Glossary
Core AI and agentic concepts behind modern customer intelligence, from standard terms like RAG and embeddings to BaseDynamics' own vocabulary for agents, evidence, and signals.
Software that can perceive context, reason about it, and take action toward a goal with limited human intervention, distinct from a chatbot that only responds when asked.
Deploys three specialized agents, Activation Specialist, Retention Guardian, and Expansion Analyst, each monitoring and acting on a specific part of the customer lifecycle.
AI systems with multiple specialized agents designed to operate with autonomy, making decisions and taking multi-step action toward a goal, rather than simply generating a single response to a single prompt.
The entire platform is built agentic-first. Agents don't just surface data, they evaluate evidence, filter noise, and execute playbooks directly.
The amount of text (measured in tokens) a language model can consider at once when generating a response. Everything outside that window is effectively invisible to the model.
Pre-baked platform metrics are compiled and structured before being passed into any model's context window, keeping responses grounded rather than relying on the model to search broadly.
A deterministic system produces the same output every time given the same input (standard code logic). A probabilistic system, like a language model, can produce varying outputs, useful for reasoning over language, risky for exact calculations.
Critical calculations like usage limits, contract values, and health scores run on deterministic code logic, while language models are restricted to parsing sentiment and text patterns.
A numerical representation of text (or other data) that captures its meaning, allowing a computer to compare how semantically similar two pieces of text are, even if they don't share any of the same words.
Powers semantic mapping of customer voice data to the correct product feature area, without requiring exact keyword matches.
A single data point (a usage anomaly, a sentiment shift, a support ticket pattern) that an agent collects while investigating a potential risk or opportunity. Evidence on its own doesn't trigger action. It has to be validated.
Agents scan the Customer Knowledge Graph for complementary evidence before elevating anything to a Signal. This is the noise-filtering step that prevents alert fatigue.
Further training an existing AI model on a specific, narrower dataset to specialize its behavior for a particular task or domain, an alternative approach to RAG for grounding a model's responses.
BaseDynamics relies primarily on RAG (structured context retrieval) rather than fine-tuning, keeping the model current with live customer data instead of a static training snapshot.
The most advanced, capable class of AI models available at a given time, as opposed to smaller, cheaper, or older model tiers used for simpler tasks.
Every agent and the co-pilot are powered by frontier reasoning models, paired with deterministic guardrails to keep outputs grounded.
Rules, code-based checks, or architectural constraints placed around an AI system to keep its outputs safe, accurate, and within defined boundaries.
Deterministic validation layers restrict language models to parsing text and sentiment, while calculations stay in traditional application code.
When an AI model generates a confident-sounding but factually incorrect or fabricated response, a known risk of ungrounded language model outputs.
Pre-baked context delivery and code-based guardrails are specifically designed to prevent hallucination drift in the co-pilot's answers.
The process of an AI model generating an output (a prediction, a response, a classification) based on its training, as opposed to the training process itself.
All customer voice and agent reasoning inference runs inside secure, private inference loops that are isolated from public model training pipelines.
A structured, interconnected representation of data and the relationships between them, designed to be queried and reasoned over, rather than a flat table or document store.
The Customer Knowledge Graph links usage telemetry, customer voice, and CRM/contract data across every account into one unified structure that the agents and co-pilot both query.
An AI model trained on large volumes of text data, capable of understanding and generating human-like language, the underlying technology behind most modern conversational AI and agents.
Frontier LLMs power both the background agents' reasoning and the conversational co-pilot, always paired with deterministic guardrails for anything involving calculation.
An AI architecture where several specialized agents, each responsible for a distinct task or domain, work in parallel or coordination, rather than a single general-purpose agent handling everything.
Three specialized agents, Activation Specialist, Retention Guardian, and Expansion Analyst, each own a distinct part of the post-sales lifecycle instead of one generic assistant trying to do all three.
The input text or instruction given to an AI model to generate a response. It can range from a single user question to a large, structured system instruction.
User questions to the co-pilot are combined with pre-compiled platform context before reaching the model, so the prompt the model sees is already grounded in real account data.
An architecture where a language model's response is grounded by first retrieving relevant, structured data and passing it into the model's context, rather than relying purely on the model's trained knowledge.
The core architecture behind the co-pilot, filtering and structuring database metrics before passing them to the model, specifically to increase accuracy without hallucination drift.
A high-trust, evidence-backed conclusion an agent reaches after corroborating multiple pieces of evidence, the point at which an agent moves from "noticed something" to "confident enough to act." A team member can approve, modify, or dismiss a Signal.
Signals are what actually trigger playbooks, task creation, and alerts. One-off anomalies with no corroborating pattern never become a Signal.
The basic unit of text an AI model processes, roughly a word or part of a word. Model usage and cost are typically measured in tokens processed.
Internal AI usage is benchmarked in tokens, but this is never exposed to customers as a billable credit. It is absorbed into the platform's included usage.
A database that stores data as embeddings and retrieves results based on meaning rather than exact keyword matches, the technology behind mapping unstructured text to the right structured category.
Runs incoming customer voice data through a vector database to automatically map conversations to the exact product module they relate to, without manual tagging.
See how BaseDynamics uses evidence, signals, and frontier models with deterministic guardrails, not ungrounded chat.
