32  How Enterprises Are Building AI Agents in 2026

Originally published January 2026.

32.1 From Experiments to Production

2025 was the year of AI agent experiments. Every enterprise tried pilots — a chatbot here, a document summarizer there, maybe a code assistant for the engineering team. Most of these experiments were promising but limited. They demonstrated potential without delivering transformation.

2026 is the year agents moved to production. Enterprises are deploying agents for real work — multi-stage workflows that handle meaningful portions of business processes. The experiments are over; the implementation has begun.

This chapter presents findings from Anthropic’s survey of over 500 technical leaders at enterprises across industries, combined with detailed case studies from organizations at the forefront of agentic AI adoption.

32.2 The Survey

Anthropic surveyed technical leaders — CTOs, VPs of Engineering, Heads of AI, and senior architects — at 500+ enterprises ranging from 1,000 to 100,000+ employees. The survey covered agent adoption patterns, use cases, challenges, and economic impact.

32.2.1 Key Findings at a Glance

Metric Finding
Multi-stage workflow deployment 57% of enterprises deploy agents for multi-stage workflows
Plans for increased complexity 81% plan more complex use cases in the next 12 months
AI for software development ~90% use AI for development
AI for production code 86% use AI for production code
Highest-impact use case Data analysis/reporting (60%)
Measurable economic returns 80% report measurable returns

32.3 Where Agents Are Being Deployed

32.3.1 Coding Leads the Way

Software development remains the dominant use case for AI agents in the enterprise:

Development Activity AI Usage Rate
Code generation (any purpose) ~90%
Code generation for production 86%
Code review and quality 74%
Test generation 71%
Documentation generation 68%
Bug fixing and debugging 65%
Dependency management 42%
Note

The fact that 86% of enterprises use AI for production code — not just prototypes or experiments — signals that AI-generated code has crossed the threshold from novelty to standard practice. The question is no longer “should we use AI for code?” but “how do we use it most effectively?”

32.3.2 Beyond Coding: Where Agents Add Value

While coding leads, agents are delivering impact across the enterprise:

Use Case % Reporting “High Impact” Description
Data analysis & reporting 60% Automated report generation, data pipeline monitoring
Customer support 55% Ticket triage, resolution suggestions, FAQ automation
Document processing 48% Contract analysis, invoice processing, compliance review
IT operations 45% Incident response, infrastructure provisioning, monitoring
Sales & marketing 42% Lead scoring, content generation, campaign analysis
HR & recruiting 35% Resume screening, onboarding automation, policy Q&A
Financial analysis 33% Risk assessment, fraud detection, financial reporting
Legal 28% Contract review, compliance checking, legal research

32.3.3 Multi-Stage Workflows

The most significant finding is that 57% of enterprises now deploy agents for multi-stage workflows — not just single-turn interactions, but complex processes that require multiple steps, tool calls, and decisions.

Examples of multi-stage workflows:

# Example: Automated invoice processing pipeline
agent = Agent(
    system_prompt="""Process incoming invoices:
    1. Extract data from invoice PDF (OCR + parsing)
    2. Validate against purchase order in ERP
    3. Check for duplicate invoices
    4. Route for approval based on amount and vendor
    5. If approved, enter into accounting system
    6. Send confirmation to requesting department
    Flag any discrepancies for human review.""",
    tools=["ocr", "erp_api", "email", "slack"],
)

32.3.4 Future Plans

81% of enterprises plan more complex agent use cases in the next 12 months. This means the current adoption is a floor, not a ceiling. The trajectory is toward:

  • More steps per workflow
  • More tools and integrations per agent
  • More autonomous decision-making
  • More departments and functions using agents
  • More agents working together (agent-to-agent communication)

32.4 Economic Impact

32.4.1 80% Report Measurable Returns

80% of enterprises report measurable economic returns from their AI agent deployments. This is a striking number — it means that agentic AI, unlike many technology trends, is delivering tangible value, not just hype.

Return Category % of Respondents
Significant ROI (>3x return) 28%
Moderate ROI (1.5-3x return) 34%
Positive ROI (1-1.5x return) 18%
Break-even 12%
Negative ROI 8%

The 8% reporting negative ROI typically had implementation challenges — poor data quality, lack of internal expertise, or unrealistic expectations — rather than fundamental technology limitations.

32.4.2 Where the Returns Come From

Source of Value Mechanism Example
Labor savings Automating repetitive tasks Invoice processing: 4 hours → 15 minutes
Quality improvement Reducing human error Data entry accuracy: 94% → 99.9%
Speed Faster processing Incident response: 5 hours → 7 minutes
Scale Handling more volume without more staff Customer support: 3x ticket volume, same headcount
New capabilities Things that weren’t possible before Real-time competitive analysis across 1000+ sources

32.5 Case Studies

32.5.1 Thomson Reuters: CoCounsel and 150 Years of Case Law

Thomson Reuters, the media and information conglomerate, built CoCounsel — an AI legal assistant that gives lawyers instant access to over 150 years of case law.

The Challenge: Legal research is time-consuming. Finding relevant precedents, analyzing case outcomes, and synthesizing legal arguments can take days of manual work.

The Solution: CoCounsel uses Claude to:

  • Search and analyze millions of legal documents
  • Identify relevant precedents and case law
  • Summarize legal opinions and rulings
  • Draft legal memos and briefs
  • Answer natural-language legal questions

The Results:

Metric Before After
Time for precedent search 2-4 hours 2-5 minutes
Documents reviewed per case 20-50 200-500
Legal memo drafting 4-8 hours 30-60 minutes
Lawyer satisfaction N/A 92% report time savings
Tip

The CoCounsel case study illustrates a key principle: AI agents are most valuable when they provide access to information that was previously inaccessible. No human lawyer can read 500 cases in an hour. An agent can. This doesn’t replace the lawyer’s judgment — it amplifies it.

32.5.2 eSentire: From 5 Hours to 7 Minutes

eSentire, a managed detection and response (MDR) company, deployed Claude-powered agents for security incident analysis and response.

The Challenge: Security analysts were spending hours investigating each alert — gathering context from multiple systems, correlating data, and determining the appropriate response.

The Solution: AI agents that:

  • Automatically triage incoming security alerts
  • Gather context from SIEM, EDR, and threat intelligence sources
  • Correlate indicators of compromise (IoCs)
  • Determine severity and recommend response actions
  • Draft incident reports

The Results: Average incident analysis time dropped from 5 hours to 7 minutes — a 43x improvement.

Metric Before After
Average incident analysis time 5 hours 7 minutes
Alerts processed per analyst/day 8-10 40-50
False positive rate 35% 12%
Mean time to response 45 minutes 8 minutes

32.5.3 Doctolib: 40% Faster

Doctolib, Europe’s leading healthcare technology company, uses AI agents to streamline clinical administrative workflows.

The Challenge: Healthcare administrators spend significant time on scheduling, documentation, and patient communication — time that could be spent on patient care.

The Solution: AI agents that handle:

  • Appointment scheduling and optimization
  • Patient communication (reminders, follow-ups)
  • Medical documentation assistance
  • Insurance pre-authorization processing

The Results: Administrative workflows are 40% faster, freeing up healthcare staff for patient-facing work.

Metric Improvement
Appointment scheduling speed 40% faster
Documentation time 35% reduction
Patient communication response time 60% faster
Insurance pre-auth processing 50% faster

32.5.4 L’Oréal: 99.9% Accuracy with 44,000 Monthly Users

L’Oréal, the global beauty and cosmetics company, deployed an AI-powered product information management agent that serves 44,000 monthly users across the organization.

The Challenge: Managing accurate, consistent product information across thousands of products, multiple languages, and dozens of regional markets is a massive operational challenge.

The Solution: AI agents that:

  • Automatically generate and update product descriptions
  • Translate and localize content for regional markets
  • Ensure regulatory compliance across jurisdictions
  • Maintain consistency across all channels (web, print, retail)

The Results:

Metric Before After
Data accuracy 94% 99.9%
Product description generation 2 hours/product 5 minutes/product
Localization time 3 days/language 4 hours/language
Monthly active users N/A 44,000
Note

L’Oréal’s 99.9% accuracy rate is particularly noteworthy. In product information management, even small errors can have significant consequences — regulatory fines, customer complaints, or lost sales. The AI agent’s ability to maintain near-perfect accuracy at scale is what made adoption possible across 44,000 users.

32.6 Challenges

Despite the success stories, enterprises face significant challenges in deploying AI agents. The top three:

32.6.1 1. Integration Complexity (46%)

The most common challenge is integrating agents with existing systems. Enterprises have dozens or hundreds of systems — ERPs, CRMs, data warehouses, custom applications — and agents need to connect to them.

Integration Challenge % Affected
Legacy systems with poor APIs 52%
Authentication and access management 47%
Data format incompatibility 43%
Network and security constraints 38%

MCP (Model Context Protocol) is emerging as a solution, providing a standardized way to connect agents to external systems. But legacy systems often require custom integration work.

32.6.2 2. Data Quality (42%)

Agents are only as good as the data they can access. Poor data quality — inconsistent formats, missing fields, duplicate records, stale information — undermines agent effectiveness.

# The data quality problem in practice
agent.run("Find all customers who haven't ordered in 90 days")

# If the customer database has:
# - Duplicate customer records (same person, different IDs)
# - Inconsistent date formats (MM/DD vs DD/MM)
# - Missing email addresses
# - Outdated company names (post-merger)
#
# The agent's results will be unreliable

Solutions include data cleaning pipelines, master data management systems, and — ironically — using AI agents to improve data quality (deduplication, normalization, validation).

32.6.3 3. Change Management (39%)

Technology is only part of the equation. People and processes must also adapt:

Change Management Challenge % Affected
Employee resistance/skepticism 48%
Skills gap (AI literacy) 45%
Process redesign requirements 41%
Governance and policy gaps 37%
Unclear ownership of agent initiatives 33%
Tip

Change management is often underestimated. The most successful enterprises invest heavily in training, communication, and gradual rollout. Start with a single department, demonstrate success, and expand. Attempting org-wide deployment on day one typically fails due to resistance and lack of readiness.

32.7 The Adoption Maturity Model

Based on the survey data, enterprises fall into four maturity levels:

Level Description % of Enterprises
Level 1: Experimental Pilots and proofs of concept, limited scope 15%
Level 2: Departmental Production agents in specific departments 35%
Level 3: Cross-functional Agents spanning multiple departments and workflows 38%
Level 4: AI-Native Agents deeply integrated into core business processes 12%

The trajectory is clear: enterprises are moving from Level 1 toward Level 4. The 81% who plan more complex use cases are investing in moving up the maturity ladder.

32.8 Best Practices from Leading Enterprises

Based on the case studies and survey data, here are the practices that distinguish successful implementations:

32.8.1 1. Start with a Clear Use Case

Don’t deploy agents because it’s trendy. Identify a specific, measurable problem where agents can deliver value.

32.8.2 2. Invest in Data Quality First

Before deploying agents, ensure the data they’ll access is clean, consistent, and accessible.

32.8.3 3. Build Incrementally

Start with a simple workflow, prove value, then add complexity. Don’t attempt a 20-step workflow on day one.

32.8.4 4. Measure Everything

Track time savings, quality improvements, cost reductions, and user satisfaction. Data drives continued investment.

32.8.5 5. Invest in People

Train your team, communicate openly about changes, and involve employees in the design process.

32.8.6 6. Plan for Scale from the Start

What works for 10 users may not work for 10,000. Design your architecture to scale from the beginning.

32.9 Looking Ahead

The survey was conducted in early 2026. Given the pace of AI advancement, the landscape will look different by year-end. But the fundamental patterns are clear:

  • Agents are moving from experiments to production
  • Multi-stage workflows are the new standard
  • Coding leads, but other use cases are catching up
  • Economic returns are real and measurable
  • Challenges remain, but they’re solvable

The enterprises that succeed will be those that approach agentic AI strategically — with clear use cases, solid data foundations, strong change management, and a commitment to measuring and improving outcomes.

32.10 Key Takeaways

  • 57% of enterprises deploy agents for multi-stage workflows, and 81% plan more complex use cases within 12 months.
  • Coding leads adoption: ~90% use AI for development, 86% for production code.
  • Data analysis and reporting is the highest-impact non-coding use case (60% report high impact).
  • 80% of enterprises report measurable economic returns from agent deployments.
  • Case studies demonstrate transformative results: Thomson Reuters (150 years of case law, 2-4hr → minutes), eSentire (5hr → 7min incident analysis), Doctolib (40% faster workflows), L’Oréal (99.9% accuracy, 44K monthly users).
  • Top challenges: integration complexity (46%), data quality (42%), and change management (39%).
  • Maturity is increasing: 38% are at cross-functional deployment, 12% are AI-native.
  • Success requires: clear use cases, data quality investment, incremental deployment, comprehensive measurement, people investment, and scalable architecture.
  • The trajectory is clear: 2026 is the year agents moved from experiments to production, and the pace is accelerating.