Where does enterprise AI agent adoption really stand?
Permalink to “Where does enterprise AI agent adoption really stand?”The headline numbers suggest broad momentum, but not scale. According to McKinsey’s State of AI 2025 survey, 88% of organizations now use AI in at least one business function, and 62% are experimenting with AI agents specifically. Yet only 23% have scaled an agentic system beyond a pilot, and fewer than 10% report agents running at scale in any single function.
PwC’s 2025 AI Agent Survey tells a similar story: 79% of senior executives say AI agents are already being adopted, but only 66% of those adopters see measurable productivity gains. The gap between “adopting” and “scaling” is the constant across every survey.
The pattern that emerges: nearly everyone is trying agents, a meaningful share is seeing early returns, but the jump from pilot to production is where programs stall. The reason is consistent across industries: agents that reach production are grounded in a governed business context, metric definitions, policy rules, lineage, and classification schemas the agent can trust. That is the gap Atlan’s Context Layer for AI closes: a governed layer that keeps those definitions, policies, and lineage current across every agent that draws on them.
What are the real AI agent use cases inside enterprises?
Permalink to “What are the real AI agent use cases inside enterprises?”Four functions account for most of the adoption enterprises report today: conversational analytics, customer service, data management, and IT and security operations. Each cluster has a different maturity level and a different set of requirements for reaching production.
Conversational analytics (talk to your data)
Permalink to “Conversational analytics (talk to your data)”This is the use case that gets executives excited first. Leaders ask a natural-language question, and an AI agent connected to their data platform returns an answer grounded in real metrics.
The appeal is obvious: no SQL, no waiting for an analyst. The reality is harder. For a conversational analytics agent to answer correctly, it needs to know what “revenue” means in this specific company, which tables to join, what “this quarter” refers to, and which definition of “active customer” is canonical.
One global online-learning company deployed a conversational analytics agent connected over MCP. Their CEO became the heaviest user, and that executive adoption pulled the rest of the company in. The agent worked because it was grounded in the company’s actual metric definitions, not generic knowledge.
The enterprises getting value from talk-to-data agents are the ones that invested in making their metric definitions, joins, and business logic machine-readable before the agent went live.
Customer service and contact centers
Permalink to “Customer service and contact centers”Customer service is the most widely deployed use case for customer-facing agents. According to Cloudera’s 2025 survey, 78% use AI agents for customer support, and 96% plan to expand agent use in the next 12 months.
The agent-native market has consolidated around three vendors: Sierra, Cresta, and Decagon. According to TechCrunch, Sierra reported $150 million in annual recurring revenue in February 2026. Axios reported that Cresta crossed $100 million in ARR that April, and Decagon was valued at $4.5 billion the same quarter.
What makes a customer service agent work in production? It needs account context (purchase history, subscription tier, prior tickets) and policy rules (refund limits, escalation thresholds, data scoping requirements). Without that context, the agent either gives generic answers or makes decisions that violate company policy.
The AI agent architecture matters here. The best deployments connect to live systems (CRM, ticketing, knowledge bases) rather than working from static exports, reasoning across those systems with the business context that tells the agent what the rules are.
Data management and enrichment
Permalink to “Data management and enrichment”This is the use case that data teams have been waiting for. AI agents can now automate metadata tagging, data lineage mapping, classification, and bulk enrichment at a scale manual processes could never match.
A global data-storage manufacturer used AI agents for bulk metadata enrichment. Their data governance team delivered an outcome that leadership had been asking for “forever,” tagging assets faster and more consistently than manual workflows ever could.
Data management agents don’t need to answer ad-hoc questions. They need to understand the structure of the data estate, apply classification rules consistently, and generate descriptions that are accurate enough for other agents (and humans) to rely on.
For enterprises exploring this path, AI agents in data management and AI agents for the data catalog cover the specific capabilities and patterns.
IT, security, and other operations
Permalink to “IT, security, and other operations”IT operations and security monitoring are natural fits for agentic AI. These environments generate high volumes of structured signals (alerts, logs, tickets) that follow patterns an agent can learn to triage and act on.
According to Anthropic and Material’s 2026 State of AI Agents report, 90% of organizations surveyed use AI to assist with coding today, and nearly 60% report that AI agents free up time across the development lifecycle.
Cloudera’s survey found that 63% of enterprises use or plan to use security monitoring agents, and 66% are investing in performance optimization bots. These are operational use cases where the agent’s context is primarily technical: system state, alert history, known playbooks, and escalation rules.
The AI agent use cases in data engineering page covers the technical patterns in more detail.
That technical specificity is why these agents often reach production faster than their business-facing counterparts: logs, alerts, and runbooks are already structured, machine-readable context, which means less translation work between what the agent needs and what already exists. The failure mode shifts from missing context to stale context, and an outdated runbook is just as dangerous as no runbook at all.
The AI Context Stack
The reference architecture for giving agents governed access to metrics, policies, and lineage.
Get the AI Context StackHow does adoption vary by industry?
Permalink to “How does adoption vary by industry?”The industry gap comes down to which sectors have the operational data and regulatory pressure to make agents trustworthy first.
According to NVIDIA’s State of AI Report 2026 (3,200+ responses across five sectors), telecommunications leads all industries in agentic AI adoption at 48%, with retail and CPG close behind at 47%.
Financial services trail in raw adoption but lead on maturity. NVIDIA puts adoption at 42%, with 21% already in production. Evident Insights found generative and agentic use cases now make up 70% of bank and insurer implementations, concentrated in compliance, reconciliation, and fraud detection.
Healthcare sits at 47% adoption per NVIDIA, though agentic AI still ranks fourth among AI workloads there; clinical documentation is further along than autonomous clinical decision-making.
Retail and customer experience match healthcare in adoption at 47%, per NVIDIA. Retailers rank process speed and efficiency (57%) and personalization (40%) as their top goals.
Manufacturing is the earliest in the curve. Deloitte frames agentic AI there as emerging, concentrated in supply chain, aftermarket service, and workforce knowledge capture. The clearest proof point is PepsiCo’s digital-twin work with Siemens and NVIDIA, which caught up to 90% of issues before they reached the physical line and lifted throughput 20%, per NVIDIA’s 2026 report.
For industry-specific agent patterns, see AI agents in finance, AI agents in healthcare, AI agents in retail, and AI agents in manufacturing.
Raw adoption counts how many pilots got started. Production counts how many industries had already codified the business rules, compliance requirements, and definitions agents need, which is why financial services and manufacturing convert pilots into production fastest even as they trail on headline adoption numbers.
Why do most agent projects stall before reaching scale?
Permalink to “Why do most agent projects stall before reaching scale?”The honest answer is not that the models are too slow or the use cases too ambitious; those were real constraints two years ago, not today.
According to McKinsey’s 2025 survey, only 39% of organizations say AI has made any real dent in their bottom line. Gartner projects that more than 40% of agentic AI projects will be canceled by 2027 due to unclear ROI, escalating costs, and inadequate controls.
The blocker that shows up across every survey, every industry, and every function is context readiness.
Anthropic’s research on building effective agents found that the most successful implementations use simple, composable patterns, not complex frameworks. The difference between a demo and a deployment is whether the agent has access to the right information, in the right format, with the right governance.
Here is what that looks like in practice:
-
A conversational analytics agent that gets “revenue” wrong because it pulls the definition from an outdated source instead of the finance team’s canonical version
-
A customer service agent that approves a refund it should not because the policy rules were not loaded into its context
-
A data enrichment agent that tags assets with outdated classifications because the glossary was last updated 18 months ago
-
An IT operations agent that escalates a known-issue alert because it does not have access to the runbook
In each case, the model is capable, the architecture is sound, and the context is missing, stale, or inconsistent.
This is the AI agent cold-start problem at enterprise scale. The agent does not fail because it cannot reason. It fails because it doesn’t understand the business context behind the task it’s working on.
What do the successful use cases have in common?
Permalink to “What do the successful use cases have in common?”Every use case that reaches production and stays there shares one trait: the agent runs on trustworthy, governed business context. Three things separate the ones that stick from the ones that stall:
-
Canonical definitions. One authoritative version of each business term, owned by one team, with a clear last-reviewed date. When the agent queries “customer,” it gets one answer, not three conflicting ones.
-
Governed context delivery. The agent retrieves context through structured protocols, like MCP, instead of scraping wikis or reading stale exports.
-
Continuous context maintenance. The context layer updates as the business changes, so definitions, policies, and lineage stay current instead of going stale.
This is what a context layer provides: the governed, versioned, machine-readable foundation every production agent draws from, not a prompt template or a RAG pipeline.
Context-Ready or Not?
Check where your agent program has context gaps before it reaches production.
Take the Readiness CheckHow does Atlan make agents production-ready?
Permalink to “How does Atlan make agents production-ready?”The three requirements above, canonical definitions, governed delivery, and continuous maintenance, describe an infrastructure layer, not a single feature. That is what Atlan’s Context Layer for AI is built to close. Mapped to the use cases above, the pattern is direct:
| Use case | What the agent needs | How Atlan makes it possible |
|---|---|---|
| Conversational analytics | Correct metric definitions, joins, and what “this quarter” means here | Context Agents generate the definitions; MCP serves them to the agent at query time |
| Data management and enrichment | Lineage, usage, and descriptions generated at scale | Context Agents run bulk enrichment continuously; 87% of customers say the output reads better than what humans wrote |
| Customer service and contact centers | Account context plus policy rules (refund limits, data scoping) | Atlan carries a governed policy context alongside the agent’s runtime context |
| Cross-functional agents at scale | One shared, consistent source of truth across every agent | Versioned, model-agnostic Context Repos give every agent the same context, regardless of framework or vendor |
| Any production agent | Proof it answers real questions before it goes live | Context Engineering Studio runs the agent’s answers against real historical questions and simulated scenarios before production |
For enterprises evaluating where to start, the context layer page explains the core concept, and why MCP matters for AI agents covers the protocol that makes context delivery work.
What are enterprises achieving with AI agents?
Permalink to “What are enterprises achieving with AI agents?”The results are real, even if they are concentrated.
“In just one year, we have onboarded 6,000 people. Conversational analytics will go even further to reach every employee.”
An EVP of Data Democratization at a global telecom operator
“We just crossed the threshold where it’s quicker to use Atlan AI than to go out of the tool.”
A data governance lead at a mortgage lender
According to the Anthropic and Material 2026 report, 80% of organizations report measurable economic returns on their AI investments.
The returns are concentrated among enterprises that invested in the infrastructure to support agents at scale. The ones that bolted agents onto existing workflows without addressing context readiness are the ones seeing projects canceled.
See the Context Layer Live
Watch Atlan turn metric definitions, policies, and lineage into context agents can query.
Watch the Live DemoWhat’s the bottom line on enterprise AI agents?
Permalink to “What’s the bottom line on enterprise AI agents?”The enterprise AI agent picture is real, growing, and uneven. Conversational analytics, customer service, data management, and IT operations are the functions where agents are most active. Finance, healthcare, retail, and manufacturing are the industries pushing fastest.
But the gap between experimentation and production is wide, and the primary reason is not model capability or framework maturity. It is context readiness.
The agents that stick understand the business context behind their tasks. The ones that stall have a capable model sitting on top of an empty, stale, or inconsistent picture of the business.
Enterprises are not short on agent ideas. They are short on the context that makes the ideas work.
FAQs about how enterprises use AI agents
Permalink to “FAQs about how enterprises use AI agents”How are enterprises using AI agents today?
Permalink to “How are enterprises using AI agents today?”Mostly in digital, repeatable functions: conversational analytics, customer service, IT and security operations, and data management. Adoption is highest in customer support (78%, Cloudera 2025) and coding assistance (90%, Anthropic 2026).
What is the most common AI agent use case?
Permalink to “What is the most common AI agent use case?”Customer service and contact centers lead in deployment volume: Sierra and Cresta have each surpassed $100 million in ARR, and Decagon is valued at $4.5 billion (2026). Conversational analytics is the fastest-growing use case among executive teams.
What share of enterprises actually run agents at scale?
Permalink to “What share of enterprises actually run agents at scale?”A small fraction. McKinsey’s State of AI 2025 survey found 62% of organizations experimenting with AI agents, but only 23% have scaled one into production.
Why do so few agent projects reach production?
Permalink to “Why do so few agent projects reach production?”The primary blocker is context and data readiness, not model capability. Gartner projects more than 40% of agentic AI projects will be canceled by 2027.
Which industries are furthest ahead in AI agent adoption?
Permalink to “Which industries are furthest ahead in AI agent adoption?”Telecommunications leads agentic AI adoption at 48%, with retail, CPG, and healthcare close behind at 47% (NVIDIA, 2026). Financial services trails on raw adoption at 42% but leads on maturity, with 70% of bank AI deployments now generative or agentic.
What do the successful AI agent use cases have in common?
Permalink to “What do the successful AI agent use cases have in common?”They run on trustworthy, governed business context: correct metric definitions, policy rules, or lineage. The context layer underneath each one stays accurate, current, and governed.
What ROI are enterprises seeing from AI agents?
Permalink to “What ROI are enterprises seeing from AI agents?”80% of organizations report measurable economic returns on AI investments (Anthropic and Material, 2026). Those returns concentrate among enterprises that invested in context infrastructure first.
How do agents get the business context they need?
Permalink to “How do agents get the business context they need?”Through a context layer: governed infrastructure that stores canonical definitions, policy rules, and lineage, then serves them to agents over structured protocols like MCP, instead of static exports or wikis.
Sources
Permalink to “Sources”- McKinsey’s State of AI 2025 survey, McKinsey. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- PwC’s 2025 AI Agent Survey, PwC. https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-agent-survey.html
- The Future of Enterprise AI Agents, Cloudera. https://www.cloudera.com/content/dam/www/marketing/resources/analyst-reports/the-future-of-enterprise-ai-agents.pdf
- The breakout year for enterprise AI agents, Cloudera. https://www.cloudera.com/blog/business/the-breakout-year-for-enterprise-ai-agents.html
- Sierra raises $950M as the race to own enterprise AI gets serious, TechCrunch. https://techcrunch.com/2026/05/04/sierra-raises-950m-as-the-race-to-own-enterprise-ai-gets-serious/
- Decagon completes first tender offer at $4.5B valuation, TechCrunch. https://techcrunch.com/2026/03/04/decagon-completes-first-tender-offer-at-4-5b-valuation/
- The 2026 State of AI Agents Report, Anthropic and Material. https://resources.anthropic.com/2026-state-of-ai-agents
- State of AI Report 2026, NVIDIA. https://blogs.nvidia.com/blog/state-of-ai-report-2026/
- State of AI in Financial Services Survey 2026, NVIDIA. https://blogs.nvidia.com/blog/ai-in-financial-services-survey-2026/
- Use Case Trends Q4 2025, Evident Insights. https://evidentinsights.com/insights/use-case-trends-q4-2025
- State of AI in Healthcare Survey 2026, NVIDIA. https://blogs.nvidia.com/blog/ai-in-healthcare-survey-2026/
- State of AI in Retail and CPG Survey 2026, NVIDIA. https://blogs.nvidia.com/blog/ai-in-retail-cpg-survey-2026/
- 2026 Manufacturing Industry Outlook, Deloitte. https://www.deloitte.com/us/en/insights/industry/manufacturing-industrial-products/manufacturing-industry-outlook.html
- Gartner predicts over 40 percent of agentic AI projects will be canceled by end of 2027, Gartner. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- Building effective agents, Anthropic. https://www.anthropic.com/engineering/building-effective-agents/
