AI agents in production: efficiency, costs and trust
How companies are deploying AI agents, how efficient LLMs are and the risks before going to production.
Today’s most important AI news
The day was dominated by a question more and more teams are seriously asking: how do you go from an AI agent that works in a demo to one that works in production? From concrete enterprise cases to academic research on efficiency and pre-deployment verification, today’s news all points to the same problem. It’s not a day of spectacular launches, but something more valuable: clear signals about what’s working and what’s still unsolved.
Endava redesigns its software delivery with agents — and the lessons are transferable
Endava, a technology consultancy with more than 11,000 employees and a strong presence in Latin America, published with OpenAI the details of how it’s restructuring its software delivery process around AI agents. The company combines ChatGPT Enterprise with Codex to automate workflows that used to require manual coordination between teams: from code generation and review to documentation and the handoff between stages of the development cycle.
What makes this case relevant isn’t the use of well-known tools, but the scale of the organizational change. Endava isn’t talking about “assistants” that help developers, but about redesigning software delivery as a discipline. That implies changes in roles, in performance metrics and in the technical team’s culture. The stated goal is to build what they call an “AI-native culture” at the enterprise level, where agents are part of the standard production flow, not isolated experiments.
For a product manager or CTO in Latin America, the Endava case is a concrete reference for what to expect when scaling beyond the pilot. The practical question that arises is whether your organization has the governance processes needed for an agent to make decisions in production without constant supervision — and that’s exactly the problem the next piece of research tackles.
Verifying agents before deploying them: the missing link in the enterprise chain
A paper published today on ArXiv proposes a framework called “ontology-grounded simulation and trust certification” for verifying AI agents before they reach production. The central argument is straightforward: current control mechanisms — post-deployment monitoring, human-in-the-loop, guardrails in prompts — offer limited guarantees once the agent is already operating. By then, the damage may already be done.
The framework proposes three combined components: an “Agent Operational Envelope” that formally defines what the agent can and can’t do, ontology-based simulation to test behavior in adverse scenarios before launch, and an auditable trust-certification mechanism. In practical terms, it’s the equivalent of the integration tests and QA processes any software team knows, but applied to systems that reason and make decisions in natural language.
For teams building or buying agentic solutions, this work points to a real gap: most organizations have no formal criteria for deciding when an agent is “ready” for production. The cost of ignoring that isn’t only technical — it’s reputational and potentially regulatory, especially in sectors like finance or health where automated decisions have direct consequences.
LLMs spend tokens very unevenly — and that has a real financial cost
OckBench, a benchmark presented on ArXiv that already includes results for GPT-5 and Gemini 3, measures something traditional benchmarks completely ignore: efficiency in token usage. The findings are striking. Two models that solve the same problem with similar accuracy can differ by up to 5x in the number of tokens they consume to reach that answer.
That difference isn’t just academic. For any company running LLMs in production at high volume, a model that uses twice as many tokens as another for the same task means, directly, paying twice as much for each inference. At current API prices, that inefficiency becomes significant at scale. OckBench proposes a metric called “token efficiency” that combines accuracy with token economy, offering a more honest view of the real cost of operating a model.
The practical implication is immediate: if your team is evaluating which model to use for a high-volume use case, adding token efficiency as a selection criterion can have a bigger impact on total operating cost than marginal differences in accuracy. Not every benchmark tells you that — OckBench does.
In a nutshell
Today’s pattern is clear: the industry is leaving the “this is possible” phase and fully entering “how do we make it reliable and sustainable?” Endava shows that organizational transformation around agents is already happening at enterprise scale. Research on pre-deployment verification shows that the tooling to do it well is still being built. And OckBench reveals that even the way we evaluate models has costly blind spots. For those making AI decisions at Latin American companies, the message is that the conversation is no longer about whether to adopt agents, but about what criteria to use — and today there are more concrete elements to answer that question.
Sources used: https://openai.com/index/endava-frontiers, https://arxiv.org/abs/2606.04037, https://arxiv.org/abs/2511.05722