There's a phrase that's stuck with me since a conversation with Martin Tombs, VP Global Go-to-Market for Analytics and Field CTO EMEA at Qlik, back in February: "boring is brilliant."
He used it to describe the unglamorous but essential work of data governance. I deployed it back at him in the context of observability log files and we laughed at the irony. But given the state of enterprise AI right now, it's also genuinely useful shorthand for where the industry needs to go – and what separates the vendors building for production from the ones still selling demos.
Tombs was speaking ahead of Qlik's announcement of the general availability of its agentic experience in Qlik Cloud, delivered through Qlik Answers as a unified conversational interface, alongside the GA of its Model Context Protocol (MCP) server. This week at Qlik Connect, a new ServiceNow partnership completes the picture. Taken together, they form a fuller architecture story – one worth unpacking carefully, because the market is drowning in agentic announcements that don't survive contact with production environments.
The unstructured data problem is still the problem
Ask Tombs where enterprise AI deployments actually fail and his answer is that it's not the model – not the interface. He explains:
Getting your unstructured data right – I think everyone's still getting their heads around this. It's not just where you store a PDF. It's what's in the PDF, who's responsible for that content in the PDF.
This tracks against what diginomica has been hearing across organizations ranging from Fortune 100 to Fortune 500. And if you're building agentic systems that need to reason across both, the governance challenge compounds. You cannot bolt it on afterward, as Tombs puts it, because the agent has to make decisions about what to trust, what to surface, and – critically – what to decline to answer.
One of the more honest design decisions Qlik has built into Answers is a hard boundary: ask it something outside its governed dataset and it will not hallucinate a response. It tells you it doesn't know. Tombs illustrates this with a deliberately absurd example – training a finance instance and then asking how to peel a banana – but the principle lands in production environments where a confidently wrong answer is categorically worse than silence. Tombs explains:
If I give you three wrong answers, you're going to be out very quickly in asking me questions. And that's really how I see the adoption of any vendor's product.
Gartner's hype cycle framing, which Tombs refers to, describes vendors still climbing the peak while enterprise consumers have started the descent into disillusionment. The gap between what AI delivers in a controlled demo and what it reliably does in a messy production environment remains significant – and it follows that this is a useful lens for evaluating everything Qlik has announced.
What the MCP launch actually does – and why the bouncer matters
Qlik's Model Context Protocol (MCP) server deserves a little more scrutiny. Originally developed by Anthropic, MCP provides a standardized way for AI assistants to discover and invoke external tools and data sources. Qlik's implementation exposes its analytics engine, tools, and governed data products to third-party AI assistants including Claude.
Tombs uses a door analogy that earns its keep. If Qlik's internal Answers capability is the front door of the house, MCP is the side door – the one you open to let external agents access what you've built. But you need a bouncer on that door. He elaborates:
By opening this front door, you've always got to have a bouncer on the door that says, 'What are you coming in for? What are you doing?' You've got to hand a menu to the MCP of what you are capable of doing, what your uniqueness is – and then other things can take advantage of that.
The governance layer has to come before the MCP exposure. The practical implication is that an organization that has done its data governance homework can make that trusted intelligence available to whatever AI assistant its teams use – without re-exposing raw data or bypassing established controls.
The distinction from a conventional API is also worth making explicit for those new to MCP: it standardizes not just the call but the capability discovery, so an external agent understands what a tool does before deciding whether to invoke it. That matters for multi-agent orchestration, where agents select tools dynamically rather than following hard-coded instructions.
Discovery Agent – proactive risk, not reactive reporting
The capability Tombs is most visibly animated about is Discovery Agent – Qlik's continuously monitoring agent that surfaces anomalies, shifts, and emerging risks in key measures, without a human having to go hunting for them. He notes:
We can proactively identify anomalies, trends, and risk. We could tell decision-makers that without me finding it all for them.
Qlik CEO Mike Capone is pointed about what enterprise boards are actually wrestling with right now: they are navigating geo-political volatility, tightening AI regulation, and relentless cost pressure – all of which changes what enterprise AI has to be: auditable, governable, and capable of acting inside real workflows. Discovery Agent is the operational result of that positioning.
The counterpoint is that automated anomaly detection is only as reliable as the model's understanding of what "normal" looks like for a given business – a data quality and contextual calibration problem that no GA announcement resolves. Tombs is candid about this: Qlik will get some things right and some things wrong in deployment, and the product is in constant iteration.
ServiceNow – closing the loop from insight to action
The Qlik Connect announcement adds a dimension that contextualizes the rest. The new ServiceNow partnership routes Qlik analytics into ServiceNow workflows and agents, while adding Qlik metadata collectors to the ServiceNow Data Catalog for discovery and lineage visibility.
An organization can have the best governed analytics layer available, but if the insight never reaches the person or agent making the operational decision, it's academic. ServiceNow is where a significant volume of enterprise work execution happens – and getting Qlik's analytics engine, aggregating cross-system context from ERP, CRM, supply chain, billing, and support data, feeding into that environment is a substantive architectural move rather than a badge-swap partnership.
Pramod Mahadevan, VP of Data and Analytics Product Ecosystem at ServiceNow, says:
The decisions people and agents make every day are only as good as the data behind them.
That's been true for thirty years. What's changed is the plumbing: the data layer, analytics layer, and workflow layer are now connectable in ways that no longer require custom integration work at every junction. The metadata collector piece reinforces this – lineage, discovery, and structure visibility for Qlik-managed assets become accessible from within ServiceNow's own governance tooling, which is a more practically grounded integration than most partnership announcements in this space manage to deliver.
My take
The combination of a governed data product layer, a reasoning engine that knows when not to answer, an MCP interface that extends trusted intelligence to external assistants, continuous monitoring via Discovery Agent, and now a workflow integration with ServiceNow represents a genuine end-to-end architecture – not a feature announcement dressed up as a strategy.
Those honest caveats remain significant, though. Deployment quality varies, unstructured data governance is hard regardless of tooling, and the cost of running agentic systems at enterprise scale is underexamined – Tombs raises cost governance explicitly, and he's right to. These are not problems Qlik's announcements solve; they're problems that good tooling makes more tractable.
The underlying positioning – govern first, agent second, trust as infrastructure – is correct. The organizations that internalize it will build AI systems that actually work in production. The ones that skip it will keep discovering, at significant expense, why promises don't match up to reality. 'Boring is brilliant' needs to be read as an engineering principle, not a marketing line.