Sales call analysis: turn a transcript into decisions, not just a summary
A transcript tells you what people said. A useful analysis tells you what changed, what is still uncertain, and what your team should review next.
Start with the decision the call should inform
Before asking AI to analyze a sales call, decide who will use the output. A seller may need a follow-up task. A revenue operations leader may need a field that accurately captures the approval process. A manager may need to know why the deal has stalled. A long, generic summary can miss all three needs.
The first useful unit is a finding linked to evidence: a short statement, the exact supporting quote, the speaker or source line, and a proposed action. That makes the output easier to check and easier to use.
Separate facts, targets and assumptions
Consider this fictional statement: ‘The budget is approved, but Legal still needs to sign off. We would like to decide by October 30.’ It contains three different types of information. Budget approval is an explicit assertion. Legal approval is an unresolved dependency. October 30 is a target, not a promised close date.
An analyst should preserve those distinctions. Turning ‘we would like to decide’ into ‘will sign on October 30’ adds certainty that the conversation never supplied. A confident-looking field can still be wrong.
Use a small, consistent extraction checklist
Start with budget, business problem, buying criteria, stakeholders, competing alternatives, approval dependencies and commitments. For each item, ask whether it was explicitly stated, reasonably inferred, or still unknown.
Next steps deserve particular care. A buyer’s request and a seller’s commitment are related, but they are not identical. Capture the named owner and stated deadline only when the conversation supports them. Leave missing information unresolved rather than completing the story for the user.
Evaluate correction effort, not just polish
Test the analysis on a small set of authorized conversations. Count unsupported findings, missed commitments, incorrect owners and dates, and suggestions that require substantial rewriting. A report that looks polished but takes ten minutes to correct may not improve the workflow.
Also ask whether anyone acted on the output. Did the follow-up get sent? Did the reviewer correct a record? Did the manager understand a risk earlier? Those actions are more useful evidence than the number of summaries generated.
Try the workflow before buying a platform
Call Refinery is a pre-launch experiment in this narrower approach. Its interactive example lets you inspect fictional buying signals, follow the supporting transcript, and edit review-ready suggestions. The initial live-analysis flow accepts text transcripts; it does not connect meeting bots or write to Salesforce.
Start with the example, then decide which outputs would be valuable enough for your own team to use repeatedly.
See the evidence for yourself.
Explore a fictional transcript and inspect the signals.
Explore the interactive demo ↗