FAQ: How does Jam use AI, and is it secure? 🇬🇧

☑️ How Does Jam’s AI System Work?

• The Jam platform is a tool designed to support human teams through the use of AI, particularly through simulations of sales conversations (”role-plays”). We place the highest priority on data protection and security.

• Brief system description:
1. Preparation (”content creation”)
• Customer company shares information about required training (scenarios, personas, "scorecards" with evaluation criteria, possibly sales materials (PDFs, guidelines)
• Jam and/or the customer company's training administrators create individual AI training scenarios based on this information
• The shared information flows exclusively as instructions into the AI models' prompts, can be viewed and modified at any time by training administrators, and is never used for "model training"
2. Conversation simulation (”training”)
• Employee conducts simulated sales conversation via a voice-to-voice (V2V) interaction powered by a large language model (LLM)
• AI system (LLM + V2V) takes on the customer's role and responds in real-time
• Conversation is recorded and transcribed
3. AI Evaluation (”analysis”)
• The transcript is analyzed by another LLM
• This LLM evaluates the transcript based on “scorecards” determined by Jam and the customer regarding, e.g, conversation skills, objection handling, product knowledge, customer orientationcreates improvement-oriented feedback for the employee
• The LLM generates personalized feedback with improvement suggestions for the user
4. Results (”feedback”)
• Employee receives audio, transcript, and improvement-oriented AI feedback
• Trainers/coaches and - if desired - supervisors can access training results with the sole purpose of supporting employees in their development

☑️ Where is AI used in Jam's coaching solution?

• “AI” is used in several places in Jam’s app.
• Speech recognition/transcription: Converting spoken language to written language
• Conversation simulation: Large Language Model (LLM) conducts conversation simulation with the user by “playing” the role of a fictitious customer
• Feedback: Large Language Model (LLM) evaluates conversation transcript based on evaluation criteria transparently communicated to users
• Content editor: LLM supports training administrators as a "copilot" in creating training content based on Jam's templates

• Important to know:
• ✓ Data storage: All data is stored in the EU
• ✓ Transparency: Employees see their own data and evaluations
• ✓ Data protection: DPAs with all providers, no use for AI training
• ✓ Co-determination options: Access by supervisors to training data can be individually granted or denied

☑️ What Data Sources Does the AI System Use - and How Does It Ensure Data Reliability and Integrity?

• The AI models used at Jam (currently primarily OpenAI's GPT models) were trained by the model providers using large amounts of publicly available texts such as books, websites, databases, etc.

• Neither these model providers nor Jam itself ever use customer data, uploaded documents, images, videos, or audio data from operational use for training the underlying AI models.

• The only way non-public information from our customer companies is used in connection with Jam's AI is through "prompts," i.e., instructions to the AI models for conversation simulations (e.g., what role the AI persona should play, what scenario is being practiced) and analysis of the resulting transcript for improvement-oriented feedback.

• What information goes to the model as instructions via prompts can be viewed and modified at any time by training administrators in our role-play editor (”Jam Studio”). Customers can determine who receives administrator rights.

• The voice training sessions conducted (audio/transcript) for real-time analysis (improvement-oriented feedback) are sent exclusively as context to the AI model during operational use. No data is incorporated into OpenAI's model training ("no data retention for training purposes" when using the API). OpenAI's Data Processing Agreement (DPA) stipulates that all data transmitted via the API is processed exclusively for service delivery but not used to improve or train the AI.

• Summary:
• Jam does not train its own AI model with customer data but uses pre-trained models from third-party providers (for AI coaching: OpenAI).
• OpenAI does not use customer data for training purposes in API operations.
• Prompt contents are under the control of our customers' training administrators.
• Evidence includes OpenAI's publicly available API Data Usage Policy and the DPA with OpenAI, which explicitly contains this regulation.

• We are happy to provide or reference the relevant passages from the DPA with OpenAI for your data protection team.

☑️ How Is the AI System Trained?

• The AI model used by Jam (OpenAI GPT) is trained exclusively by OpenAI continuously based on publicly available data and licensed material. Customer data from the Jam system is neither permanently stored nor used for training, model improvement, or research by OpenAI. For real-time analysis, data is only processed temporarily ("volatile") and not incorporated into the training set.

• Improvement of our services occurs through manual creation, curation, and annotation of synthetic data that Jam creates internally and makes available to models exclusively temporarily via few-shot prompting.

• We consciously avoid a self-learning, autonomous system and prioritize human control ("human in the loop")

☑️ Does the AI System Learn Autonomously or Only Under Supervision?

• No, our system is not an autonomous (self-learning) one.

• We deliberately rely on human control ("human in the loop"). Any changes after operation begins are under the full control of human users.

• These changes in LLM outputs occur mediated through the prompts of the AI conversation simulation (role-play prompt) and the AI feedback system. These are executed by training administrators via a CMS/LMS and can be viewed there at any time.

☑️ Is There a System Description That Makes the AI Model, Its Logic, Functionality, Scope, and Data Sources Understandable?

• Yes, such a system description exists, see answers above and additional documents.

• In a nutshell:
• The system consists of conversation simulations with feedback for training conversation skills. Users conduct role-plays via voice interaction and receive improvement-oriented feedback based on the resulting transcript.
• The role-play is based on voice-based interaction with a large language model (LLM) via speech-to-text (STT) and text-to-speech (TTS) technology.
• The feedback is also based on an LLM that analyzes transcripts from role-plays and outputs structured text feedback.
• The audio recording of the role-play interaction is also stored for the user to listen to but is not otherwise processed.
• Both LLMs (for conversation and feedback) are instructed through prompts.
• The prompts are under full human control. No third-party data sources are used.
• The system's scope is limited insofar as it serves exclusively training purposes as a coaching tool and makes no autonomous decisions about the user.
• Users also have the option at any time to provide feedback in the app on whether they agree with the AI feedback or not.
📎
Additional Documents - please request access if not yet granted:
• List of Jam’s Subprocessors
Jam Subprocessors.xlsx
12.2 KiB
• System Description with Notes on Security and Data Protection
Security_Trust_Deck_20250924_EN.pdf
874.4 KiB