.avif){kind=logo}
1. Governance is the most important thing nobody wants to talk about
Governance kept coming up as the foundational condition for everything else on the agenda. AI readiness, data products, semantic layers, agent reliability — every one of them collapses without governed data underneath. And yet, governance still does not have a seat at the main stage. It arrives as a footnote. A prerequisite. The thing you are supposed to have already sorted out.
The honest observation from the room was this: governance is not unsexy because it is unimportant. It is unsexy because it is hard, it is invisible when it works, and it requires convincing people to invest in infrastructure whose value only becomes visible when something goes wrong.
The framing that resonated most was rebranding governance as AI readiness infrastructure. You are not building a governance programme. You are building the control plane that makes your AI decisions trustworthy. That is a conversation the business is ready to have in a way it was not three years ago.
If you want to go deeper on what enforcing data governance actually looks like in practice, we have covered it.
2. Dashboards are dying. Again. But not entirely.
The dashboard death narrative had a strong showing in London. The argument goes: why build a static view of data when you can query it live, in natural language, and get exactly the answer you need at the moment you need it? Conversational data access, powered by LLMs sitting on top of well-structured data, is genuinely changing how some teams work.
We think this is right, and it is also incomplete.
Dashboards will survive in the places where they were always strongest: executive reporting, operational monitoring, regulatory submission. Where they will shrink is everywhere else — the one-off queries, the ad hoc analysis, the "can you pull a number for this" requests that consumed data team capacity. That work is moving to natural language interfaces. The dashboard that remains will be more intentional, more curated, and more important — not less.
The implication is not that you should stop building dashboards. It is that the data feeding them needs to be more reliably governed than ever, because when it is the only view that matters, a broken number in an exec dashboard is a much bigger problem. We wrote about protecting your executive dashboard for exactly this reason.
3. Semantic layers are no longer optional
Every product at the summit had a semantic layer story. Every session touched on the importance of shared definitions. The word "semantic" was used so frequently it started to lose its meaning, which is either ironic or appropriate depending on your mood.
The underlying point is real and important. As AI systems make more decisions on behalf of your organisation, the gap between what the data says and what it means becomes existential. A model that does not know the difference between "revenue" as defined by finance and "revenue" as computed by the sales system is not a useful model. It is a confident wrong answer delivered at scale.
Our honest advice from what we heard: do not get seduced by which product has the most elegant semantic layer implementation. Get clear on what your organisation actually means by the terms it uses, document it in one place, and make that documentation the source of truth for every system that touches it. The technology is a vessel. The thinking is the work.
And the thinking only holds if the data feeding those semantic definitions is clean and trusted. That is where data quality monitoring and a solid data catalog become the foundation rather than the afterthought.
If you want to hear this argument made properly: our Head of Product Laura gave a keynote talk at the summit on exactly this — the gap between a data stack that understands syntax and one that understands semantics. The session was one of the most attended of the event. We are making the recording available again shortly. Register to be notified when it goes live →
4. Everything is an agent, and that should make you nervous in a productive way
The agentic AI conversation at Gartner has matured past "what is an agent" into "how do we govern a world where agents call agents." Most major data platforms now have agent capability. Orchestration layers are multiplying. The question is no longer whether your stack will include autonomous systems acting on your behalf. It is whether the data those systems act on is trustworthy enough to let them.
The concern that surfaced most in the hallways was not about AI safety in the abstract. It was operational and concrete: if an agent calls another agent to make a pricing decision, or trigger a workflow, or generate a disclosure, and the data feeding that chain is stale, or wrong, or ungoverned, the blast radius is not a single bad report. It is a cascade of automated decisions built on a broken foundation.
AI-ready reliability is not a marketing phrase. It describes the specific infrastructure problem of making sure the inputs to autonomous systems are as governed as the decisions those systems make. We built our AI agents — Sentinel, Sage, and Forge — precisely because observability in an agentic world has to be agentic too.
5. Costs are going to rise, and AI is only half the reason
There was a lot of honest conversation in London about the economics of AI adoption. The headline story is familiar: inference costs, GPU costs, the unit economics of running LLMs at scale. But the more interesting conversation was about second-order cost pressure.
More data is being stored because AI systems need more context. More processing is being triggered because LLMs need their inputs freshly computed. Data warehouse bills are going up not just because teams are running more queries, but because the queries themselves are getting more expensive to feed. Storage and compute costs, long on a downward curve, are bending back upward.
The implication for data teams is that cost governance and data quality governance are now the same conversation. Bad data that gets processed through an LLM pipeline does not just produce a wrong answer. It produces an expensive wrong answer. The cost of ungoverned data is no longer just a trust problem. It is a line item.
What all five have in common
Read them together and the pattern is clear. Governance, semantic consistency, agent reliability, dashboard trust, cost discipline — they are all downstream of the same question: is the data your systems are acting on actually trustworthy?
That is not a new question. What is new is the consequence of getting it wrong. When decisions were made by people looking at dashboards, a data quality problem produced a bad slide in a meeting. When decisions are made by agents calling agents at machine speed, the same data quality problem produces something much harder to recover from.
The control plane for your data is not a nice-to-have in that world. It is the infrastructure everything else runs on.
Where do you actually stand?
We built a short assessment — five dimensions, ten questions, three minutes — that scores your organisation's readiness across exactly the themes this summit surfaced: data quality, lineage, incident management, AI governance, and infrastructure maturity.
You get a personalised readiness score, a per-dimension breakdown, and a narrative that tells you where your exposure is highest relative to your regulatory context and the direction the market is moving.
Take the AI Readiness assessment →
No slides. No sales call required to see your results. Just an honest picture of where you are.
Sifflet is the control plane for Data and AI — continuous monitoring, end-to-end lineage, and governed inputs for every system that acts on your behalf. See how it works →
-p-500.png)
