The rollout of autonomous AI agents poses major opportunities to organizations.
However, without the right foundations and approach, the risks are vast when there is no way to guarantee that agents will make correct, reliable decisions.
Article continues below
Global AI Practice Leader at Informatica from Salesforce.
When data quality is poor, an agent will make bad decisions, and there is no human around to identify the mistake and correct it.
The good news is that there are data strategies that can mitigate the risks.
Agents of chaos
In a traditional Machine Learning (ML) system using automation, data problems can reduce accuracy. But in an agentic ecosystem, the actions of one agent can have catastrophic downstream impacts.
At worst, a rogue agent could cause a data cascade in which one error sparks a chain reaction of flawed outputs that get stored, treated as truth, and then reused.
In large organizations, these failures don’t look dramatic at first. Everything seems normal.
Until someone realizes the system has been acting on the wrong frame for weeks. By then, it’s too late and other agents are acting out based on the mistakes made upstream.
A lie whispered into the system becomes a command shouted out of the other side.
Data quality is necessary, but not sufficient. Agents need decision quality data to mitigate the risk of poor outcomes
Data vs reality
Given my role as an AI data leader working with global enterprises, I was reflecting on a key data problem during a recent trip to London. I stopped to eat a famous British dish and when I asked for some chips, I was confident I wouldn’t be served a plateful of silicon.
The data I had gathered about my own personal situation and the context – sitting in a restaurant in the UK capital – made me confident the waiter would bring me some cooked potatoes served alongside fish, not a handful of CPUs. Analysis of local data told me what kind of food I should expect. If I had made the wrong judgment and gathered or interpreted data incorrectly, it would have been a disappointing dinner.
A New Zealand supermarket recently provided another excellent – and hilarious – illustration of the challenge of interpreting data without all the right pieces in place. It created an AI recipe builder to help customers use up their leftovers, inviting people to type in the ingredients they have available and have the bot generate recipes.
Inevitably, people started asking it to make dishes with bleach, ant killer and other dangerous ingredients, so the AI began generating less-than-lovely-sounding suggestions like glue sandwiches and French toast flavored with a soupçon of methanol.
The ingredients were complete and the instructions were correct, but the AI only understood the structure of a recipe – not the purpose or intent of using ingredients to make nourishing, rather than noxious, food.
Now imagine if that AI had been agentic and tasked with, for example, instructing a food assembly plant to generate recipes and ingredient boxes to be sent to customers. The story above was funny – the nightmare scenario of poison being mailed to customers is anything but.
A source of truth for agents
That tale is a clear lesson. If organizations haven’t established a trusted foundation for their data, the jump to agentic AI is extremely risky. This doesn’t magically appear when agents are deployed – it comes from governance, metadata, lineage and understanding not just what the data says, but where it came from and why it exists.
And what should that data look like? It needs to be authoritative and trustworthy, as well as comprehensive and up-to-date to inform timely, complete decisions. It must also be responsible – meaning agents can safely act on it – and secure to prevent misuse.
Finally, context must be considered at every stage. It is the secret sauce that ties all these aspects of good data management together.
Big decisions on data
As they roll out autonomous AI tools, organizations are discovering that accuracy alone is not enough. Agentic systems don’t just predict – they act – and those actions compound. That means missing context is far more dangerous than it was in earlier generations of AI.
The organizations that succeed will be those that treat governance, metadata, and lineage not as an annoying requirement from their compliance teams, but as a strong foundation for their agents. When data drifts away from the truth, that movement doesn’t stop – it gets worse. Building a solid “truth layer” will help to stop that and prevent systems from failing down the line.
Many IT management teams won’t know where this layer is. Maybe they will point vaguely to a warehouse. But they need to know where the truth their agents rely on can be found, because if it’s not in place upstream, then everything else is at risk.
Right now, this appears like a relatively small issue. But as agents take on more and more critical roles, it’s going to become a big one – so getting the basics right today is not just a prudent decision, but an unavoidable hedge against future chaos.
https://cdn.mos.cms.futurecdn.net/2UMvPDp3snEwaGbRuCivjE-970-80.jpg
Source link
