Core Concepts
I’ve defined a shared vocabulary for how sales conversations are ingested, assessed, evaluated, and coached. This feels non-technical, but it is critical infrastructure for an AI system.
This section defines the core terms used throughout the system.
Transcript
A Transcript is the canonical representation of a sales conversation.
It is the primary input to the system and represents what actually happened during a call, demo, or discovery session.
A transcript includes:
- Speaker turns (e.g. sales rep, prospect)
- Timestamps or ordering
- Optional metadata (organization, representative, call context)
Key principle:
Transcripts are immutable historical records. We never “fix” a transcript; we only re-evaluate it as prompts, rubrics, or models evolve.
Business rule validation, enforcing constraints such as score ranges, required fields, and completeness
Assessment
An Assessment is a single, structured evaluation of a transcript.
It represents the system’s judgment of how well a conversation was conducted, based on a defined framework (in this case SPIN).
An assessment:
- Is generated by an LLM (prompt-driven)
- Produces structured output
- Is tied to a specific prompt template and model version
Multiple assessments can exist for the same transcripts with different prompt template or model version.
SPIN Dimension
The system evaluates conversations using the SPIN selling framework, broken into four independent dimensions.
Each dimension is scored and reasoned about separately.
Situation
How effectively the representative establishes context about the prospect’s current state, environment, and constraints.
Problem
How well the representative uncovers explicit pain points, challenges, or inefficiencies.
Implication
How clearly the representative explores the consequences of those problems if left unresolved.
Need-Payoff
How effectively the representative helps the prospect articulate the value, outcomes, or benefits of solving the problem.
Key principle:
Each SPIN dimension is scored independently, not combined into one overall score.
Coaching Output
Coaching Output is the human-readable layer built on top of an assessment.
It translates structured scores and rationales into actionable guidance for a sales representative.
Typical coaching output includes:
- Summary – a concise, high-level read of the conversation
- Wins – what was done well and should be repeated
- Gaps – missed opportunities or weak signals
- Next Actions – concrete suggestions for improvement
This is where the system creates real business value.
Prompt Template
A Prompt Template defines how the LLM is instructed to perform an assessment.
It includes:
- Role instructions (e.g. “senior sales coach”)
- Scoring rubric definitions
- Output schema constraints
- Formatting and validation rules
Prompt templates are versioned artifacts. Changing a prompt does not overwrite prior assessments; it creates new ones.
Evaluation Dataset
An Evaluation Dataset is a curated collection of transcripts used to measure system quality.
These datasets are used to:
- Compare prompt versions
- Compare models (prompt-driven vs fine-tuned)
- Track regression or improvement over time
Key principle:
Evaluation datasets exist to measure system behavior, not to train intuition.
Evaluation Run
An Evaluation Run is a repeatable execution of the system over an evaluation dataset.
An evaluation run captures:
- Prompt version
- Model identity
- Scores and outputs
- Aggregate metrics (e.g. correlation, variance, agreement)
Evaluation runs allow the system to answer:
"Is this version better than the last?"
Organization
An Organization represents a tenant in the system.
Organization define:
- Their own scoring preferences
- Prompt customizations
- Evaluation datasets
- Access boundaries
Organizations are isolated by design.
Representative
A Representative is the individual being assessed and coached.
Assessments are always contextual:
- A score is not “good” or “bad” in isolation
- It is interpreted relative to a representative’s history, role, and goals