TL;DR
- Docker’s 2026 State of Agentic AI report surveyed 805 technical practitioners. Sixty percent of them already run AI agents in production.
- Read one paragraph further and the story flips. Almost every one of those agents points inward, at DevOps and internal productivity. Almost none of them touch a customer.
- The thing holding agents back is trust, not intelligence. Forty percent name security as the number one barrier, 33% cite technical complexity, and 85% know the Model Context Protocol but call it not enterprise-ready.
- The first agents that talk to buyers will not be the smartest ones in the room. They will be the most governed ones.
- If you run go-to-market, this is your problem too. Build the trust layer now, or watch a competitor’s agent get to your buyer first.
Docker’s new State of Agentic AI report is easy to misread. The headline number says sixty percent of organizations already run AI agents in production, out of 805 technical practitioners surveyed across North America, Europe, and Asia-Pacific. The first pass reads like the agent era finally showed up.
Then you read one more paragraph and the story flips. Docker’s own framing is that adoption today is “focused on internal workflows that boost team productivity while minimizing business risk.” The most common jobs these agents do are DevOps and CI/CD (38%), security automation (35%), and general process automation (34%). Those are back-office jobs. Agents that clean up pipelines, write internal tickets, and watch dashboards. They do not talk to buyers, and nobody is letting them near revenue. Adoption also skews hard toward the industries with the most to automate: telecom at 67%, financial services at 58%, technology at 54%, and retail at 50%.
Everyone Has an Agent. Almost Nobody Has a Customer-Facing One
That gap is the real story, and it maps cleanly onto go-to-market teams. Marketing and sales are where an agent would matter most, because that is where the work is repetitive, high-volume, and drowning in context. Research. Signal capture. List building. First-draft outreach. Each of those is an obvious agent job and I know exactly why none of them have one yet. The moment an agent touches a customer, the risk profile changes.
A broken internal agent is a shrug. It writes a sloppy summary, a teammate fixes it on Friday, and nobody outside the building ever knows. A broken customer-facing agent is an incident. It emails the wrong person, invents a claim, or quotes the wrong price, and now the trust damage is public. That is why every agent in that survey points inward. It is not a technology gap. It is a governance gap wearing a technology costume.
The Barrier Is Trust, Not Intelligence
Docker’s numbers make this concrete. Forty percent of respondents name security as their top challenge in scaling agentic AI. Forty-five percent say their hardest problem is guaranteeing that the tools their agents use are secure, trusted, and enterprise-ready. One in three (33%) name technical complexity, and the report singles out orchestration as the hardest part of the entire agent lifecycle. Then there is the line that should stop you cold. The survey found teams operating in “leap-of-faith mode,” adopting the technology without the security guarantees they would demand from any other enterprise system.
I think most AI marketing content misses this. It treats capability as the constraint and spends its time on prompts and model choice. Every team in that survey already has a model that can write a competent email. The constraint is whether they can prove to a security team that the agent will do it safely, every time, inside the rules. The teams that solve for that do not get a better model. They get permission, and permission is the only thing standing between an agent and your buyer.
The MCP Problem Is Bigger Than the Agent
Docker spends real time on the Model Context Protocol, the emerging standard for connecting an agent to the outside world, meaning your CRM, your data, and your tools. Eighty-five percent of teams know it. Two thirds use it. And it is not ready. Teams report 42% operational overhead, 41% security and compliance concerns, and 41% struggling with installation and configuration. The security fears are specific, and 46% named them: indirect prompt injection, tool poisoning, and rug pulls.
Why should a go-to-market leader care about a protocol? Because MCP is how an agent reaches buyer data, a content library, and the outreach stack. Every one of those connectors is a place where an agent can be turned against you. The report’s own framing is the one worth stealing. MCP is “a promising foundation that’s not yet secure and enterprise-ready,” and it has to mature in three places before it scales. Discovery, because 44% of organizations cannot find trustworthy MCP servers. Manageability, because teams cannot afford configuration that adds overhead. And security and governance, secure-by-default, with visibility, auditability, and policy enforcement. Hold onto that last one. That is the whole game.
Agents are already good enough to do the work. They are not yet trusted enough to do the work that touches a buyer. The winner will not be the team with the best model. It will be the team that builds the trust layer first.
The Three Gates Before an Agent Touches a Customer
If you want an agent working on revenue, the report hands you the sequence. Three gates, in order.
1. Put it on a boring, proven substrate
Ninety-four percent of organizations containerize their agents, and 98% use the same cloud-native workflows they already use for traditional software. That is the least glamorous finding in the report and one of the most important. The teams that will scale agents treat them like normal software rather than magic. An agent that lives in a chat window with no deployment story is not going to a customer.
2. Codify security as architecture, not a checklist
This is Docker’s own recommendation, and it matches the pattern behind the numbers. Sandboxing, credential handling, and policy enforcement have to be built into how the agent runs, not bolted on at the end in a review. The teams stuck at 60% adoption point inward because their governance is a document. The teams that break through build governance into the runtime, so the safe path is the default path. Only 2% of organizations run a single model, and 61% combine cloud-hosted and local models, which means the average agent is already juggling between four and six models. Governance cannot be a slide in that world.
3. Make the agent portable and inspectable
There is no standard yet for packaging or sharing agents, so every team is inventing its own. Portability is what lets you move an agent from an internal sandbox to a customer-facing process without rebuilding trust from zero. Signed, inspectable artifacts mean a security reviewer can see what the agent does before it does it. That is the road from internal productivity to revenue.
What I Would Do First
I would start with one internal agent and make it boringly governed, then earn the trust before letting it anywhere near a prospect. I would pick the highest-volume, lowest-risk task the team does by hand, run it in a sandbox with a full audit trail, and move it one step closer to the buyer. The step I care about most is research and preparation: read the signal, assemble the context, draft the first line, and leave a rep in the loop for the moment of contact. That is a customer-adjacent agent with a blast radius I can live with.
Then I would measure one number: how many meetings the agent helped book, and how many of those were with real ICP buyers. If an agent cannot move that number, it is an expensive toy no matter how well it is containerized.
Docker calls this a decade-long transformation rather than a year of the agents. That is the most useful line in the document. The businesses that win the next two years will not be the ones that adopted agents fastest. They will be the ones that built enough trust to point them at the parts of the business that make money. The substrate is ready, and 79% of teams are already running agents across two or more environments. The question I keep asking is whether the governance is.
Agents are in production at 60% of companies, and almost all of them are pointed away from customers. That is a trust gap, not a capability gap. Build the governance, the security, and the portability now, and you get to be the team that puts an agent in front of a buyer on purpose instead of by accident.
Source: Docker, “The State of Agentic AI Report” (2026). Survey of 805 technical and IT decision-makers across North America, Europe, and Asia-Pacific. Note: Docker intentionally sampled technical practitioners, so these figures reflect the leading edge of adoption, not the broader market average.
