A survey of 306 engineers shows that the AI agents that work in production have less autonomy, more control and more custom engineering than the market imagines.
During the last two years, the technology industry has been going through the phase in which agentic AI software is moving beyond proofs of concept (POCs) that overpromise and turning into workflows for real and complex scenarios. High failure rates and frustration accompanied this shift, as already evidenced by the extensively cited MIT State of AI in Business in 2025 study.
On the executive side, there is pressure for returns on high investment and great expectations for end-to-end process automation. On the technical side, a growing view is that AI agent software that goes to production does not fail due to technology limitations, but from the lack of engineering needed to combine language models with deterministic processes, proper measurement of results, and implementation of methods for agent evolution and learning.
In the first experiments with the technology in industry, the modus operandi was to treat this software as black boxes, expecting results to be generated magically, with quality and accuracy. The bet was on fully autonomous AI systems, with language models given total control over the process. At first glance, solutions of this type are tempting for their ease and speed of implementation. However, this method underestimates something intrinsic to the nature of these models: hallucinations.
What the research reveals about AI agents for business
To understand what separates projects that work from those that fail, a study conducted by researchers from Berkeley, Stanford and IBM offers one of the most concrete answers to date.
Measuring Agents in Production (MAP) is a systematic survey on production-deployed AI agents. Differentiating itself from previous studies such as MIT NANDA (addressed in an article by Danilo Toledo, co-founder of Taqtile), MAP stands out not only for the volume of cases included, but also for restricting itself to projects that went to production and for going to the people building and running these systems, rather than executives.
From a questionnaire with four questions, answered by 306 engineers who implement and operate agents on a daily basis, distributed across 26 market niches, the study gathered evidence on which practices are most effective in developing agent-based systems. The questionnaire included the following questions:
- What are the applications, users and requirements of the agents?
- What models, architectures and techniques are used to build agents that go to production?
- How are agents evaluated before deployment (going to production)?
- What are the main challenges in building agents deployed in production?
The main findings
The most prominent finding from MAP was that successful production agents do not have high autonomy. Being deliberately conservative, they rely on human validation during processes (human-in-the-loop) to ensure result reliability and learning.
Although 52% of projects use the LLM-as-a-judge technique, 74% primarily depend on human evaluation. Among the systems studied, 68% execute at most ten steps before requiring human intervention, with half executing fewer than five.
This is what we see in practice. In the AI agent that runs car subscription sales we built, human supervision existed from the prototype, autonomy grew in stages as the agent proved it was getting things right, and the credit decision stayed with a person.
The study investigated how teams measure the value of delivered products. Most opted for experiments within their own ecosystem, conducting A/B tests, manual feedback and their own business metrics, such as time per task, rework rate, escalation-to-human rate, cost per resolved case and impact on operational indicators (SLA, conversion, losses avoided).
They did not rely on ready-made market benchmarks. This reinforces that, in the current scenario, "right" and "wrong" are operational and business definitions, not academic ones.
These results echo what MIT NANDA (State of AI in Business in 2025) had already shown: most projects fail because they do not learn from feedback, are not properly measured, do not fit into real business flows, and are not reliable because sensitive processes are executed in a non-deterministic way.
Another notable result is that 70% of projects use ready-made models without fine tuning, focusing only on prompt engineering. The study also mapped the low adoption of ready-made agent frameworks: 85% of cases built custom applications to gain control and predictability. It is the same logic as the car subscription agent: price, term and payment conditions come from rules defined with the business, and the language model comes in to converse and qualify. Regarding the balance between output quality and response time, 66% of participants chose to prioritize quality, even if that meant waiting minutes for a response.
The success conditions for AI-driven process automation
The research points to the conditions in which agents have the greatest potential to deliver results in production:
- There are repetitive tasks with high human cost; 73% of respondents developed agents for manual task automation and efficiency gains.
- There is a verifiable "operational truth", even if partial.
- The process accepts human-in-the-loop as a control step.
The study also provided information about the market niches in which systems were deployed, and examples of applications. We included the chart and table below for better consolidation and visualization of the information:
The field is still consolidating, but MAP data on enterprise AI agents shows that the direction is clearer than it seems: less autonomy, more control; less benchmark, more proprietary measurement; less ready-made framework, more custom engineering. Companies that understand this first have a real advantage, not over the technology, but over how to use it.
This article is part of what we are learning as we develop agentic AI software, from the perspective of those who build: experiments, real cases and reflections on AI in the enterprise. If you want to explore how these learnings apply to your context, talk to our team.
Notes
LLM-as-a-judge: use of an LLM to evaluate the system's responses and outputs.
Human-in-the-loop: a flow in which humans validate or take over critical steps (final approval, exceptions), reducing risk and increasing reliability in production.
Prompt engineering: building and iterating the instructions in the prompt and context to guide the model.
Fine tuning: adjusting model weights with proprietary data.
References
- Pan, Melissa Z.; Arabzadeh, Negar; Cogo, Riccardo; et al., Measuring Agents in Production, December 2025.
- The GenAI Divide (MIT NANDA), State of AI in Business 2025 report.
- Alvarez-Teleña, S.; Díez-Fernández, M., Advances in Agentic AI: Back to the Future, December 2025.