also found 52% of workers lacked the AI training they needed, making management development and role-specific upskilling a practical constraint on AI adoption.
