Since 2012, Alation has built a business around its enterprise Data Intelligence platform that provides tools to support metadata management, data governance, data lineage and data quality. A year ago, it acquired Numbers Station which has an agentic AI approach, honed at Stanford, to automate complex data workflows and generate SQL queries from natural language. 

Just over a year of R&D later the combination of Alation’s metadata foundation with Number Station’s agents has led to the launch of AIOS as a system to allow agents to access data and apps that also ensures accuracy and counteracts issues such as drift, essentially a new operating system for enterprise AI. 

An operating system for agents

Understandably enterprises are cautious about deploying production level AI because it involves unleashing non-deterministic software on enterprise data in a dynamic business context. What could possibly go wrong? As enterprises are increasingly aware, a lot could go wrong: the data the agent is acting on could be outdated or incorrect, it could misunderstand the context because the business logic has changed or it has misread the definition. And then it could drift because the instructions, tools or training it has is no longer relevant to the environment it is operating in. 

Nevertheless, your agentic Pollyanna lacks human self-doubt and nor does it generate a software error; it moves confidently to continue to carry out its activities providing incorrect answers that then drive your business decisions. Satyen Sangani, CEO and co-founder of Alation, explains that AIOS is designed to address these failure points by combining data, context and agents into a unified operating system that is open, governed and self-improving. 

Does this mean that you no longer need a human in the loop? Sangani says:

It depends on the significance of the accuracy that the enterprise defines via its policy. An agent codifies a prompt and then the agent may change because the world around it changes, but the key thing is to ensure that the outputs remain accurate. Part of the problem is that enterprises have piece-part technologies to create systems that loosely talk to each other.

In other words, for many use cases a human is still required as the final arbiter but AIOS is a layer working across disparate software as a steadying system on the outputs generated.

So why is Alation pitching the product as an operating system. According to Sangani, it was going to be called an engine or another well-used industry term, but a customer suggested operating system:

The customer said, ‘I want to build data apps and I want AI portability so I can swap things in and out such as Chat GPT, Open Source, DataBricks, Snowflake and so on. I want to use many semantic layers including Tableau and any search engine. I want to swap these components in and out’.

 AIOS achieves this by ensuring all the layers are working together at the metadata level.

When 85% right can be 100% wrong in some domains

Alation has over 500 customers, including Fortune 1000 organizations, and Sangani says that what he sees around agentic AI is that:

People are struggling with the question of trust and accuracy and in many domains, people are being cautious about adoption, because for those domains 85% right is 100% wrong resulting in a lawsuit or lost business. They need certainty.

Alation currently has customers trying out AIOS in sectors such as Supply Chain, Healthcare, Retail, and Financial Services, as Sangani dubs them, “the sectors which are more circumspect or cautious about engaging with agentic AI.”

What then do these enterprises need to make the best use of AIOS? Sangani says:

It is not about technographic capabilities, it is more about the mindset of the organization. Do they have a well-defined, high stakes business problem to solve? It needs to be high stakes to get C-level sponsorship and funding. But they do need an experimental mindset so that they can change and adapt quickly.

AIOS helps you build an agent to which you attach an eval and then build cases with the agent giving a response that a customer thinks is correct. It is the silent, confident failures in the catch that are most interesting when you catch the actions on the other side. Once there is a robust case, we then publish via an MCP client as a series of steps that hook into a live process, or hand off to another agent. This may not guarantee fidelity so you need to observe what is going on over time and decide on an acceptable error rate.

AIOS does not provide observability to see how models think and reason, rather it offers the packaged model and all the output from the context the enterprise data provides, so that organizations can see and manage the output via the eval testing capability. AIOS is best used when you understand the issue you are trying to solve so that it can provide the right context development capability.

Sangani says that what customers need help with at the moment is how to set up, iterate and build agents. While Alation offers guidance and best practices, it also has partners in place such as Capgemini, TCS, Accenture (via its Redkite acquisition) and phData to help with this.

My take

This year, many people have emphasised that agents are not a system of record, and that this is one of the biggest challenges in overcoming a lack of enterprise trust in scaling up agentic AI deployment. Alation has taken the view that agents and AI adoption is a system problem that can be remedied by AIOS as a layer above fragmented tooling that can synchronise agents, contexts, data and governance as the environment changes.  

Creating this operating system for agents is a step towards building more reliable agentic output. It is to be welcomed since every tool that launches and pushes the market towards trusted agentic AI will help settle enterprise nerves about wider deployment of the technology.