AI-powered Insights Tailored to the Consulting Sector

Accuracy. Human review of AI model outputs plays an important role in our validation process, and we’re committed to being transparent about our accuracy rates.

Bias mitigation. We test our models in multiple languages and across countries to ensure that any training data skews used by a third-party algorithm are not contributing to biased results.  We also build guardrails into our model applications to mitigate hallucinations.

Duty of care. We are cautious about using AI in any autonomous action towards human beings. For example, our AI-based probing tool, QProbe is designed to ensure that survey participants are not probed on sensitive topics they may have mentioned in their response, unless that topic is crucial to the survey (of which the participant has already been briefed). During testing, we also ran analyses to ensure that this prompt was not served to indicate a skew toward any gender, age, ethnicity, or education level.

We add the symbol to indicate the use of AI in our DIY survey platform and in client reports. In data files, we create a separate field that indicates whether the response is synthetic or not.

We assess the overall quality of our training data in four ways:

1. Accuracy checks that ensure data correctly represents real-world scenarios. We use a combination of automated tools and manual reviews of large samples to do this. Accuracy is checked both on the input and output of the AI model.
2. Completeness evaluation that identify any gaps in our data. We define initial expected clusters and balance the data between these clusters. This often leads to more data collection or covering missing areas in existing data sets. The latter is achieved through a combination of automated tools and manual sampling.
3. Compliance with data privacy laws to ensure all data complies with applicable privacy laws and regulations. We are committed to protecting personal information and respecting user privacy, as well as being transparent on the use of individual personal data. We use anonymization techniques when training models. In the case of third-party models, we ensure they don’t use information we submit to train their models.
4. Continuous monitoring and adaptation to new insights or changes in the real world. This enables us to maintain a high standard of data quality throughout the lifecycle of our AI models. Monitoring is done by automated anomalies detection, client/user feedback, and frequent manual checks of random samples.

1. Guardrails: This includes restrictions on what the AI can generate or decide, with monitoring to detect and alert on unusual behavior. It also includes mechanisms that allow human intervention if the AI behaves unexpectedly, which helps us maintain control over the AI’s actions and ensures it operates within the ethical and operational boundaries we’ve set.
2. Noise and disruption management: We use techniques like data augmentation and regularization to make our models less sensitive to noise and more capable of handling real-world variability.
3. Fallback plans: In case of a significant problem, we have recovery strategies and backups in place to restore functionality. This ensures minimal disruption to our services and protects data integrity. AI products are subject to secure software review, architecture, data governance, and third-party assessments. They are checked for prompt injection, OWASP  top 10 for LLM models.We also ensure AI safety and security assurance by training our teams on these topics. In addition, we check for existing vulnerabilities in LLM models.