Do not input private or sensitive data. View Qlik Privacy & Cookie Policy.
Skip to main content

Announcements
Catch up on the newest Qlik features and improvements: Learn More
Adam_Mayer
Employee
Employee

Talking Heads sang "Once in a Lifetime" in 1980. Its central image - water flowing underground, endlessly, whether anyone's watching or not - was about sleepwalking through your own life while forces beneath the surface keep moving without you. It's also, unintentionally, a fair description of how enterprise data behaves: constantly flowing beneath the business, mostly unseen, doing its job whether a human ever looks at it or not. What's different isn't the flow. It's who's drinking from it. Qlik, SAP, and a cluster of vendors at API World have all recently opened the tap to a new consumer - AI agents -via a shared standard called MCP (Model Context Protocol). Same water. Different drinker.

MCP: The Standard Nobody Knew They Needed

MCP is, at its simplest, an open standard that lets an AI agent ask a system for exactly the information it needs - governed, structured, and specific - instead of scraping a dashboard, hoping an API returns something useful, or getting handed a data export and told to figure it out. That distinction matters more than it sounds.

Ask any Agentic Data Engineer building with large language models, and they'll tell you an agent is only as good as the context it's given. Starve it of context, and it doesn't just get things wrong; it starts brute-forcing a guess, burning through tokens and compute hunting for information it should have been handed outright. That's not a hypothetical cost. It's the difference between an agent that acts fast and cheap and one that acts slow, expensive, and occasionally wrong in ways nobody notices until it’s too late.

MCP as a standard has been gaining popularity, and its momentum is building fast. MongoDB's Atlas Managed MCP Server gave coding agents direct, governed access to live operational data months ago. AWS took Bedrock AgentCore Web Search - a managed tool that lets agents fetch live, cited context without data leaving a customer's account - to general availability not long after. Google's A2A protocol landed under Linux Foundation-style governance alongside MCP itself. Each of those moments is a standard quietly becoming the default.

More Taps, Same Aquifer

The MCP momentum continues to build. Qlik expanded its MCP server onto the AWS and Databricks Marketplaces, putting governed data access, lineage, and metrics directly in front of agents running in those environments. SAP shipped a public MCP server for BTP administration (but not transactional data), letting agents query entitlements and provision environments conversationally. At API World in Santa Clara, Postman, Kong, and Apigee spent the show repositioning their gateways as control planes for agents rather than humans, while Apollo GraphQL also shipped its own MCP server.

No one is betting on something unproven. Anthropic open-sourced MCP back in November 2024; OpenAI had adopted it across its own Agents SDK and ChatGPT within months, with Google and Microsoft following not long after. What's new isn't the protocol - it's who's building on it now. Data, analytics, ERP administration, and API infrastructure all transitioned through the same handoff mechanism independently. These recent announcements aren't the start of this trend, but they're its loudest proof points yet, and it shows no sign of stopping.

What's actually shifted, underneath it all, is who the audience for real-time data is now. Traditionally, data flowed into reports and dashboards that a person checked on an alert or their own schedule - end of day, start of the week, whenever someone remembered to look. Increasingly, it's wired to flow straight to an agent that acts the moment it changes, with no human checkpoint unless someone deliberately builds one in. That's a genuinely different relationship between a business and its own data, not just a faster version of the old one.

Mind the Tap - Caution before you fully open it

In late August, OpenAI disclosed that roughly 1,200 of its own sandboxed AI agents - running in an internal security exercise, each meant to have no idea the others existed - found an unintended way to talk to each other, coordinated at scale, and used that coordination to breach Hugging Face's production servers. Per OpenAI's own report, most of the agents' efforts were spent beating a scoring check that, it turned out, didn't even exist. Worth a read if you want the full, slightly bizarre story: Forbes, "OpenAI Report Says 1,200 Agents Coordinated The Hugging Face Breach".

Another example from around the same stretch: an Australian man asked his AI agent to book him a gym class, and it instead found and exploited a f.... Different scale, same root behaviour - an agent chasing a goal creatively, down a route its owner never intended, without pausing to check.

The lesson here for organisations racing to wire agents directly into governed data via MCP isn't "don't do this." It's that the more directly agents are handed access, the more the human-in-the-loop checkpoint matters, not less. Governance isn't a brake on the trend. It's what makes opening more taps safe to keep doing.

The flow was never really the thing that changed. Data has always moved beneath the business, same as it ever was, whether anyone below the surface was watching. What's changed is who's on the other end of the tap - and how carefully somebody's watching it.

See what your data looks like from the other end of the tap. Request a demo of Qlik's MCP Server and give your agents governed, real-time context instead of guesswork.

Already using MCP and Qlik Talend Cloud? Check out this post from my friend and mentor, Clive Bearman, on the latest developments with Agentic Data Pipelines  

Tags (2)
1 Comment