Agentic AI is autonomous software that plans, chooses tools, executes multi-step tasks, and reviews outcomes with limited human supervision. It matters because enterprise work is fragmented across systems, data, and approvals, and agents can coordinate those steps faster than manual processes. When governed well, they automate high-volume workflows, reduce routine workload, and force better collaboration, accountability, and controls around machine autonomy.
Enterprise leaders face a paradox. Over half of organisations already deploy agents in production, yet Gartner predicts 40% of agentic AI projects will be cancelled by 2027.[1] The gap between hype and reality demands clarity.
What Agentic AI Is
Most AI systems today are reactive: ask a question, get an answer. Request a task, receive a result. The AI waits for your instruction, then responds.
Agentic AI operates differently. These systems work more like autonomous colleagues than responsive tools. Give an agent a goal—"Process this insurance claim" or "Analyse this customer support issue"—and it:
- Plans an approach to achieve the goal
- Decides which tools and data sources to use
- Executes multiple steps autonomously
- Adapts when it encounters obstacles
- Evaluates its own outputs before presenting results
The key distinction: agentic systems pursue goals through multi-step reasoning rather than responding to individual prompts.
How Agents Actually Work
Modern AI agents combine three core capabilities:
Autonomous Decision-Making: Agents operate within defined parameters but choose their own path to accomplish goals. They don't require human approval at each step.
Tool Integration: Agents connect to databases, APIs, knowledge bases, and other systems. They query information, trigger actions, and coordinate across platforms.
Multi-Step Planning: When faced with complex tasks, agents break them into subtasks, sequence them logically, and adjust plans based on intermediate results.
Unlike traditional automation that follows predetermined rules, agents use language models as reasoning engines to determine which actions to take and in which order.
Multi-Agent Systems: Collaborative Intelligence
Single agents handle specific tasks. Multi-agent systems tackle enterprise complexity by coordinating multiple specialised agents.
For example, processing a complex customer service request might involve:
- A triage agent that analyses the issue
- A knowledge agent that searches documentation
- A transaction agent that accesses order history
- A resolution agent that determines the appropriate fix
- An orchestrator agent that coordinates the workflow
Frameworks like AutoGen enable agents to communicate in natural language to solve problems collaboratively, while CrewAI takes a role-based approach treating agent teams like human work teams with defined responsibilities.
Why Most Projects Will Fail
Most agentic AI initiatives will not succeed.
Gartner's Sobering Prediction
Gartner predicts over 40% of agentic AI projects will be cancelled by 2027, citing escalating costs, unclear business value, and inadequate risk controls.[1]
Anushree Verma, Senior Director Analyst at Gartner, stated bluntly: "Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied."[1]
Only 19% of organisations made significant investments in agentic AI, while 42% made conservative investments, signaling caution. The remaining 39% are waiting or unsure.[1]
Most damning: widespread "agent washing" plagues the market. Of thousands of vendors claiming agentic solutions, only about 130 offer genuine autonomous capabilities.[1] The rest rebrand chatbots.
McKinsey's Scale Problem
The failure to scale isn't unique to agentic AI—it's endemic to AI initiatives generally. Fewer than 10% of use cases make it past the pilot stage. Over 80% of companies report no material earnings contribution from gen AI initiatives.[2]
The problem isn't technical. McKinsey identifies the real challenge: "coordination, judgment, and trust." As agents evolve from passive tools to proactive actors, organisational complexity increases dramatically.[2]
Three challenges dominate:
- Workflow integration: How humans and agents coexist in day-to-day work
- Governance: Establishing control over systems that act autonomously
- Agent sprawl: Preventing chaos as agent creation becomes democratised
Unlike traditional AI tools that plug into existing workflows, agents demand fundamental process redesign. Organisations treating them as "just another tool" fail to capture value.
Frameworks Worth Adopting
Three production-ready frameworks dominate the landscape:
LangChain/LangGraph: The most established ecosystem with extensive documentation and multi-provider support. LangGraph (the newer iteration) offers fine-grained control over agent behaviour. Best for document analysis, RAG systems, and complex orchestration across tools. If you need flexibility to switch between OpenAI, Anthropic, or open-source models, start here.
AutoGen: Microsoft's conversation-first framework where agents communicate in natural language to solve problems collaboratively. Version 0.4 brings event-driven architecture designed for scale. Includes AutoGen Studio (no-code GUI) and AutoGen Bench (performance evaluation). Best for code generation, research workflows, and Microsoft-centric organisations.
CrewAI: Role-based coordination treating agents like team members. Define roles (Researcher, Analyst, Writer), assign tasks, let the framework handle the rest. Simpler learning curve and 100+ pre-built integrations. Best for content workflows, competitive intelligence, and rapid prototyping.
Over half of organisations already use agents in production, with 78% planning expansion.[3] The framework choice matters less than execution discipline—most successful teams start with one and iterate based on results, not theory.
Enterprise Implementations
Wiley: 213% ROI in Publishing Customer Service
Global publisher Wiley faced a common challenge: seasonal spikes in customer service requests at the start of each semester without proportionally increasing headcount.[4]
Wiley deployed Salesforce Agentforce for dynamic, conversational self-service. The agent answers questions using Wiley's knowledge base and automatically resolves account access issues and password resets.
Results:[4]
- 40% improvement in case resolution rates
- 50% faster onboarding for seasonal agents
- 213% return on investment
- $230,000 in documented savings
The agent didn't replace human agents—it handled routine inquiries, allowing humans to focus on complex issues. During peak periods, this hybrid approach maintained service quality without scaling headcount proportionally.
Salesforce Agentforce: 10,000+ Agents at Launch
In September 2024, Salesforce unveiled Agentforce—autonomous agents that handle tasks across service, sales, marketing, and commerce. Unlike chatbots that respond to prompts, Agentforce agents retrieve data, reason through complex tasks, and independently execute decisions using the Atlas Reasoning Engine.
At Dreamforce 2024, customers built over 10,000 autonomous agents to address specific business problems.[5]
Early adopters include:
- OpenTable: Triaging incoming reservation requests
- Saks Fifth Avenue: Automating product returns and exchanges
- Disney: Customer service applications
Global system integrators—Accenture, Deloitte Digital, Capgemini, IBM Consulting, PwC, and Slalom—are deploying Agentforce for enterprise clients.
Implementation Patterns
What distinguishes successful projects from failures:
Start With High-Volume, Rule-Based Workflows
Agents excel at tasks with:
- Clear success criteria: You can define what "done" looks like
- High volume: Enough repetition to justify development effort
- Existing rules: Current processes you can encode as guardrails
- Access to data: The agent can query necessary information
Examples that work well:
- Customer service triage and resolution (Wiley's use case)
- Document processing and data extraction
- Lead qualification and routing
- Research and summarisation tasks
Examples that typically fail:
- Processes requiring significant judgment calls without clear criteria
- Low-volume tasks where manual handling is more efficient
- Workflows where failure has severe consequences (without extensive human oversight)
- Tasks requiring access to data that's siloed or unavailable
Design for Human-Agent Collaboration
The most successful implementations don't eliminate humans—they create effective human-agent teams.
McKinsey identifies the organisational challenge: "how humans and agents cohabit day-to-day workflows." Organisations that succeed:
- Define clear handoff points: Agents handle routine tasks; humans manage exceptions
- Make agent reasoning transparent: Humans can see why agents made specific decisions
- Enable easy escalation: Agents recognise their limits and hand off to humans
- Collect feedback continuously: Human corrections improve agent performance
The Wiley case exemplifies this: agents resolve straightforward issues, while complex inquiries route to human agents who now have time to handle them properly.
Expect Organisational Challenges, Not Just Technical Ones
McKinsey emphasizes: "unlike gen AI tools that could be easily plugged into existing workflows, AI agents demand a more foundational shift." This requires rethinking business processes, not just adding new technology.
Common organisational barriers:
- Process redesign resistance: Teams accustomed to current workflows resist fundamental changes
- Governance gaps: Existing approval processes don't account for autonomous systems
- Skill gaps: Teams lack experience monitoring and managing agent behaviour
- Trust deficits: Stakeholders uncomfortable delegating decisions to AI
Address these before building. The best technical implementation fails if the organisation isn't ready. This is exactly the kind of groundwork that a structured ML implementation process is designed to handle.
Monitor Relentlessly (Or Accept Failure)
As agents scale, monitoring becomes critical. Rolling out hundreds of agents creates a challenge: when mistakes happen, diagnosing what went wrong becomes difficult.
Essential monitoring practices:
- Log every agent decision with reasoning context
- Track success rates for different task types
- Identify edge cases where agents struggle
- Measure business impact, not just technical metrics
- Establish alerts for unusual behaviour patterns
Organisations treating agents as "set and forget" automation inevitably encounter failures they can't diagnose or prevent.
The Market in 2025
The Model Provider Landscape Is Shifting
Anthropic now holds 32% of the enterprise LLM market share by usage, while OpenAI holds 25%—a dramatic shift from 2023, when OpenAI commanded 50% and Anthropic just 12%.[6] For coding specifically, Anthropic's enterprise usage is more than double OpenAI's.[6]
This matters for agentic AI because underlying language models power agent reasoning. Enterprises are selecting based on reasoning capabilities, not just text generation. Anthropic's Claude Opus 4 and Sonnet 4 specifically target advanced reasoning and agentic workflows.
Three Philosophical Approaches
The major AI labs pursue distinct visions:
OpenAI: Launched Operator in January 2025, powered by the Computer-Using Agent (CUA) model. Introduced AgentKit in October 2025, packaging agent building blocks to reduce orchestration complexity. Philosophy: Programmable substrate for developers.
Anthropic: Introduced Computer Use for Claude 3.5 Sonnet in October 2024, enabling agents to interact with computer interfaces like humans would. Philosophy: Human-in-the-loop app creation with safety guardrails.
Google: Focus on enterprise-scale governance and integration with Google Cloud and Workspace. Philosophy: Governed enterprise scale with security and compliance built in.
Different approaches suit different organisational needs. Enterprises will likely use multiple platforms.
Market Projections vs. Reality
The AI agent market is projected to reach $47.1 billion by 2030. But temper this optimism with Gartner's 40% cancellation prediction. The gap between projected market size and actual enterprise success reveals significant hype. Many announced "agent" products are rebranded automation tools.
Gartner's prediction that 15% of day-to-day work decisions will be made autonomously by 2028 (up from 0% in 2024) is more grounded—real but measured progress, not overnight transformation.
The Strategic Question for Leaders
Agentic AI is simultaneously overhyped and strategically important.
It's overhyped because:
- 40% of projects will fail
- Most deployments never scale
- Vendor "agent washing" creates false expectations
- Organisational challenges exceed technical challenges
It's strategically important because:
- 51% of organisations already deploy agents in production
- 78% plan to expand deployment
- Real frameworks exist (LangChain, AutoGen, CrewAI)
- Verified enterprise successes demonstrate achievable value
The strategic question isn't "Should we explore agentic AI?" Most competitors already are. The question is: "How do we succeed where 40% will fail?"
What Distinguishes Success from Failure
Successful organisations:
- Start with constrained, high-value use cases (like Wiley's customer service)
- Design for human-agent collaboration, not replacement
- Address organisational challenges before technical implementation
- Monitor relentlessly and iterate based on real performance
- Set realistic expectations about what agents can and cannot do
- Choose frameworks based on organisational context, not hype
Organisations likely to fail:
- Pursue broad transformation without focused pilots
- Treat agents as cost-cutting tools to eliminate headcount
- Underestimate organisational change management requirements
- Deploy without adequate monitoring and governance
- Expect perfect autonomous operation from day one
- Select vendors based on marketing rather than capabilities
The Timeline Question
Should you start now or wait for maturity?
The case for starting now:
- Frameworks are production-ready
- Over half of organisations already operate in production
- Learning curves take time—competitors starting now gain experience
- Limited, well-scoped pilots have acceptable risk profiles
The case for waiting:
- 40% of projects will be cancelled—avoid becoming a statistic
- Organisational challenges often exceed readiness
- Frameworks evolve rapidly
- The ecosystem isn't yet settled
Start with small, constrained pilots in high-value areas where you can accept failure. Avoid betting the enterprise on agentic AI transformation. Learn, iterate, and scale when you have evidence of value. For a real-world example of this measured approach delivering results, see how Spartan Waterproofing deployed AI to transform their lead response process.
Organisations waiting for perfect maturity will lag behind. Organisations rushing into broad transformation will likely fail. The middle path—measured experimentation with realistic expectations—positions you to capture value while managing risk.
What Matters and What Doesn't
The technology works. Production frameworks exist, enterprises are deploying successfully, and verified ROI exists. But vendor hype obscures real limitations.
What works: Agents autonomously handle research, summarisation, customer service triage, and data processing. Companies like Wiley document 213% ROI. Over half of organisations run agents in production today.
What doesn't: Universal transformation claims. Only ~130 of thousands of vendors claiming "agentic AI" offer genuine autonomous capabilities—the rest rebrand chatbots. Over 80% of companies report no material earnings from gen AI initiatives. Fewer than 10% of use cases scale past pilots.
The 40% project cancellation rate Gartner predicts stems from organisational readiness failures, not technical limitations. The technology evolved faster than enterprise capabilities to deploy it effectively.
Getting Started: A Practical Roadmap
Select constrained, high-value use cases: Target high-volume workflows with clear success criteria, existing business rules, and accessible data. Score candidates on business impact, technical feasibility, data availability, and organisational readiness. Avoid transformation initiatives and mission-critical processes without extensive oversight.
Build a minimum viable pilot: Choose your framework (don't overthink it—most teams use multiple eventually). Start with a single workflow, human-in-the-loop approvals, comprehensive logging, and clear escalation paths. Test with 10-50 real workflows, monitor closely, and document failure modes.
Scale deliberately or learn from failure: If successful, expand scope incrementally and build monitoring before scaling. If unsuccessful, diagnose whether the issue is use case selection, implementation, or organisational readiness. 40% of projects fail—treat it as data, not defeat.
Optimise continuously: Monitor weekly, gather user feedback, update based on real interactions, and expand capabilities gradually. Budget 10-20% of development effort for ongoing optimisation. Agents that stagnate eventually fail.
The Bottom Line
Agentic AI is real. Production frameworks exist. Enterprises deploy agents successfully. Over half are already in production.
But success requires realistic expectations and disciplined execution:
- 40% of projects will fail
- Fewer than 10% will scale past pilots
- Organisational challenges exceed technical challenges
- Most companies see no material earnings impact from gen AI yet
The strategic opportunity exists for organisations that:
- Start with constrained, high-value pilots
- Design for human-agent collaboration
- Address organisational readiness honestly
- Monitor relentlessly and iterate based on evidence
- Set realistic expectations about capabilities and timelines
The organisations that succeed won't be those that started first or spent most. They'll be those that learned fastest from focused experimentation — following a methodology that prioritises learning over perfection.
Agentic AI isn't magic. It's engineering with realistic constraints, organisational change management, and disciplined execution. Approached this way, it delivers meaningful value. Approached as a silver bullet, it joins the 40% of cancelled projects.
References
Gartner Predicts 40% of Agentic AI Projects Will Be Abandoned by 2027 - Gartner Press Release (December 2024)
https://www.gartner.com/en/newsroom/press-releases/2024-12-02-gartner-predicts-40-percent-of-agentic-ai-projects-will-be-abandoned-by-2027The State of AI in 2024 - McKinsey Global Survey (2024)
https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-aiSalesforce State of AI Report - Salesforce Research (2024)
https://www.salesforce.com/resources/research-reports/state-of-ai/Wiley Customer Success Story - Salesforce Case Study (2024)
https://www.salesforce.com/customer-success-stories/wiley/Agentforce Launch Zone: Dreamforce 2024 - Salesforce Newsroom (September 2024)
https://www.salesforce.com/news/stories/agentforce-launch-zone/2025 Mid-Year LLM Market Update: Anthropic Leads Enterprise Market Share - Menlo Ventures (July 2025)
https://menlovc.com/perspective/2025-mid-year-llm-market-update/
Design Your Agentic AI Strategy
Agentic AI offers genuine opportunities to automate complex workflows and augment your team's capabilities. But success demands strategic planning, appropriate use case selection, and organisational readiness.
Our team helps enterprises design agentic AI strategies that navigate the gap between hype and reality. We assess workflows, identify high-impact opportunities, select appropriate frameworks, and guide implementation from pilot to production.
Whether you're exploring your first agent deployment or working to scale existing implementations, we help you avoid the pitfalls that lead to the 40% cancellation rate while capturing real value.
Start your agentic AI initiative
Frequently Asked Questions
What is agentic AI and how does it differ from chatbots?
Agentic AI is autonomous software that plans, selects tools, executes multi-step tasks, and reviews its own outcomes. Unlike chatbots that respond to single prompts, agents orchestrate complex workflows across multiple systems with minimal human supervision.
Which frameworks are used to build enterprise AI agents?
The most widely adopted frameworks include LangChain, LangGraph, CrewAI, and AutoGen. Each offers different trade-offs between flexibility, multi-agent orchestration, and production readiness.
Why do most enterprise agentic AI projects fail?
Most failures stem from insufficient governance, unclear task boundaries, and poor integration with existing business processes. Agents need well-defined scopes, human oversight checkpoints, and robust error handling to succeed in production.
What business processes are best suited for agentic AI?
High-volume workflows with structured steps and clear success criteria work best. Examples include document processing pipelines, customer onboarding sequences, compliance checking, and multi-system data reconciliation.
How should enterprises start with agentic AI implementation?
Start with a single, well-scoped workflow where the current manual process is documented and measurable. Build governance controls from day one, measure against the manual baseline, and expand only after proving reliability in production.