Scaling agentic AI without losing control: why guardrails become an advantage
- Augusto Saúde
- Jul 28
- 2 min read
Most companies want agents in production. Only 11% got there. The bottleneck isn't the model — it's control.
There's a number that sums up the state of agentic AI in 2026: only 11% of companies have put agents into production (Deloitte, State of AI 2026), and Gartner estimates that 89% of pilots never scale. Almost everyone wants it; almost no one runs it. And the bottleneck isn't the AI model — which keeps getting better and cheaper. It's control: data, governance and security.
Deloitte itself coined the phrase that defines the moment: "AI agents are scaling faster than their guardrails." In the same study, only 21% of companies say they have a mature governance model to manage agent risk — roughly 80% are moving ahead without the brakes in place.
It makes sense. A copilot that suggests errs low: the human filters. An agent that acts errs high: it executes. Putting a system that perceives, decides and acts inside a financial institution — with a license, suitability, an audit trail and the regulator watching — demands a rigor most AI solutions weren't built for.
This is exactly where, for me, the conversation flips. A guardrail isn't a brake. It's what lets you accelerate.
In practice, institution-grade guardrails mean four things:
Granular permissions — the agent only does what it's authorized to.
Audit trail — every action logged: who, what, when.
Data under your license — isolated per institution.
Human-in-the-loop — human approval points where risk and regulation require it.
The brokerage that adopts with these controls scales. The one that improvises stalls out of fear — or gets burned in production. McKinsey frames the shift well: leading companies stop treating compliance as a cost and start treating trust as an enabler of scale — a "new era of trusted AI." And trust, in finance, is infrastructure, not marketing.
That's the difference Lumes brings: real agentic AI — agents that execute — inside a robust package, with the security guardrails few deliver, and under the brokerage's own license. It's not software with AI; it's the operation run by agents, with control.
If you lead a brokerage and want to enter the agentic era without stepping into the dark, Lumes runs a 20-minute diagnostic of your operation — which agents come in, where, and with which controls.
→ 20-min diagnostic: the agentic layer, with guardrails, applied to your brokerage. lumes.io/broker
Informational material. Lumes provides technology; investment products are offered by the contracting institution, under its license and responsibility (CVM). Not an offer, recommendation, or promise of results.
Comments