If you’ve attended any tech conference in the past twelve months, you’ve been told that AI agents are going to transform everything. Customer service, software development, financial planning, supply chain management — name an enterprise function and someone has a slide deck showing how autonomous agents will revolutionise it.

The reality on the ground is considerably more nuanced. After speaking with dozens of enterprise tech leaders and reviewing deployment data from the past year, here’s where AI agents actually stand in production — and where the gap between promise and delivery remains wide.

What’s Actually Working

Let’s start with the genuine success stories, because they exist and they matter.

Customer service triage and resolution. This is the clearest winner. AI agents handling first-tier customer inquiries — order status, password resets, FAQ responses, simple troubleshooting — are now standard at most large consumer-facing companies. The technology isn’t new (chatbots have existed for years), but the quality has improved dramatically with large language models. Resolution rates for these simple queries are hitting 70-80% without human intervention at well-implemented deployments.

Companies like Intercom have built agent-first customer service platforms that handle the simple stuff automatically and route complex issues to humans with full context. The ROI here is straightforward and measurable.

Code generation and development assistance. GitHub Copilot, Cursor, and similar tools have crossed the threshold from novelty to essential workflow tool for many development teams. They’re not autonomous agents in the strict sense — they operate within human-directed workflows — but they’re the closest thing to a broadly deployed “AI agent” in knowledge work. Survey data from Stack Overflow suggests that 60-70% of professional developers now use AI coding assistants regularly.

Document processing and extraction. Pulling structured data from unstructured documents — invoices, contracts, medical records, regulatory filings — is another area where AI agents are delivering measurable value. What used to require teams of manual data entry operators is now handled by AI systems with human review for edge cases.

What’s Partially Working

These areas show promise but aren’t yet delivering consistent enterprise-grade results.

Sales and marketing automation. AI agents can draft emails, personalise outreach, score leads, and generate content. But the “fully autonomous SDR” that several startups promised? Not quite there. The best implementations have AI handling research and draft creation with humans reviewing and sending. Fully autonomous outreach tends to produce content that’s technically competent but lacks the specificity and relationship awareness that closes deals.

IT operations (AIOps). Using AI to monitor systems, correlate alerts, and diagnose issues is working well at the detection stage. Where it falls down is in autonomous remediation — actually fixing problems without human approval. Most enterprises are (rightly) cautious about letting AI agents make changes to production infrastructure, so the agent detects and recommends while humans decide and act.

Financial analysis and reporting. AI agents can pull data, generate charts, write narrative commentary, and flag anomalies. But the output still requires expert review before it goes to stakeholders. The risk of a confidently wrong number in a financial report is too high for fully autonomous operation.

What’s Still Mostly Hype

Autonomous enterprise decision-making. The vision of AI agents making significant business decisions independently — pricing, hiring, vendor selection, capital allocation — remains far from reality. And frankly, it should. These decisions involve judgment, context, stakeholder management, and accountability that can’t be delegated to a system that doesn’t understand consequences.

Multi-agent orchestration at scale. The idea of multiple specialised AI agents collaborating on complex tasks — one researching, one analysing, one writing, one reviewing — sounds elegant in conference demos. In practice, error propagation between agents, inconsistent context handling, and the complexity of coordinating multiple autonomous systems make this fragile in production. Some AI consultancy teams are making progress on specific, well-defined multi-agent workflows, but the general case remains unsolved.

Fully autonomous software development. Despite impressive demos of AI systems building entire applications from natural language descriptions, the reality in professional software development is that these tools are assistants, not replacements. They excel at boilerplate and well-defined patterns. They struggle with architecture decisions, edge cases, and the kind of systems thinking that experienced engineers bring.

The Pattern That’s Emerging

If you step back and look at where AI agents are succeeding versus failing, a clear pattern emerges:

Works well:

• High-volume, repetitive tasks
• Well-defined success criteria
• Low cost of individual errors
• Structured inputs and outputs

Struggles with:

• Novel situations requiring judgment
• High-stakes decisions with significant consequences
• Tasks requiring cross-domain context
• Interactions where trust and relationships matter

This isn’t surprising. It’s the same pattern we’ve seen with every automation technology in history. The interesting question isn’t whether AI agents will eventually handle more complex tasks — they will — but how fast the boundary between “works well” and “struggles with” moves.

What Enterprise Leaders Should Actually Do

Based on what I’ve seen working and failing, here’s pragmatic advice for 2026:

1. Deploy agents for clear, measurable use cases first. Customer service triage, document processing, and code assistance have proven ROI. Start there if you haven’t already.

2. Human-in-the-loop is not a weakness. It’s good engineering. Design your agent workflows with human review at decision points, especially for anything customer-facing or financially significant.

3. Measure ruthlessly. Agent performance degrades over time as edge cases accumulate. Monitor resolution rates, error rates, escalation rates, and customer satisfaction continuously.

4. Don’t reorganise around agents yet. The technology is evolving too fast to restructure your organisation around current capabilities. Build agent capabilities alongside existing teams rather than replacing them.

5. Budget for iteration. Your first agent deployment will not be your best. Plan for 3-6 months of tuning, prompt engineering, and workflow adjustment before you see optimal performance.

The AI agent revolution is real, but it’s happening as incremental productivity gains across many functions rather than as dramatic autonomous takeovers of entire departments. That’s less exciting than the conference keynotes suggest, but it’s far more useful.