GWI’s Tom Smith on Agent Spark, agentic AI and why 2026 is the year the insights industry has to move
The GWI founder and CEO on supercharging the business around AI agents, why most LLMs produce “friendly rubbish” without proper data, and what going agentic first actually requires.
The GWI founder and CEO on supercharging the business around AI agents, why most LLMs produce “friendly rubbish” without proper data, and what going agentic first actually requires.
When Tom Smith and I spoke at Cannes Lions last year, GWI’s AI ambitions were already clear enough. The company had built foundational infrastructure to run natural-language queries across its dataset, and Smith was carefully considering how consumer insights could flow into the tools marketing teams use. That conversation was barely nine months ago. A lot has changed.
Agent Spark launched in January 2026, integrating GWI’s proprietary dataset, 35 billion data points drawn from 1.4 million annual surveys across 52 countries, directly into ChatGPT, Claude, and Microsoft Copilot. The idea is to put verified, human-grounded consumer insights into the AI tools marketing teams are already using, without the detour through a separate analytics platform or a request to a research team that’s already at capacity. More on how it works at Agent Spark.
Smith is direct about what prompted the move. “We could see that what people wanted was to connect insights to proper output and answers,” he says. “And one of the technologies that transformed the ability to do that was MCP — Model Context Protocol — which wasn’t really around even a year ago.”
MCP is what makes Agent Spark work at this level. It allows GWI’s agent to communicate directly with other agents, such as Claude, ChatGPT, and Copilot, so that insights are automatically pulled into the workflow a marketer or sales team is already running. What was previously a separate step, requiring a trip to the research department, becomes part of the same conversation.
The lean insight team’s problem
Smith’s argument for Agent Spark rests on something that rarely gets discussed openly: insight teams inside large organisations are extraordinarily lean. “It’s not untypical that you might have two researchers supporting a thousand salespeople,” he says. “You might have a hundred thousand employees and a hundred people working in consumer insights. But the demand for that data just can’t be satisfied at that scale.”
Rather than positioning Agent Spark as a cost-reduction tool, Smith frames it as a growth play. More people getting insights means more decisions grounded in data. Sales teams pitch faster. Marketing strategies get stress-tested before they go live. The research function’s reach expands without requiring additional headcount.
One of his early case studies is from a major electronics goods company. The shift from a handful of expert users of GWI’s products to a full marketing organisation illustrates the three things Smith says Agent Spark is actually solving:
Speed: more than a hundred marketers now get answers directly, without routing through the research team
Language: the agent works in Cantonese or whichever language the team is operating in, removing a significant barrier for globally distributed organisations
Scale: what previously required specialist knowledge to extract and interpret is now available to anyone with a business question
“That’s speeding up workflows by a thousand times,” Smith says.
Omnicom is another reference point. Bharat Khatri, Chief Digital Officer APAC at Omnicom Media Group, has described testing Agent Spark in the agency’s autonomous-agent workflows, with plans to integrate GWI audience taxonomies into end-to-end media activation and creative iteration. Smith explains what that looks like in practice: when a platform needs to tailor a global campaign across 25 markets, it can now automatically pull audience context from GWI, feeding the data into creative models as input for localisation. A process that previously required considerable human coordination can now occur in near real time.
Why generic AI gives you “friendly rubbish”
Accuracy is where the conversation gets interesting. GWI’s own research suggests 85 per cent of marketers use AI to support their work — but most of that is generic AI, without access to properly structured audience data.
“If the data isn’t there, ChatGPT and Claude produce friendly rubbish,” Smith says. “They produce really generic, inaccurate output that is difficult to act on, and can be quite dangerous if you use it. Most consumers aren’t represented in the content that trains large language models.”
I pushed him on the hallucination risk, the concern that the AI presentation layer might introduce inaccuracies even when the underlying GWI data is correct. His answer comes down to the difference between open-ended and concrete questions. Ask Agent Spark a broad question, such as the ten most unexpected things about Xbox gamers in Germany, and then interrogate the output. Ask how many people in Germany use Xbox, and the answer will be correct every time because the model pulls from a deterministic data layer rather than generating text via pattern matching.
“As long as the question is right, the answers will come back accurate,” he says. “We’ve spent a lot of time testing and calibrating. If you bring an agent to market, you have to validate this.”
Data company, AI company, or both?
I asked him the obvious question: Is GWI still a data business, or has the product evolved into something else? He doesn’t think the distinction holds any more.
“We can’t operate like a software business from ten years ago,” he says. “Our whole strategy is about being an agentic-first business. Even if you come into our platform now, the dashboards are still there, but over time, everything will evolve into an environment where you use the agent to do any task.”
What Smith is describing positions GWI not as a competitor to the LLMs, but as the data infrastructure they need to be useful in a marketing context. The LLMs handle the interface and reasoning; GWI provides the human truth beneath the surface. That’s a bet that proprietary, first-party data becomes more valuable as AI spreads, not less, because without it, the outputs are, as he puts it, friendly rubbish.
He points to companies building similar models in adjacent sectors — structured, proprietary data used to ground AI outputs for specific professional contexts — as evidence that the model works. “LLMs and agents are incredibly accurate if the data and context are well-defined and accurate,” he says. “The biggest challenge is having clean data in the right format. That’s our focus.”
2026: the agentic pivot year
Smith is confident this is the year the agentic model breaks through in agencies. The past twelve months have been build-out — platforms constructed in the background, partnerships established, pipelines developed. This year, those systems will go public.
“Every agency will need that kind of system in order to be relevant and competitive,” he says. “I think this is the big year where it breaks out.”
On the industry’s shape by 2027, he is more cautious. The mega-mergers — Omnicom’s acquisition of IPG being the most significant — are partly about building the scale to invest in these platforms. Smaller boutique agencies may find that AI gives them the reach that their size wouldn’t previously have allowed. The squeeze, he thinks, will fall hardest on mid-tier groups: large enough to need the infrastructure, not always large enough to fund it.
There’s also a counterintuitive point he makes about workloads. In creative production, the ability to generate infinite variants of an ad doesn’t reduce demand — it tends to increase it. Clients who once wanted three ideas now expect ten. “AI can make more work, not less,” he says.
GWI is reorganising its entire business around agents rather than attaching them to existing structures. Platform, product, commercial model, user experience — all of it. Smith says companies that approach this by “just building a couple of chatbots” won’t cut it. Whether that assessment ages well will depend on how quickly the rest of the market catches up.
