How we use AI at GWI

All of GWI’s services - including databases that support our AI features - are hosted on Google Cloud Platform (GCP), specifically within the europe-west1 data center located in Brussels. For more information, see Google Cloud's data center locations.

This is currently the only hosting location we use. However, GWI retains the option to add additional regions in the future, for example to support disaster recovery, performance optimization, or other operational needs.

Access to logs and data within GCP is strictly controlled and protected by encryption both in transit and at rest. By default, Google-managed encryption keys are used,  including customer-managed encryption keys (CMEK), as part of its internal security architecture where appropriate.. For further technical details, please refer to Google Cloud’s encryption documentation.

All outputs from Agent Spark (either: in the platform, Spark MCP and Spark API) follow the same GWI-side process below. For integrations, the GWI side flow remains the same but the difference is that the integrating LLM will render the final response (LLM Interpretation).

Agent Spark in the platform includes a thinking and source overview. In practice, this means the user can see a background actions summary together with structured metadata about the data used, including fields such as audience, topic, timeframe, location, and dataset, so the user can understand how the result was produced and what underlying GWI data it is drawing from.

1. User Input: The user submits their input via the text box or uses a pre-selected audience. For example, a user can type: "What were Gen Z's interests in energy drinks in 2025 for the UK?"
2. Prompt Transformation: The text-box user input is interpreted by the agent so it can be turned into a queryable plan against GWI data. For example: "Audience: Gen Z; Topic: Energy drinks; Timeframe: 2025; Location: UK". For broad or multi-topic questions, the request can be split into subtopics and make multiple calls to the insights tool.
3. API Calls: The transformed request or pre-selected audience is passed into the Spark/Insights API flow, which then interacts with the underlying data services.
4. Database Query and Calculations: Each insights-tool path calls calculations first. The system retrieves structured results for the relevant audience, topic, timeframe, location, and dataset scope, such as interest levels, index scores against the average population, depth-of-interest metrics, and awareness figures for each league within the defined audience.
5. LLM Interpretation: The insights engine then prompts the LLM with structured data derived from the calculations context so the output can be turned into a readable narrative response.
6. Output: The user receives generated insights (e.g. “56% of Gen Z in the UK drink energy drinks at least once a month”) together with related origin metadata and source links to the underlying GWI Data used to generate the output, allowing the user to verify how each insight was derived.

Agent Spark has a documented evaluation framework covering end to end agent evaluation, audience-tool evaluation, insights-tool evaluation, and component-level evaluations.

GWI documents agent evaluation criteria that include query success, reasoning clarity, safety compliance and policy alignment, hallucination detection, chat coherence, and audience quality.

GWI also documents a dedicated grounding evaluation path that checks generated insights against underlying source data for numerical accuracy and text hallucinations.

GWI has also built in transparency by design in the documented flow:

• the insights pipeline returns origin metadata for generated insights, including audiences, attributes, datasets, waves, and locations

• the agent documentation states that citations are converted into interactive source links for the user

• GWI also documents a dedicated sources-evaluation framework that checks source assignment accuracy, number preservation accuracy, citation format accuracy, and source embedding similarity

The main documented checks currently available are:

1. End to end agent evaluations

2. tool and component level evaluations

3. grounding checks against source data

4. Source attribution and citation quality checks

These checks operate after or around generation as evaluation and validation layers over generated outputs.

The fact that GWI documents hallucination detection, numerical grounding checks, and source-attribution evaluation means the system is designed to detect and reduce risks such as unsupported claims, number mismatches, or incorrect source assignment. These are the kinds of issues the documented evaluation layers are intended to catch.

Given Agent Spark’s single source architecture and its grounding in real respondent data, the primary risk in this context would be the misrepresentation or misinterpretation of statistics, rather than the invention of entirely fabricated data. Nevertheless, users are always able to verify outputs, as all GWI insights are accompanied by a clear reference that can be easily accessed to cross check the underlying data and confirm consistency between the output and the source.

Both integration services proxy the insights generation to GWI's Spark Service Layer, following the same underlying flow as Agent Spark in the platform, intent extraction, database query, statistical calculation, and LLM interpretation of structured results.

• Spark MCP: The host LLM, which can be ChatGPT, Claude, Microsoft Copilot, Gemini CLI, or a client's own direct LLM integration, acts as the agent. The host LLM receives the user's natural language prompt, calls the relevant GWI Spark tools to retrieve insights, and assembles the final response. The output is then rendered by the host LLM and returned to the user.
• Spark API: The Spark API is designed for clients who wish to integrate GWI insights directly into their own tools without a host LLM. Clients submit natural language questions via the API, which proxies the request to the Spark Service Layer. The structured insights, including audience size, audience percentage, and index figures, are returned directly in the API response, along with origin metadata and source links so results can be verified.

LLM:

1. User Input: The user submits their input via the text box or uses a pre-selected audience. For example, a user can type: "What were Gen Z's interests in energy drinks in 2025 for the UK?"
2. Prompt Transformation: The text-box user input is processed by the LLM to generate a prompt that can effectively query the database. For example, "Audience: Gen Z; Topic: Energy drinks; Timeframe: 2025; Location: UK".

Spark MCP/Spark API:

1. Call: The transformed prompt or the pre-selected audience is fed into the relevant GWI Spark tools, which then interact with the database.
2. Database Query: The database fetches the data that answers the prompt. For example, "Year: 2025; Topic: Energy drinks monthly; Gen Z UK %: 56%".

LLM:

1. Output: The LLM takes the answer from the database via the relevant GWI Spark tools and turns it into text that the user can easily read. For example, "56% of Gen Z in the UK drink energy drinks at least once a month". The text is presented to the user in a clear and understandable format.

For clarification, the original prompt entered into the LLM such as ChatGPT (e.g., "Give me an analysis of how millennials and Gen Z prefer watching Premier League"), does not reach GWI's input database. GWI only collects the transformed prompts generated by the LLM (e.g. "search GWI for millennial watching Premier League").

As set out above, for platform services, GWI relies on ChatGPT for its infrastructure and ChatGPT confirms that API data is retained for up to 30 days.

For API services, your LLM of choice will set the retention period.

The AI model is hosted by OpenAI. GWI uses GPT-5 via API, along with fine-tuned versions of GPT-4o and GPT-4o mini.

GPT-4 model card is here

GPT 5 model card is here

Yes, GWI may fine tune responses received based on Agent Spark's inputs via the Platform. For example when a user searches for "insights about men in the US who own dogs" and GWI provides an insight, GWI may use the search to improve the accuracy and relevance of the insights generated. However, where Agent Spark is accessed via third-party tools, the prompts are not used to fine tune Agent Spark.

The Input database is used to power GWI Canvas, so this feature benefits from the fine tuning to identify the insights most useful to each audience.

GWI’s infrastructure and multi-layered safety mechanisms are designed to ensure that all AI-generated outputs are responsible, relevant, and aligned with the intended use of the platform.

When a user submits a prompt, it’s evaluated through a series of automated guardrails, including rule-based filters, OpenAI’s Moderation API, and LLM based relevance and safety classifiers. These checks help prevent misuse, ensure prompts are on topic, and harmful or inappropriate inputs are blocked.

Once cleared, GWI data (which is anonymized and aggregated) and the prompt are processed by the LLM. The model is used strictly to reformat inputs in order to surface insights from GWI data and ensure outputs are readable while respecting ethical and compliance boundaries.

Outputs are never generated from external or unvetted sources, and the system is continuously reviewed by our Product and Legal teams to maintain alignment with evolving ethical standards, regulatory requirements, and industry best practices.

Yes. AI features include built in safeguards to ensure responsible and stable use of the system. Each user is currently limited to one prompt per second and a maximum of 100 prompts per hour, helping to prevent misuse and ensure fair access across users. These thresholds may be adjusted as needed to maintain platform performance and reliability.

In addition, output length and complexity are also constrained by system level controls to ensure that results remain relevant, manageable, and aligned with the intended use of the platform. These measures form part of GWI’s broader commitment to responsible AI deployment.

GWI is implementing a multi-layered set of safeguards to prevent malicious or inappropriate use of free text prompting in AgentSpark, which is the only AI feature on the platform that accepts open ended inputs. These guardrails include:

1. Rule based protections, such as a blacklist of harmful or inappropriate terms and character limits designed to restrict overly long inputs.
2. OpenAI's Moderation API, which automatically flags and blocks inputs that may violate safety policies.
3. Custom LLM based classifiers, including:

• A relevance model, which detects and blocks off-topic or spam prompts before they reach the AI system.
• A safety model, designed to prevent attempts at jailbreaking or misuse of the AI.

These guardrails are being continually developed and strengthened to reflect best practices in AI safety. If a user input fails any of the GWI checks, it's not processed by the AI and the user is prompted to revise their input. Only inputs that pass all safety and relevance checks are handled by the AI agent.