Dispatches from the Gartner Data & Analytics Summit 2026: The Noise, the Slop, and the Signal
Amid alarming blogs announcing that AI is coming for our jobs, I landed in Orlando armed with my own questions about how AI was being adopted by data leaders and demonstrating real benefits. Since late 2025, AI models have become so good that they can automate many of our tasks. Coding is solved. Singularity is here. We may as well have achieved artificial general intelligence (AGI). AI models now have capability overhang; a theory that AI capability outstrips humans’ ability to even fathom it.
Gartner analysts Adam Ronthal and Georgia O’Callaghan’s keynote was quite apt. They used the theme of white water rafting and sure enough, the rapids are hitting us from every direction and it perpetually feels like our boats are about to capsize. Adding to the confusion of AI’s ability to perform the most complex tasks is its ability to fail at the most basic understanding of human-centric tasks.
The Urgency
I find myself fortunate to be attending this Summit at this inflection point where AI is finally starting to make an impact. The urgency in the hallways of the Gaylord Palms was palpable. Not panic, but a clear recognition that the rules are being rewritten in real time. I was surprised to see how many conversations turned to building one’s own brand. After having attended this conference for over 10 years, this was the first time such a topic came up repeatedly.
When I read last year’s blog it stands out like a time capsule of what were the most pressing problems. To be honest, those problems sound rather quaint. In this blog, I capture the zeitgeist of the lives of data practitioners as of early 2026. A year of urgency to make AI mainstream and deliver sustained returns. How much difference one year makes! It is like we are living in dog years — what used to be a seven-year cycle is now compressed into less than a year.
My three hot topics before showing up to the event were: context platforms, agents, and governance. However, I had an epiphany that all three topics are driving a resurgence of operational databases. So, let’s start with that.
Everyone Fixated on the Lake. Agents Are Living in the Database
Operational databases are the bread-and-butter systems that run every organization’s business. Their uptime is far more critical than that of analytical databases. Yet, for the last few years we have fixated on data engineering pipelines, Iceberg format, lakehouse, governance, data quality, observability, etc. — to the detriment of operational databases.
However, most uses of agents I see are geared towards detecting anomalies, generating predictions, and automating processes that rely on transactional data. Yes, this data needs to be integrated with the rest of an organization’s enterprise data to drive additional AI-driven use cases, which requires the analytical data stores. But agents are shining a light on where and how data is being sourced.
Before we go deeper into this topic, I want to clarify that agents play two roles in databases. One is where agents are being used to automate the inner DBA-like activities of optimization, monitoring, etc. The second use of agents is where they consume data for the purpose of meeting actual business needs. In this section, I am more interested in exploring the latter business-focused use cases.
In the traditional use cases, data was created once at the source and then moved out to the consumers. But in the new world, agents are constantly creating new data, chain of thought, prompts, etc. This additional data is used by the models for reasoning, planning, and action. It needs to be stored for audit and lineage reasons in real time. This is a write-heavy workload for which operational databases are best suited, compared to read-optimized analytical datastores. The write-heavy workloads will put more pressure on the database.
A real-world example brings this home. OpenAI recently published how it scaled PostgreSQL to support 800 million ChatGPT users using a single primary instance and nearly 50 read replicas, rather than migrating to a distributed database. Their workload is predominantly read-heavy, which is why this architecture held. But they were candid that write pressure is where PostgreSQL struggles, and they moved write-heavy workloads to a different operational database, Azure CosmosDB.
The dominant pattern right now is files: CLAUDE.md, AGENTS.md, markdown memory folders, plain-text context files committed to git. This works well at small scale because LLMs are remarkably tolerant of messy, denormalized representations on the read path. After all, they were trained on the entire messy internet! But files have no concurrency control. When a single agent reads sequentially from a markdown file, there is no problem. When multiple agents write to shared state consisting of context summaries, execution traces, RAG knowledge bases, you enter a regime of arbitrary interleavings. Unlike schema ambiguity, which an LLM can interpolate around, a race condition produces outcomes that are arbitrarily wrong, with no learned prior to fall back on.
Another new pressure on operational databases involves ingesting unstructured data, performing real-time embedding, creating vector indexes fast, and enabling hybrid search (lexical and semantic).
Not only are agents performing operational tasks, they are personalized based on the individual data they read. However, a lot of this “trajectory” data is generated from users’ behaviors and preferences, which is also stored in the database. Agents further learn user persona and behavior during the course of interactions. In the future these agents may continually improve from their own execution traces, creating newer forms of data that also need to be stored and managed. In some ways, these databases are acting as feature stores — the repositories that supply behavioral signals to models at inference time — without having been designed for that purpose.
While all this sounds great in theory, what happens to the cost of storing and retrieving data? Data storage costs are being addressed by storing data in object stores or on NVMe drives, but this is still a small fraction of the cost compared to retrieval, or inference. One of the biggest costs of running agents is inference generation via GPU. Caching is one way to reduce both cost and latency by redirecting common queries and query paths away from the GPU entirely. Operational databases now act as a “semantic” cache. The idea is that if agents ask the same question, the operational databases can reduce token cost and latency by avoiding excessive calls to an LLM. The query results will still be different for each user and will be served out of the database.
In addition to agents querying the databases, the existing UI for the operational databases will need to be augmented with chatbots and assistants as the front end. Orchestration of agents will become critical when we have multiple agents. If organizations use multiple models, an external orchestrator will be needed. However, models are becoming good enough to perform orchestration of agents on their own.
AI governance will become more critical on operational databases as well like detecting prompt injection and setting up guardrails. Organizations will need to set up red teaming processes to improve the reliability of agents.
The next section looks into the governance learnings from the conference.
Governance Has a Scope Problem
I heard plenty of discussions on how to go “from data governance to AI governance.” First, I think this approach sets up the wrong expectations. Semantics matter! This statement makes it look like data and AI governance are two separate initiatives. One attendee tried to explain his way out of this dichotomy by saying they are two sides of the same coin. But the problem is that each side of a coin has never seen and never will see the other side.
The term I arrived at was “upskilling data governance to AI governance.” Data governance is the foundation on which we build specific controls that AI workloads need. However, the problem lies in how we define the boundaries of those specific controls. The biggest challenge for an organization is to clearly define the “scope” and then address it carefully.
Another problem is to come to consensus on what AI governance is. Not an easy task, as my head started spinning as I asked various people what the process meant to them. Here is my running list of what needs to be governed:
- Model inputs: Guardrails, as mentioned earlier, need to be in place to ensure that inputs to AI are monitored and policies are in place to ensure correct use of AI. This catches problems like prompt injection.
- Model outputs: In deterministic systems, governing inputs is sufficient because outputs are predictable. In AI, outputs are probabilistic, meaning the same input can produce different results depending on context, model state, and data. This makes post-generation evaluation as important as pre-generation guardrails. Organizations need continuous evaluation pipelines and, where possible, reinforcement learning from human or automated feedback to catch drift, hallucinations, and policy violations before they compound.
- Model training: Most models do not share the dataset they were trained on, and hence organizations have the least control over whether the model provider used any intellectual property (IP) or not. In the early days, many providers offered indemnity in case the model underperformed, but I haven’t seen this term being used lately.
- Networking: MCP is an excellent protocol to call tools, but its proliferation is a new area that needs to be governed. Agents can call other agents using standards like A2A, which must also be governed.
- Data quality: In structured data, well-known attributes like duplicates, missing values, etc. suffice. But in an unstructured (and multimodal) world, data quality is subjective. It also includes bias and harmful content. In other words, it is contextual! An excellent point for us to jump to the hottest topic at the conference by far.
But before we end this section, it’s important to note how seriously Gartner is taking AI governance as they are coming out with a brand new Magic Quadrant in this space this year. While we are on the topic of Magic Quadrants, in 2020 Gartner had consolidated operational and analytical database MQs into a single Cloud DBMS MQ. In 2026, they plan to split this MQ into operational and lakehouse ones.
Context Graphs: The Hottest Idea Nobody Can Define
As the community knows well, Foundation Capital’s “AI’s Trillion-Dollar Opportunity: Context Graphs” took social media by storm. Before we knew it, many companies were claiming to be building exactly that. Never mind that context graphs are not even defined consistently.
What is old is new again. Metadata, ontology, taxonomy, knowledge graphs, and semantics have been debated ad nauseam for decades. Now comes context. In fact, Gartner has declared 2026 to be the year of context.
Of all the topics I explored at the event, this was by far the most controversial. There is no argument that context is needed to improve LLMs’ reliability. However, the notion that all the contextual information — written and tacit — will somehow be captured into a logically unified system for the LLMs to reason from seems like a huge undertaking. Especially when similar efforts to build semantic layers and knowledge graphs have not been hugely successful.
Two concerns worth raising before we leave this topic.
- I am not convinced that context needs to be a separate “layer” at all. Treating it as one complicates the architecture and risks creating yet another silo, attracting a new wave of standalone vendors. This is precisely the problem we created with the modern data stack and have spent years trying to unwind the spaghetti architecture. Adding a dedicated context layer on top of an already fragmented data estate may solve one problem while introducing another.
- It is worth being precise about terminology. Semantics is what humans use to add business meaning to data artifacts using definitions, metrics, and/or rules. Context adds situational awareness and is consumed by agents at runtime to understand what a question actually means in a given business scenario. These are related but distinct. However, maintaining separate stores for each seems like architectural overkill. The more pragmatic path is to extend existing semantic infrastructure rather than build parallel systems alongside it.
This topic needs a whole blog by itself, so more on that later.
AI Use Cases: In Search of Real ROI
My biggest desire going into the event was to understand from end users if they were using agents and, if so, was there any ROI they could point to. I can summarize my findings as:
- A lot of what businesses told me were agents turned out to be assistants. They were, at best, semi-autonomous. However, I wouldn’t hold this against enterprises. We are in such an early journey that it is laudable that organizations are seriously progressing on the path.
- Organizations are starting to see productivity gains, but they don’t necessarily want to go on record by publishing commercial success when so many jobs are getting lost.
I saw many fascinating examples of how AI was being used. A huge new successful AI use case I am seeing more is speech to text, e.g., call centers. Not just customer service but also healthcare. By law, Medicare patient calls need to be recorded in archaic CMS systems. If this is not done adequately, heavy penalties are applied. These calls are very difficult on the caseworkers and take a toll on them. The workers then have to create care plans by looking into case and pharma history databases.
Insurance companies are now able to use AI to transcribe calls with 95% accuracy and create care plans within minutes of receiving calls, rather than weeks. I was impressed to see that organizations are not just interested in ROI but are also focused on the cost. They are experimenting with on-premises GPUs, cloud AI services, and models like ElevenLabs to optimize their costs.
As the world’s population ages, healthcare may be the biggest beneficiary of AI advancements. In fact, AI couldn’t have come at a better time as we enter a world with more older people than caregivers. I was in a session delivered by an insurance company, where I learned that life expectancy differs by 20–30 years between the wealthiest and the poorest neighborhoods of Chicago.
The Uncomfortable Section: What Happens to Analyst Relations Now
I saved this topic for last because it is the most uncomfortable to write about, as it questions the entire space of analyst research.
The numbers are stark. Gartner’s stock has dropped roughly 71% from its November 2024 high, with its market cap collapsing from $45B+ to around $12B. Forrester, doing nearly $400M in revenue, now has a market cap of just ~$120M, about 0.3x revenue, a valuation the market typically reserves for businesses in structural decline.
What does this mean for those of us who work in the analyst relations ecosystem, whether as vendors navigating research reports, independents like myself, or the analysts that work at these firms? I offer a few honest observations.
There is no denying that enterprise budgets are tightening and AI is playing whack-a-mole with most professions, including ours. However, Gartner has advantages that AI cannot offer: proprietary data, original analysis, and community. Consider what AI did to Stack Overflow but not Reddit. The era of question and answer has been replaced by AI, but human judgment and trust have not.
Gartner should lean into exactly that. It has worked hard building barriers around its analysts, but as the premium shifts to judgment, relationships, and trust, those barriers are becoming liabilities. Gartner has remarkable talent whose voices are more valuable than ever precisely because of the noise and slop AI has created. Letting analysts be more market-facing, and gatekeeping less of their content, would be a stronger business strategy than the current posture. The one bright spot in Gartner’s recent results was its Conferences segment, up nearly 14% year-over-year with 51% margins. That is not a coincidence. It is the community argument made in financial terms.
Independent analysis is finding its moment, but it comes with its own pressures. AI is compressing the commodity end of independent research just as it is the institutional end. Buyers are looking for more nimble, specialized, and frankly more affordable voices. That creates opportunity and real challenges simultaneously. This year, for the first time in many years, I did not publish my annual trends and predictions. Technology was changing faster than I could pen it down, and there was already too much content on the same topics for another take to feel original.
AR teams are asking me what happens to their profession. My honest answer is that the role is bifurcating. If someone’s AR job was primarily logistics and relationship maintenance without strategic depth, that role is at risk. But understanding that a particular analyst is skeptical of a vendor’s go-to-market even when writing positively about the technology, or reading the internal dynamics around a particular research process is incredibly important. That kind of intelligence becomes more valuable, not less, because vendors who engage thoughtfully get more airtime and more meaningful dialogue.
The AR professionals who survive and thrive will be the ones who can walk into a product roadmap meeting and say, “Here is how analysts are thinking about this capability category and here is what our narrative gap is,” not the ones scheduling the next briefing. They will also need to understand how buyers actually use analyst content today, which is not the same as five years ago. Buyers are triangulating analyst research with AI-generated summaries, peer reviews, and community voices like podcasts. AR strategy has to account for all of that.
Conclusion
These five topics, operational databases, governance, context graphs, real ROI, and the future of analysis itself, are not separate conversations. They are one in the same, just viewed from different angles. AI is not just changing what we build. It is changing how fast we have to think, how quickly assumptions expire, and who gets to be the expert in the room.
A year ago, I left the Summit with a list of problems to solve. This year, I left with more questions than answers. That is not a step backward. It means the problems got harder and more interesting. The next trillion-dollar opportunity lies in how data platforms built for humans evolve in an era where humans and agents work together.
