The Year Agentic AI Got Real

At the start of 2025, Sam Altman told us AI agents would "join the workforce." Gartner predicted they'd handle 15% of day-to-day work decisions by 2028. Venture capitalists flooded the space with $2.8 billion in just the first half of the year.

And you know what? Some of it actually happened.

But if you were expecting a wave of fully autonomous AI employees taking over entire departments, 2025 probably disappointed you. If you were building practical systems that augment human work and deliver measurable ROI, this was your year.

Let me walk you through what really happened.

Here's the stat that tells the whole story: 60-70% of enterprises experimented with agentic AI in 2025. Only 15-20% deployed agents in production workflows touching real customers or critical business processes.

That gap isn't about technology capability. It's about three brutal realities:

First, reliability. A 5% error rate sounds acceptable until you're running a business. One corrupted database entry shuts down operations. One wrong order costs a customer relationship. The models got better in 2025, but "better" still isn't "reliable enough" for most critical business processes.

Second, integration complexity. I watched dozens of companies build impressive agent demos in weeks, then spend months trying to integrate with enterprise apps, legacy databases, security protocols, and compliance requirements. The technical debt exceeded the expected value in most cases. The demo-to-production gap killed more pilots than any technical limitation.

Third, cost at scale. Multiple clients piloted customer service agents with great results, then discovered that scaling to all customer interactions would cost more than their entire contact center budget. Token usage adds up fast when agents make multiple LLM calls per task.

The companies that closed this gap didn't have better technology. They had better strategy, realistic expectations, and organizational patience.

Despite the challenges, four categories saw genuine production success:

1. Customer Service (With Guardrails)

The winning pattern wasn't autonomous agents replacing humans. It was agents handling tier-one requests, performing triage, and pulling information for human agents. Capital One's Chat Concierge for auto dealerships achieved 55% better conversion rates. Salesforce Agentforce closed 18,000 deals by year-end.

The key insight: augmentation beats replacement every time.

2. Developer Assistance

GitHub Copilot and Cursor became standard tools. Why did this work? Developers can immediately evaluate output, mistake stakes are lower, and productivity gains are measurable. If you're not using AI-assisted coding by now, you're the outlier.

3. Data Analysis and BI

Agents that query databases, generate reports, and answer natural language questions succeeded because they're primarily read-only with easy verification steps. Low risk, high value.

4. Back-Office Operations

Document processing, invoice handling, compliance checking. One insurance company deployed agents for claims intake and processed 100,000+ claims in six months. Adjusters spent 40% less time on routine work.

Less glamorous than customer-facing AI. Better ROI and risk profile.

Notice what's missing? Fully autonomous decision-making. Strategic planning agents. Creative work replacement. Those remained experimental.

Mid-year, everyone started talking about multi-agent systems. Instead of one generalist agent trying to do everything, you'd have specialist agents working together: research agents, analysis agents, writing agents.

This happened because building one super-capable generalist proved extremely hard. Building three focused agents and orchestrating them often proved easier.

But here's what the successful deployments taught us: most use cases don't need multiple agents.

Before jumping to multi-agent architecture, ask: can one well-designed agent with good tool access handle this? Often, yes. When you do need multiple agents, keep teams small (3-5, not 20), maintain clear specialization, and expect significant orchestration complexity.

The companies that succeeded with multi-agent systems treated it as building a management layer for agents, not just building more agents. That requires different engineering, different monitoring, and different operational expertise.

If you deployed agents at scale in 2025, you hit governance challenges fast.

The EU AI Act started affecting real project timelines. Enterprises faced practical questions about prompt injection, data privacy across multiple systems, audit trails for agent decisions, and liability when things went wrong.

Organizations that succeeded built proactive frameworks:

The pattern that worked: start with more oversight than you think you need, then relax gradually as you build confidence. The opposite approach (deploy permissively, tighten when problems occur) led to incidents and rollbacks.

Positive surprises:

Model costs dropped faster than expected. What cost $100 to run in January might cost $30 by December. That made previously uneconomical use cases viable.

Code generation quality became legitimately good. Not perfect, but good enough that experienced developers saw measurable productivity gains.

Vertical applications in medicine, legal, and scientific domains delivered real value even when general-purpose agents struggled.

Disappointments:

Fully autonomous agents working independently for hours or days didn't materialize. Most successful agents still need frequent human check-ins or handle bounded tasks only.

The "AI replacing jobs" narrative clashed with reality. Tasks got automated, roles evolved, but wholesale job replacement didn't happen. That's probably better for society but meant ROI calculations were often wrong.

Reasoning over long horizons remained weak. Agents struggled with tasks requiring 20+ steps with branching logic. This limitation held back ambitious deployments all year.

The biggest deals signaled where the smart money went:

Salesforce made multiple AI-related acquisitions in 2025, focused heavily on data, agentic AI, and autonomous marketing and analytics capabilities:

·      Informatica – Enterprise AI-powered data management, integration, governance, and MDM, acquired for about $8B

·      Convergernce AI – UK-based AI agent company building adaptive agents that handle complex, multi-step workflows across changing digital interfaces

·      Spindle AI – Analytics and forecasting startup combining AI agents and ML with data modeling to simulate agentic scenarios and forecast business outcomes

·      Qualified – Agentic AI marketing/sales startup focused on AI agents that engage website visitors, qualify leads, and schedule meetings, integrating deeply with Salesforce in a deal reportedly valued around $1-1.5B

ServiceNow acquired Moveworks for $2.85B. Workday bought Sana for $1.1B. Traditional enterprise software companies were buying their way into agents rather than building from scratch.

Security became critical. Palo Alto Networks pursued CyberArk for $25B. F5 bought CalypsoAI for $180M. Everyone recognized agents create new attack surfaces.

The Microsoft-OpenAI partnership restructured completely in October. Microsoft's stake went to 27% (valued at $135B), but they lost exclusive cloud provider status. OpenAI committed to $250B in Azure purchases but can now deploy anywhere. Both companies need flexibility as Google's Gemini surged to 650 million monthly users.

And finally, actual standards emerged. OpenAI, Anthropic, Microsoft, and Google formed the Agentic AI Foundation in December. Agents MD was adopted by 60,000+ open-source projects within months. Industry alignment on interoperability matters more than any single technical breakthrough.

Based on what worked in 2025, here's where to put your energy:

1. Pick specific, bounded use cases. Resist the temptation to boil the ocean. Choose one or two workflows where agents can deliver measurable value. Get those to production quality. Learn from them before expanding.

2. Build governance now. Don't wait for regulations or incidents. Establish monitoring, accountability, and safety practices while you're still small-scale. The cost of retrofitting governance is exponentially higher.

3. Develop hybrid talent. You need people who understand both your business processes and technical capabilities. Start developing these skills internally because there's a massive talent shortage.

4. Plan for iteration, not transformation. The most successful deployments took sustained effort over months. Quick pilots enable learning, but real value comes from commitment to refinement, scaling what works, and killing what doesn't.

We moved from concept to operational reality in 2025. Not everywhere, not for everything, but in enough places with enough success to prove this is the future of work.

The challenges ahead are less about technical capability and more about implementation, governance, and organizational change. Those are solvable problems. We know how to do integration, build governance frameworks, and manage change. It's hard work, but it's familiar work.

2026 will separate the organizations that learned from 2025's hard lessons from those that are still chasing the hype.

Ready to move from experimentation to implementation?

I'm working with a small number of organizations on their 2026 AI strategy. If you're leading AI initiatives and want to avoid the mistakes that killed most 2025 pilots, let's talk.

My AI Blueprint service helps you identify high-value use cases, design realistic deployment roadmaps, and build the governance frameworks that make agents work at scale.

Schedule a strategy session or reply to this email to start a conversation about your specific challenges.

What worked (or didn't work) for you in 2025? Hit reply and let me know. I read every response and often feature real-world examples in future newsletters.

P.S. This week's Disambiguation podcast is a deep dive on this same topic—30+ minutes breaking down the year in agentic AI. Watch here if you want the full story.