Updated October 2026.
Agentic AI means AI systems that plan steps, call tools and take actions toward a goal with limited human input, rather than just answering a question. Every vendor now calls its product agentic. Gartner estimates that only about 130 of the thousands of agentic AI companies are genuine; the rest are “agent washing” existing chatbots and automation tools. So this post ignores the demos and looks at agentic AI use cases only where agents are running on live data, for real users, today.
Why the surveys disagree on how many companies use AI agents
Ask three research firms how many companies run AI agents in production and you get three very different numbers.
| Source | Who they asked | What they found |
|---|---|---|
| LangChain, State of Agent Engineering (Jun 2026) | 1,340 people building agents; 63% from tech companies, 49% from companies under 100 people | 57% have agents in production; 67% at companies with 10,000+ employees |
| McKinsey, State of AI 2026 (Aug 2026) | 1,719 respondents across 97 countries and all industries | About 2 in 10 are scaling agents across the enterprise; 40% of $1B+ companies are scaling agents in at least one function, versus 22% of smaller ones |
| Deloitte, Tech Trends 2026 | Enterprise leaders | 11% have agents in production; 38% are piloting |
None of these is wrong. LangChain mostly surveyed people who already build agents, so its number runs high. McKinsey asks about scaling across a business, which is a higher bar than one agent running somewhere. The honest summary: most technology companies now run at least one agent, and most other companies don’t run any at scale yet.
The gap also comes down to what “in production” means. We use four tests. The agent works on live data, for real users. Someone gets paged when it breaks. It’s measured on a business number, not a demo score. And it keeps running after the person who built it moves on.
Agentic AI use cases by function: where agents run today
This is the state of play in October 2026, combining the McKinsey and LangChain data with what we see in our own AI agent deployments.
| Function | What the agent does in production | Maturity | Where a human stays in the loop | Evidence |
|---|---|---|---|---|
| Software engineering | Writes, tests and reviews code; opens pull requests | Scaled | Code review and merge | McKinsey: about 2 in 10 organisations scale coding agents, 31% of large ones. LangChain: coding agents are the most-used agents day to day |
| Customer service | Resolves routine chats and tickets, looks up orders, hands off the rest | Scaled for routine queries | Refunds, exceptions, complaints | LangChain: the most common primary use case (26.5%) |
| IT operations and service desk | Triages tickets, resets access, runs standard fixes | Scaled | Changes to production systems | McKinsey: IT is one of the functions where agents are most often scaled |
| Knowledge management and research | Searches internal documents, summarises, answers staff questions | Scaled | The decision made on the answer | McKinsey names knowledge management; LangChain: research and data analysis 24.4% |
| Marketing and sales | Drafts content, researches accounts, qualifies leads | Scaling, led by consumer goods and retail | Brand sign-off, pricing | McKinsey industry breakdown |
| Supply chain and manufacturing | Watches inventory, flags demand shifts, handles process exceptions | Early scaling at advanced manufacturers | Orders and production changes | McKinsey: automotive, aerospace and semiconductor respondents |
| Back office: procurement, tenders, invoices | Reads tenders and invoices, extracts terms, drafts responses, reconciles | Pilot to early production | Approvals and payments | Our field view: see our procurement agent and tender agent case studies |
| Decisions with money, safety or legal consequences | Approves credit, changes prices, acts on safety systems | Rare | Almost every step | Gartner expects 15% of day-to-day work decisions to be made autonomously by 2028, up from none in 2024 |
Two patterns stand out. First, the scaled rows are high-volume, text-heavy and forgiving: a bad first draft of code or a wrong answer to a routine question gets caught cheaply. Second, as the cost of a mistake rises, the human moves closer to the action, not further away.
What enterprise AI agents in production have in common
The agents we’ve seen survive contact with real operations share six traits. None of them is about which model you pick.
- A narrow job with a clear finish line. “Resolve password resets” ships. “Handle IT” doesn’t.
- Read widely, write narrowly. They can see a lot of systems but change very few. Write access is the thing security reviews care about, and rightly so.
- A human at the irreversible step. The agent drafts the refund; a person clicks approve. Once the error rate is known, that checkpoint can move.
- Rules where rules work. The predictable parts run as ordinary code. The model handles the judgement calls in between. Agents built entirely on the model are the ones that work 80% of the time and fail strangely on the rest.
- Watched and tested. Among teams with agents in production, 94% have observability in place. But only 52% of all teams run offline evaluations against known cases, according to LangChain. Tracing tells you what happened. Evaluations tell you whether it was right. Most teams have the first and skip the second.
- An owner and a number. Someone in the business is measured on whether it works: tickets closed, hours per tender, time to resolution.
Where agentic AI still stalls, and why
Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear value or weak risk controls. The practitioner data shows what sits behind that forecast.
- Quality is the main blocker. A third of LangChain’s respondents named quality as their biggest barrier to production: accuracy, consistency and staying within policy. At companies with 10,000+ employees, hallucinations and inconsistent outputs topped the write-in answers.
- Security rises with company size. Among enterprises with 2,000+ employees, security was the second biggest concern (24.9%), ahead of latency.
- Governance lags adoption. Deloitte found just 21% of organisations had a mature governance model for agentic AI, in a survey of 3,235 leaders across 24 countries.
- Speed matters in front of customers. Latency was the second biggest barrier overall (20%). Multi-step agents are slower than a single model call, and customers notice.
- Running costs are starting to bite. About one in ten McKinsey respondents said costs, including tokens, had limited their use of AI agents.
We covered the general version of this problem in Why AI Pilots Stall Before Production.
When not to use an AI agent
An agent is the expensive, non-deterministic option. Use it when the work genuinely needs judgement. Tick every statement that’s true for the process you have in mind:
- The process follows fixed rules from start to finish.
- A wrong action can’t be undone, and there’s no human checkpoint before it.
- The agent can’t get clean access to the systems or data it needs.
- You can’t write down 50 examples of a right and a wrong result.
- Users need an answer in under a second.
- The task happens a few times a week, not hundreds of times a day.
- Nobody in the business will own the outcome.
Two or more ticks means you probably want ordinary workflow automation, a better form or a fixed integration first. That’s not a failure. Some of the best agent projects we’ve done started as a rules engine and added a model only for the cases the rules couldn’t handle. We’ll go deeper on this choice in our November post on agentic AI vs automation.
How to pick your first production agent
Score each candidate process from 1 to 3 on five questions:
| Question | 1 | 2 | 3 |
|---|---|---|---|
| How often does it happen? | Weekly | Daily | Hundreds of times a day |
| How costly is a mistake? | Expensive or irreversible | Fixable with effort | Cheap to catch and correct |
| Can the agent reach the data and tools? | No API, messy data | Partial access | Clean read access, limited write |
| Is success measurable today? | No baseline | Rough baseline | A number the business already tracks |
| Is there a business owner? | No | Shared or unclear | One named person |
12 to 15: a strong first candidate. 9 to 11: fix the weakest score before you build. 8 or below: not yet.
What this means for mid-sized companies
The McKinsey split is the part to worry about. Large companies went from 27% to 40% scaling agents in one year; smaller organisations stayed at 22%. LangChain’s numbers look better for small firms, but its small respondents are mostly tech companies building agents into their own products.
For a mid-sized manufacturer, distributor or financial services firm, the blocker is rarely the idea. It’s engineering capacity: someone who can connect an agent to an ERP, write the evaluations and stay until the operations team trusts it. That’s the gap our forward deployed engineers fill, whether you need a full AI agent development team or one senior engineer embedded with yours.
Frequently asked questions
What is agentic AI?
AI systems that plan multi-step tasks, use tools such as APIs and databases, and take actions toward a goal with limited human input. A chatbot answers a question; an agent completes the task the question was about.
What are the most common agentic AI use cases in enterprises?
In 2026 the most common production use cases are coding agents in software engineering, customer service, IT operations and internal knowledge work such as research and document search. Back-office work like procurement and invoices is moving from pilot to production.
How many companies have AI agents in production?
It depends on the survey. LangChain found 57% of agent builders have agents in production. McKinsey found about 2 in 10 organisations scaling agents across the enterprise, and Deloitte found 11% with agents in production.
Why do agentic AI projects fail?
Mostly because of output quality, security concerns, weak governance and unclear business value, not model capability. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027.
When should you not use an AI agent?
When the process follows fixed rules, a wrong action can’t be undone without a human check, the agent can’t reach clean data, answers are needed in under a second, the task is rare, or nobody in the business owns the outcome. Ordinary workflow automation is usually the better first step there.
What is the difference between an AI agent and automation?
Automation follows fixed rules you define in advance. An AI agent decides its own steps, which handles messy cases better but makes it less predictable and more expensive to run. Many production systems combine both.
How do I choose an AI agent development company?
Ask to see agents they’ve put into production that are still running, how they evaluate quality, and who owns the system after launch. See how KnackLabs builds production AI agents.
Kartik Bansal is the Founder and CEO of KnackLabs, an AI transformation firm that embeds senior forward deployed engineers inside businesses to build, ship and run production AI systems. KnackLabs spent seven years as Haptik’s delivery arm, shipping conversational AI across 15B+ conversations. Follow Kartik on LinkedIn.
Book a 30-minute call and we’ll scope your first production agent.
Sources
- LangChain, State of Agent Engineering (12 Jun 2026; 1,340 respondents surveyed 18 Nov to 2 Dec 2025)
- McKinsey, The state of AI in 2026 (Aug 2026)
- Deloitte, Tech Trends 2026
- Deloitte, State of AI in the Enterprise 2026 (3,235 leaders, 24 countries)
- Gartner, prediction on agentic AI project cancellations by end of 2027 (Jun 2025)
- Gartner, estimate of genuine agentic AI vendors (“agent washing”) and autonomous work decisions by 2028
- Agentic AI
- AI agents
- Enterprise AI
- Production AI
- AI governance
