For two years the conversation about AI agents has been driven by demos. An agent that answers email, another that builds a report, another that “works on its own”. If you run an operation, the question that matters is different: what happens when the agent leaves the pilot and joins the day-to-day of the business?
Dynatrace asked the people who are already there. Its report The pulse of agentic AI in 2026 surveyed 919 leaders at large enterprises, and the picture is less about technology and more about trust, rules and visibility. Here are the numbers that matter most and what we take from them.
What the survey measured
Respondents were VPs and directors of engineering, platform engineering leaders, executives who own AI investment and risk, automation leads and SRE and observability teams, in financial services, retail, government, utilities, transportation and software. The sample spans the US, Canada, Mexico, Brazil, the UK, Germany, France and the rest of Europe, the Gulf, Japan, India and Australia.
A fair warning: Dynatrace sells observability, and the report’s conclusion leans that way. The numbers still speak for themselves, and they match what we see inside companies.
Agents have left the lab
Half of enterprises already run agents in production for specific use cases, 44% have broad adoption in some departments and 23% have integrated agents across the whole company. Nearly three in four (72%) are running 2 to 10 projects at once.
The money follows. Most current budgets sit between $2M and $5M, and 74% expect to spend more next year, with nearly half planning another $2M to $5M.
IT leads because the work there is the most repetitive and the easiest to measure. But the fastest-growing use cases over the next few years face the customer, and legal, the least common today, is expected to triple. Read the list as a snapshot, not a boundary: if a task repeats, it can be automated, in any function.
The model is not what blocks you
Ask what stops a project from reaching production and the answers barely mention model quality. They are about control.
Drill into the second barrier and you find the most useful line in the report: 45% cannot set clear rules for when an agent may act on its own and when it needs human approval. Right behind it: limited visibility into how agents make decisions (41%), trouble tracking cost and performance with every new version (40%) and difficulty tracing the knock-on effect of a mistake in other systems (36%).
The bottleneck is design, not technology. Before you switch an agent on, someone has to write the rule: what it does alone, what it only suggests and what it never does without a person.
The criteria for approving production follow the same line: security and data privacy (59%) and output accuracy (55%) first, monitoring and controls over model behavior (44%) right after. Depending on where you operate, that means GDPR, CCPA and the sector rules on top.
Humans in the loop is a choice, not a lag
69% of agent decisions are still verified by a person. And nobody treats that as a problem to remove. Only 13% of companies build fully autonomous agents exclusively; 64% build a mix of autonomous and supervised agents, and 23% build supervised agents only.
Even the future targets keep people in. In IT and routine support, companies aim for roughly a 50/50 split between AI and humans; in business functions such as finance, HR, legal and sales, closer to 60/40 with more human weight.
The split is simple: AI executes the repeatable and the time-sensitive, people set the goal, draw the boundaries and own the outcome. It is the same design we argue for in our piece on human-in-the-loop: human review at the right points, not everywhere.
How companies measure (and where they still do it by hand)
Agent success is measured first by technical performance (60%), operational efficiency (56%) and developer efficiency (54%). Business outcomes come later (48%). Among ROI metrics, productivity (45%) and customer satisfaction (43%) lead, while daily usage comes last (23%).
The number that stands out is a different one: 44% still review agent-to-agent communication by hand. When one agent hands work to another, someone reads the transcript to check it went fine. That works with three agents. It does not work with thirty.
And 30% say they struggle to tie what an agent did to a business result. Measuring model latency is easy; saying how much it cut rework, cycle time or cost is not. That is why we insist on agreeing the metrics before the work starts, not after.
The control plane
The report’s central argument is that observability, a real-time record of what every part of the system is doing, has to stop being just a dashboard and become a control plane: the layer that feeds agents reliable facts to decide on, enforces limits that do not depend on the model’s mood and records every action so someone can audit it later.
In practice that means seeing seven layers at once: the business outcome, the user experience, orchestration, agent-to-agent traffic, the model, data retrieval (RAG) and infrastructure. Observability is already used by 69% during implementation, 57% in operations and 54% in development, but only 40% track agent-to-agent communication and MCP traffic.
For teams getting started, the report recommends a staged path, and it is the part worth taking into your next planning meeting:
- Start with preventive operations and workflows where the agent recommends and a person decides.
- Supervise: with a human in the loop, widen autonomy step by step as data, systems and signals become reliable.
- Graduate to full autonomy only where you can verify, in real time, what each component is doing.
For what to log and track, see our guide to AI agent observability and our piece on taking an AI pilot to production.
What this changes in practice
The survey looks at large enterprises, but the lesson holds at any size: an agent that works in a demo needs three things to become part of the operation.
- Written autonomy rules. For every step: the agent acts, the agent suggests, or the agent stops and asks. It is the number one barrier in the survey, and it is solved in the design, before any code.
- A record of everything. Every decision, every tool call, every handoff between agents, in one place where someone can look and audit.
- Metrics agreed up front. Errors, rework, cycle time and cost, measured before and after. Without that, the ROI conversation is just opinion.
That is the work we do with clients. Our AI Analyst maps the routine, finds what repeats and writes the rules and the metrics. Our Forward Deployed Engineer puts agents inside the systems the company already runs, with logging and human approval at the right points. And side-by-side training takes care of the people who will supervise them.
Agents are already in production. The companies that write the rules, record every step and measure the outcome are the ones that can safely give them more room.
Enjoyed this article?
Get the next one before it hits the blog.
A short email per issue with what actually matters about AI agents at work. No spam.