Generative AI
Inside AI/works™: An agentic development platform
Disclaimer: AI-generated summaries may contain errors, omissions, or misinterpretations. For the full context please read the content below.
Autonomous AI is moving from experiments into enterprise operations, creating new opportunities for speed, scale and smarter decision-making. The real winners will be the companies that build the right architecture, governance and data foundations before AI agents start acting on their behalf.
Autonomous AI is starting to look less like a helpful chatbot and more like a new class of digital worker, one that can plan, use tools, make decisions and take action across business systems.
That is a huge opportunity. It is also a very good reason for leaders to pause before letting AI agents loose inside the enterprise.
For the past few years, most organizations have used AI as an assistant. A person asks a question, the system suggests an answer and the human decides what to do next. Agentic and autonomous AI moves us into different territory. These systems can orchestrate work across multiple steps, pull in context, interact with tools and, in some cases, execute tasks with limited or no human involvement.
In a recent conversation with Shayan Mohanty, Chief Data and AI Officer at Thoughtworks, we explored what this shift means for enterprise leaders. His message was clear. The technology is developing quickly, but the real challenge is organizational readiness.
As Mohanty put it, when it comes to fully autonomous AI, “from a technical perspective, the technology already exists.” The harder part is everything around it: governance, data, architecture, accountability and the operating model needed to make it safe and useful.
Watch the full conversation
Hear Bernard Marr and Shayan Mohanty discuss what it really takes to
prepare enterprises for autonomous AI.
The risk is real, but often misunderstood
Much of the anxiety around autonomous AI comes from high-profile examples of agents doing things they should never have been allowed to do, such as deleting the entire production data at PocketOS. These stories are alarming, but Mohanty argues that leaders need the right mental model.
He said, “the agent didn't decide to wipe the database.” Under the hood, these systems are still driven by models predicting outputs against an objective. When those outputs become commands, and when the system has excessive permissions, things can go badly wrong.
His point was important: “This isn't a runaway AI story, it's a missing enforcement story.”
That distinction moves the conversation away from fear and toward better design. Leaders do not need to avoid autonomous AI, they need to build proper enforcement around it.
A well-designed enterprise AI system should know the difference between reading data, writing data, changing records and deleting critical assets. It should have pre-action checks, permission boundaries, human approvals where needed, audit trails and observability. Asking the AI nicely in a prompt to avoid dangerous actions is not governance. It is wishful thinking with a user interface.
This is where many organizations need to mature quickly. As AI systems become more capable, governance has to move from policy documents into the architecture itself.
Governance must be built in from the start
One of Mohanty’s clearest warnings was that governance cannot be treated as an afterthought.
“We don't believe you can bolt it on post-hoc,” he told me.
That is a key lesson for every enterprise now experimenting with agents. A single pilot can feel manageable. One team builds one workflow for one problem. Then the pilot works. Other teams want their own agents. Before long, the organization has dozens of agents running across functions, each with different assumptions, permissions and connections.
At that point, the problem becomes much harder. Who owns each agent? What data can it access? Can it create sub-agents? Does it act on behalf of a human, or does it have its own identity? What happens when the employee who created it leaves the company?
These questions shape how accountability works in practice, especially when AI systems begin making decisions and taking action across the enterprise.
Mohanty explained that governance should be part of the “original DNA of the runtime.” In plain English, that means the operating environment for agents needs built-in controls for identity, permissions, observability, cost management and human escalation.
The enterprise challenge is rarely a lack of ambition. Most leaders can see the promise, but the real gap is between ambition and implementation. Agentic AI only creates real value when it is connected to business systems, governed responsibly and designed to keep improving as models, regulations, threats and business needs change.
The best use cases redesign work
A common mistake is to use AI to automate old processes without asking whether those processes should be redesigned.
Mohanty shared examples of some recent work Thoughtworks did in the fields of life sciences, financial services and embodied AI. In life sciences, agentic systems can help researchers review literature, synthesize findings and accelerate the early stages of drug discovery. In financial services, agents can become more intelligent interfaces between companies and clients, able to reason across services and take useful action. In robotics, autonomous AI raises even deeper questions because software decisions can have physical consequences.
The common thread is that the most valuable use cases are not simple task automation. They change the economics of work.
In a market increasingly cluttered with endless pilots, enterprises need a practical route from a promising idea to a production system. Many companies are busy proving that AI can do impressive things in controlled settings. Far fewer are building the foundations that allow AI to operate safely and repeatedly in the real business.
Competitive advantage will come from execution
When I asked Mohanty whether advantage will come from better models, better workflows or stronger data ecosystems, he answered directly: “It's orchestration. It's tooling.”
That rings true. Foundation models will continue to improve, but access to powerful models is becoming more widely available. The real differentiator is how well companies connect those models to their data, workflows, systems and people.
In other words, the moat is execution.
Mohanty described this as the ability to move quickly as conditions change. Technical advantages may be short-lived. Organizational capability lasts longer. Success with autonomous AI depends on how quickly an organization can integrate the technology, govern it properly and adapt as models, workflows and business needs change.
That requires AI-ready data with provenance. It requires a runtime that is independent enough to work across changing model and platform ecosystems. It requires leaders who understand AI as a business transformation rather than a software feature.
It also requires a new human role. As Mohanty put it, humans “go up the abstraction chain.” Instead of spending time on every small approval, people increasingly define the rules, constraints and objectives that guide autonomous systems, while stepping in when context, risk or judgment call for human oversight.
Stop funding demos and build the substrate
Perhaps Mohanty’s most practical advice was aimed at CEOs.
“If you're thinking about AI as an experiment within your organization, you're already doing it wrong,” he said.
That does not mean experimentation has no value. It means experiments should lead toward a scalable architecture. Too many organizations are funding demos instead of building the substrate, the data, governance, tooling and operating model that allow AI to move into production.
This is where enterprise leaders need to be honest. Autonomous AI will not deliver meaningful value if it remains trapped in isolated proof-of-concept projects. It needs to become part of how the business builds software, serves customers, manages operations and makes decisions.
Autonomous AI is here. The winners will be the organizations that build the right foundations now: governed architecture, AI-ready data, scalable engineering practices and leaders prepared to rethink work around intelligent systems.
If you're looking to explore these ideas further, The agentic enterprise: Building an ecosystem of continuous evolution and reliable impact, developed by Thoughtworks and AWS, takes a deeper look at how organizations can move from isolated AI pilots to governed, scalable agentic systems.
And because autonomous AI is only as effective as the data that powers it, our new playbook, How data platforms and governance must evolve for an agent-driven world with Databricks explores how to build modern, AI-ready data platforms that provide the trusted foundation for the agentic era.
Together, they offer practical guidance to help organizations rewire for agentic and autonomous AI - building the architecture, governance and data foundations needed to deliver AI that works.
