Autonomy is a funny thing. On a good day, an assistant’s autonomous decisions can save time. On a bad day, a toddler’s autonomy can lead to 500 purchases of Elmo merch.
For Richard Bownes, chief data scientist for Europe at consultancy Valtech, deploying generative AI for clients requires “handbrakes”—processes that slow down the speedy process to ensure an agent’s decisions don’t lead to catastrophe.
Bownes recently helped a public services org—one dealing with people’s finances, he said—to use GenAI to “speed up their case management.”
“It would be inconceivably irresponsible to create a system in which decisions could be made by a public services organization, in the name of the public good, that would make up an answer, that was actually an artifact of hallucination or bias, or just a model not really understanding the question,” Bownes said.
What’s an agent? Bownes’s definition of an agent sounds more like a thermostat, at least on the surface. It is, according to Bownes, “a model which can take an action within its environment, with some level of autonomy.”
Agentic actions today could include the drafting of a meeting agenda, monitoring machinery to predict failures, or providing personalized recommendations. And plenty of vendors, from Microsoft to Salesforce to Google, are getting in on the agentic action.
Them’s the brakes. At the heart of an agent is an LLM, which orchestrates and outputs a step-by-step plan from an input prompt.
To prevent irresponsible consequences of rogue AI agents, Bownes recommends sequestering the outputted data for review before influencing decisions or customers. Once isolated, consider these handbrakes, he said:
- Constrained autonomy: Agent owners should pre-define critical functions that must be shown to a human before executing, Bownes recommended in a follow-up email. In the case of a booked flight (or ordered Elmo, for that matter), for example, an agent should not execute the purchase without first showing you the prices.
- The four eyes principle: To avoid biases in inputs and outputs, engineers should call out to an additional model, Bownes said.
- Modularity: There are fewer opportunities for errors when a task like “book flight” gets broken down into component actions that are their own little agents: One for getting location data, one for finding the nearest airport, one for checking available flights for a desired day, and one for selecting a seat. Breaking the main action down into subtasks will improve workflow execution, Bownes told us.
Top insights for IT pros
From cybersecurity and big data to cloud computing, IT Brew covers the latest trends shaping business tech in our 4x weekly newsletter, virtual events with industry experts, and digital guides.
A 2024 Q4 report from Deloitte, released in January 2025, found that 26% of 2,773 business leaders said their orgs “were already exploring autonomous agent development to a large or very large extent.” Some companies recently shared with Business Insider how they’re using agentic capabilities for tasks like drafting a personalized reminder for invoices or crafting customer-support responses.
“There’s going to be a lot of discovery in what the appropriate balance is between letting the agents do their thing, while having human intervention and oversight,” Chanley Howell, partner and intellectual property lawyer with Foley & Lardner, told us.
When considering the critical importance of adding handbrakes to an AI model, Bownes considered the adage often attributed to a 1970s IBM training manual: A computer can never be held accountable. Therefore, a computer must never make a management decision.
“That’s extra true with autonomous agents because they are, in some ways, enabled to make decisions through their autonomy, because they can call tools which can have an action.”