We’re increasingly hearing the question, “can’t I just drop survey data into ChatGPT, Claude, or Copilot?”. The short answer, yes, you can. But you’ll encounter serious problems with data security, numerical accuracy and analytical consistency. 

Having spent the past few years building agentic systems for market research, I’ve learned what works and what doesn’t. Here are six principles that should guide any implementation. 

The core difference: generative AI models are language predictors. Purpose-built agentic solutions use those models for orchestration while keeping data processing separate. When you ask “What drives brand consideration?”, a generative model produces plausible-sounding text. A properly architected agentic system searches existing analysis, runs validated tabulation, applies statistical tests and presents findings with traceable sources. One gives you content. The other gives you answers you can defend. 

1 ) Keep your data out of public endpoints 

Pasting data into a public model endpoint means losing control over what it may train, who accesses it and where it’s stored. This can breach processing agreements and industry codes. 

Well-managed enterprise deployments (such as Copilot within a Microsoft 365 tenant) may address data governance by keeping data within organisational boundaries. However, numerical accuracy and consistency issues remain: the model still processes your data as tokens rather than calling statistical tools. Secure hosting is necessary but not sufficient. 

2 ) Never let the model see raw numbers 

When you paste data into a generative model, everything becomes tokens. The model doesn’t compute statistics, it predicts what plausible-looking numbers might be. This leads to percentages that don’t sum correctly, fabricated base sizes and insights that contradict actual data. You cannot verify where a number came from. 

The right approach: the model should never consume raw data as tokens. Instead, give it access to tabulation engines tested in market research for decades. When an agent reports that 32% of prospects are aware of Brand X, that number should come from validated statistical computation, not a language model estimate. Every insight must trace back to a reproducible run, auditable by design. 

3 ) Preserve your curated table libraries 

Research teams invest significant effort building validated table specifications: correct variables, consistent filtering, proper weighting. Dropping raw data into a generative model discards all that intellectual property. 

The right approach is to build systems where the agent can search and run from pre-built, validated table specifications. Any data point should be traceable to source. Better yet, enable discovery of previous analysis through conversation history at the organisational level. This institutional knowledge compounds over time. 

4 ) Demand statistical rigour 

Survey analysis requires demographic weighting, subgroup filters and significance testing. Generative models have no native capacity for applying weighting schemes consistently, filtering to target populations, or warning when base sizes are too small. They cannot reliably execute the same analysis twice. 

The right approach is to combine dynamic spec generation by models with a proper cross-tabulation engine. Users instruct in conversational English; models generate coding syntax and pass it to a validated engine. The statistical computation must be deterministic and reproducible. 

5 ) Embed sector-specific expertise 

Generative models know nothing about your organisation, your sector or your projects. Every conversation starts from zero. 

This is where purpose-built agentic solutions differ fundamentally. These are custom systems built on top of foundation models, adding memory, context awareness and pre-built workflows. The agent should know your organisation, your projects, your data structure, and your previous analyses. Pre-built workflows for segmentation, TURF analysis, brand tracking and verbatim coding encode lessons from years of practice. Multiple team members receive consistent analysis, not whatever the model generates that day. 

6 ) Optimise for speed to insight 

Using a generative model for survey analysis requires extensive prompt engineering and manual verification. Not every stakeholder has time to coach a model into usefulness, repeatedly, across every conversation. 

The right approach is to eliminate ramp-up time. Pre-built table libraries mean common questions resolve in seconds. Outputs should be client-ready: branded presentations, interactive dashboards, not chat text needing reformatting. Most importantly, implement market research-grade guardrails.  The system should refuse to fabricate quotes, validate numbers against source tables and separate technical statistics from executive summaries. 

Summary 

Accessing the model is the easy part. The tooling around it is where value is created. Any solution worth considering should answer: 

1. Where does my data go, and who controls it? 
2. How are numbers computed, and can I trace them to source? 
3. Does the system understand my existing analysis, and analysis run without the model? 

4. What safeguards prevent fabricated insights reaching stakeholders? 

Market research has built its reputation on transparent methodology and data governance over decades. If a significant portion of our sector is using public AI models without MR-grade safeguards, we don’t just risk individual credibility, we risk the trust that underpins our entire industry. 

Matt Gibbs is chief executive and founder at Bayes Price