General-purpose AI tools often win analytics pilots and then break down at scale, WisdomAI's Geetesh Iyer told COLLIDE Data + AI Conference 2026. He argued the missing piece is managed enterprise context, and that the people best suited to own it, AI context engineers, are already on data teams.
Key takeaways
- Pilots break at scale. Iyer said accuracy, governance, cost and consistency all suffer as AI analytics spreads across an enterprise.
- Models and more data aren't the fix. Conflicting definitions, like sales and finance measuring ARR differently, still produce different answers.
- Semantic layers capture the what, not the why. Enterprise context also needs decision traces, precedents and policies.
- Context needs an engine. Life cycle tooling and a runtime harness matter more than a context repository alone.
- A new role is emerging. Iyer said AI context engineers encode context, run evals and govern AI, and are already on the payroll.
Data teams are under constant pressure to use AI for everything, and the general-purpose AI tools that win pilots often break down at enterprise scale, according to Geetesh Iyer, Founding Product Manager at WisdomAI. In the session "Meet the Modern Data Team" at COLLIDE Data + AI Conference 2026 in Atlanta, Iyer argued that the fix is not a better model or more data but better-managed enterprise context, and a new role to own it.
Iyer, who said he previously worked on data at Apple and Amazon, described WisdomAI as an agentic analytics platform.
Why do AI analytics pilots break down at scale?
Iyer started with a typical request: a leader asks why gross margin declined in North America. In the old model, an analyst found the right data sources, wrote SQL, built pipelines and a dashboard, and went back and forth with the business. Time to value was measured in weeks, he said.
Today, the first instinct is to ask a general-purpose LLM. The pilot often succeeds, and leaders love it. But "the moment you hit scale, reality hits," Iyer said. Different departments run their own AI setups with different tools and skills, two people asking the same question get different numbers, and the data team has no visibility into why.
As more users, data sets and context are added, he said, success drops across four dimensions: accuracy, governance, cost and consistency. Iyer said WisdomAI's benchmarking with customers found general-purpose AI harnesses do research, coding and exploration well but grow expensive at scale because they take a brute-force approach to context each time.
Is a semantic layer enough for AI analytics?
Iyer walked through the usual next steps. Changing models doesn't solve it. Adding more data doesn't either. Then teams discover that sales and finance define a metric like ARR differently, and the AI returns two different answers.
Semantic layers help and are "extremely important in data heavy environments," he said, but they capture definitions, the what. What is still missing, according to Iyer, is a living record of an organization's decision traces: why the business made a choice, its precedents and policies, and the knowledge stuck in an analyst's head. He called that enterprise context.
What is a context engine?
Iyer criticized the common framing that a company needs only data, a context layer and user-facing surfaces. That view treats context as a repository and misses two blind spots, he said.
The first is life cycle tooling, which he called the context development life cycle: building, managing, governing and porting context, and bootstrapping it from places where it already exists, such as GitHub, query logs, catalogs and semantic layers.
The second is the harness, which he said few people talk about. It covers planning, entity resolution, intent clarification and classification, verification and repair at runtime, and learning from the process. Without that tooling and harness, he said, "just talking about context is absolutely useless."
"The context layer is an important but small part of the system," Iyer said. The larger system is the context repository plus the engine.
What is an AI context engineer?
The people who know which tables to trust and which definitions are right are already in the organization, Iyer said: data analysts, data engineers and business analysts. WisdomAI is promoting a credential for this work, which it calls the AI context engineer, or ACE.
In the traditional flow, a business question went to an analyst who wrote SQL and built dashboards over weeks to months. In Iyer's model, the context engineer has already encoded business and data context, set up evals, validated outputs and handled ongoing governance, so the tooling can answer questions from that foundation.
Because models now handle much of the repeatable coding work, he said, the role moves up. An ACE understands analytics, data architecture and governance and acts as a strategic business advisor. Iyer said ACEs can come from business teams as well as data teams. WisdomAI runs a certification program for the role and has completed three cohorts, he said.
How is the modern data team changing?
Iyer opened the session with a video of a singer performing and returned to it at the end. The performance is the accuracy of AI output, and the person urging the singer on is the stream of new models and tools leaders want teams to adopt. The key figure is the sound engineer at the console, who knows what works and what filters to set. That, he said, is the AI context engineer, and the console is the context engine.
"The assets best positioned to take you to the next phase of agentic analytics are already on your payroll," Iyer said. "The question is whether you'll redefine the role before the competitors do."
More on data teams and AI at COLLIDE 2027
COLLIDE, the largest AI conference in the Southern U.S., brings more than 1,000 enterprise data and AI leaders to Atlanta each year. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center. Between conferences, Data Science Connect's free AI × Enterprise Panel Series webinars bring data and AI leaders together online.
Frequently asked questions
What is an AI context engineer?
As Iyer described it, a data or business practitioner who encodes business and data context, runs evals, validates AI outputs and governs AI analytics, acting as the human in the loop.
Why isn't a semantic layer enough for AI analytics?
According to Iyer, semantic layers capture metric definitions but not why a business made decisions, its precedents and policies, or the knowledge in analysts' heads.
What is a context engine?
In Iyer's framing, the context repository plus the tooling and runtime harness for planning, entity resolution, intent classification, verification, repair and learning.
Why do AI analytics pilots fail at scale?
Iyer said teams using separate AI tools get inconsistent answers, and accuracy, governance, cost and consistency decline as more users and data are added.
Watch: Meet the Modern Data Team on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
