Validating AI results in Spotlight.ai

Spotlight.ai uses AI to extract structured insights ("Signals") from meeting transcripts (for example: expected production date, number of licenses, customer pains). While the system is designed to be accurate and evidence-based, AI outputs should be treated as assisted analysis rather than ground truth. Customers play an essential role in ensuring that finalized records are correct and aligned with their internal definitions and sales process.

Clear separation between AI answers and user answers

Users can always distinguish between what the AI produced and what a human entered or confirmed. The UI clearly labels:

• AI-generated answers (machine suggested)
• User-confirmed or user-entered answers (human authoritative)

This separation ensures transparency and prevents accidental acceptance of AI content without review.

Evidence-based answers for trust and auditability

Every AI answer is accompanied by evidence from the transcript. Users can view the specific transcript sentences that support the AI’s conclusion, highlighted directly in the transcript view. This lets reviewers quickly validate whether:

• the AI interpreted the conversation correctly,
• the answer is supported by what was actually said,
• the context reflects a confirmed plan vs. a hypothetical or early-stage estimate.

Evidence visibility is a core part of the workflow: it enables fast verification and provides an audit trail for why the AI suggested a particular value.

Users remain in control

AI answers are always editable. Users can:

• overwrite an AI answer with a corrected value,
• update the extended explanation if needed,
• confirm the answer as final once it matches their understanding and internal definitions.

This ensures that the customer’s final dataset reflects human judgment and business context, not only automated extraction.

Review expectation

Customers should establish internal expectations for review, especially for fields that are high impact (dates, quantities, commitments, pricing-related values). While AI reduces manual work, the best outcomes occur when teams treat AI suggestions as a starting point and apply quick validation before finalizing.

How Machine Learning reviews, matches, and corrects Signals

To maximize accuracy and reduce incorrect or unsupported outputs, our system uses a two-step review approach when generating Signal answers from transcripts. This approach combines broad discovery (where is the evidence?) with focused answering (what exactly is the answer?). The goal is to produce outputs that are both accurate and explainable.

Step 1: Identify the most relevant evidence

First, the system scans the transcript to find the sections most likely to contain the answer to each question. For each Signal, it selects the most relevant parts of the transcript, or determines that the transcript does not contain enough information to answer confidently.

This step is designed to ensure that answers are based on the right part of the conversation, rather than guessing from unrelated context.

Step 2: Extract the precise answer from the evidence

Next, the system uses only the selected evidence to produce the final result for each Signal. This output includes the exact answer in the expected format (date, number, pick-list option, etc.), an extended explanation of the interpretation, and highlighted sentences in the transcript showing exactly what the AI relied on.

By focusing on a limited, relevant subset of the transcript, the system reduces the chance of pulling in incorrect context or unrelated statements and increases the likelihood that the final answer is grounded in what was actually said.

Why this improves correctness

This two-step approach helps ensure quality in three ways:

1. Grounding: answers are tied directly to transcript evidence, not assumptions.
2. Consistency: extraction is guided by the most relevant sections instead of the entire conversation.
3. Reviewability: users can quickly see why the AI produced each answer and correct it when necessary.

Together, these steps create a practical balance: AI delivers speed and structure, while customers maintain final control and can validate every result with clear evidence.