Why Is AI Governance the Real Bottleneck for Enterprise ROI?

In the previous episode of the Tech Transformed podcast, EM360Tech's Shubhangi Dua sat down with David Hsu, founder and CEO of Retool, to tackle an uncomfortable truth for CIOs: vibe coding is already happening inside your organisation, whether you've sanctioned it or not. 

From customer data turning up in AI-built apps on the public internet, to duplicate tools showing conflicting numbers, to AI bills climbing with no clear view of the return, David made the case that the answer isn't to ban AI-built software but to put guardrails around it. One central platform, with controlled access to trusted data, gives IT visibility, auditability, and control, because "with governance comes freedom." He finished with a bold prediction: by 2030, most of the work in your company, and the CIO's role, will become closer to a chief people officer for AI. 

In this Big Data London special edition, host John Santaferraro picks up where that conversation left off with Ryan Wong, Head of Engineering at Retool. As AI moves from generating apps and content to agents that take real action, what does governing them actually take?

Why is AI Harness Important?

When Santaferraro asks Wong to set straight the confusion around AI harnesses in the market, Wong explains that the harness is the controllable layer around the model that surrounds context, identity, policy, action, and observability. It paves the way for an enterprise to manage and govern AI effectively and efficiently without relying on model improvement alone. 

When thinking about what goes into agentic harness, he says that it starts with context. “Business context, skills, and tools are some of the most important components for your harness or your model, your agent, to actually be successful.” If the components don’t have the same context, Wong rhetorically says, “How is the AI going to act in the way you expect it to?” 

The second imperative feature is identity. “Models are only as valuable as the actions they can take,” he tells Ferraro. These actions are necessary to gain a sense of identity and authority. In case a model is reading from a database, the question is: “By what identity and authority is it undertaking that task? Is that a delegated authority? Is that a service account authority? Is that, in a bad sense, a God mode account that is functionally a backdoor?

The third key feature of agentic harness is policy. Retool’s Head of Engineering says policy in this scenario is about how you apply policy and governance to the model and how it is taking those actions. Retool believes that while it’s important to create policy and manage policy, policy doesn’t necessarily dictate that you’re managing individual action. 

“If what you're doing functionally is every little step that an agent takes needs a human in the loop, then what are you really automating at the end of the day?” Wong voices that a conservative approach can be expanded as the understanding of how effective the model is grows. “We really want you to be able to employ policy properly.”

The fourth aspect is action. Oftentimes, Wong tells Ferraro that ‘action’ is conflated with context. As we think about the various capabilities of a model, action is what an agentic model takes. He advises minimising non-determinism by keeping AI actions strictly scoped and predictable to maintain a clean abstraction process. 

Wong further says to ensure observability. Every action taken by the model must be recorded, auditable, and tied to a specific authority so that it can be properly governed. 

How Should Enterprises Govern AI Harness?

Governing AI harness is tricky because, as Wong puts it, “you can’t govern what you can’t see.” From an observability point of view, understanding and ensuring that everything that’s happening from a model standpoint is recorded, auditable, observable, and extensible. It’s also important to remember to question under what authority that happens, alluding to the identity aspect. 

Wong also suggests bringing all AI tools together into one central hub rather than letting individual teams use separate, disconnected systems. Just as a central front door makes a building easier to secure, a single hub makes it much simpler for companies to control, monitor, and protect their AI usage.

“One of the first forays that most enterprises had with LLMs was to bring in Claude and ChatGPT into their business.”That, Wong says, “is a very distributed way to run these Large Language Models (LLMs) and a much harder way to govern them.

To centralise the harness, the Retool engineering head says, is actually a simple software engineering practice or web application architecture best practice that’s quite analogous. It’s the concept of an API gateway. 

“An API gateway is a very common mechanism to be an abstraction layer on top of a database, such as a relational database.

Just like an API gateway acts as a single, controlled access point between various services and a database to manage security and logic, an AI harness should be centralised. Instead of letting different parts of a business connect AI models directly to company data—which is impossible to govern safely—a centralised harness provides a single, secure layer to manage access, policy, and guardrails.

“In your harness, you can enforce policy as a function of the resource, but enforcing that as a function of the resource is a much more scalable way to think about your governance. Embedding policy enforcement within your agent harness is essential for capturing the user's original intent. It’s a key requirement for effective governance.

Ultimately, focusing on the business harness is the only way enterprises will progress. Wong says, “At the end of the day, that’s the part we do control.” 

“The model is not perfect now; it may not be perfect in six months ot two years, but it will get better,” he adds. “ It doesn’t change the value of what you’re building from a harness standpoint. 

He tells Ferraro that models will continue growing, but the harness will last many years, and that will not change. “You're not going to regret investing in that aspect of your business and that aspect of your architecture. AI harness is the key for you to be flexible and adaptable as long as you are forward-looking enough to see around that corner.”

By centralising governance and aligning AI capabilities with business needs, enterprises can harness AI effectively for real-world applications.

Takeaways

• The harness is the layer around the model that a business controls. 
• AI harnesses context, identity, policy, action, and observability.
• Governance, not model quality, stops agents from emailing or touching sensitive data.
• Employees connecting Claude or ChatGPT to company data is hard to govern, so centralise access like an API gateway.
• Enforce data policy at the resource, not app by app, so employees can vibe code freely.
• Start agents with tight limits, then widen them. Log every action with the identity behind it.
• Vibe coding made building cheap. Deployment, access control, and observability are now the hard part.
• Retool says its platform has long handled those layers.
• Retool’s Assist tool struggled because it used a proprietary language. Building closer to what LLMs know works better.
• Models will change; the harness will last for years.

Chapters

00:00 Introduction to AI harness and its significance

01:08 Defining the components of an AI harness: context, identity, policy, action, observability

05:12 The role of centralization in AI harness architecture

06:44 Governance challenges in multi-agent AI environments

17:48 Implementing a five-step governance process

22:20 Why focusing on the harness is key for AI progress

27:17 The importance of governance as AI agents gain authority

28:24 Final thoughts on controlling autonomous AI agents

Learn more about Retool: https://retool.com