How to build a conversation intelligence to CRM sync

What matters here

• Extract what a rep would act on, not everything a model can find. Twenty new fields means nobody reads any of them.
• Link every claim to its moment in the call. An insight nobody can verify is one nobody should act on.
• Never let a model set a close date, a competitor or a forecast category. Those are commercial claims and they need a person.
• Write next steps and attendees automatically. Those are factual, low risk and the ones reps skip.
• Measure whether an extracted field ever changed a decision. If not, stop writing it.

Who this is for

You run revenue operations. Calls are recorded and analysed, the insight sits in a separate tool, and the CRM record is still whatever the rep typed afterwards.

How it works in practice

What happens between a call ending and the record being worth reading.

1. 1

   The call is matched to an opportunity and its attendees

   By calendar invite and participant email instead of by guessing from a title.

2. 2

   A small set of things is extracted

   Next steps agreed, attendees and roles, competitors mentioned, objections raised, and any date committed to.

3. 3

   Every claim carries a link to its moment

   So anybody reading it can hear the sentence it came from.

4. 4

   Factual fields are written automatically

   Attendees, next step, call summary. Low risk and the ones reps most often skip.

5. 5

   Commercial claims are proposed, not written

   Competitor, close date, forecast signal. The rep confirms with one click or corrects.

6. 6

   Usage is measured per field

   If nobody ever acts on a field, it stops being written.

Paste these into Claude Code or Codex with the Tray Headless plugin installed. Each stage runs on its own. The systems named in them are the worked example rather than a requirement, and every prompt says so.

Once per project, run /tray-workflows:set-workspace to pick the workspace these build in. Point it at a sandbox first.

1. 1

   Set up and choose five things worth extracting

   Choosing what not to extract is most of the design.

   Headless skills build-workflow

   Copy

   Use build-workflow. The systems in play are Gong and Salesforce, or
   whatever we run in those seats.
   
   Before building, choose at most six things to extract, each passing one
   test: would a rep or a manager behave differently if this field were
   populated.
   
   My starting set: next steps agreed with owners and dates, who was on the
   call and their roles, competitors named, objections raised, and any
   timeline the customer committed to.
   
   Everything else a model can find stays in the conversation tool. A CRM
   record with twenty new AI fields is one where nobody reads any of
   them.

2. 2

   Match the call to the right opportunity

   A misattributed call is worse than an unattributed one.

   Headless skills build-workflow

   Copy

   Use build-workflow. Match each call to an opportunity by the calendar
   invite first, then by participant email domain to account, then by the
   rep and timing.
   
   Where several open opportunities exist on one account, do not guess. Ask
   the rep, once, in the channel where they already work.
   
   A call attributed to the wrong opportunity is worse than one attributed
   to none, because it silently pollutes a record somebody is forecasting
   from.
   
   Record the attendees against the opportunity contacts, creating contacts
   where somebody new joined. That alone is worth the integration: the new
   person on the call is frequently the actual decision maker and almost
   never gets added by hand.

3. 3

   Link every claim to its moment

   An insight nobody can check is one nobody should act on.

   Copy

   For every extracted claim, store a deep link to the point in the
   recording it came from, with a short verbatim quote.
   
   Write the claim and the evidence together. A field saying the customer
   mentioned a competitor is an assertion; the same field with the sentence
   and a link is something a rep can act on and a manager can check.
   
   Where extraction confidence is low, say so on the field rather than
   writing it flatly. A hedged claim a person can verify is useful. A
   confident wrong claim gets believed and forecast on.

4. 4

   Gate the commercial fields

   A model should not be setting a forecast.

   Headless skills tray-gotchas

   Copy

   Use tray-gotchas, then split what gets written:
   
   Written automatically, because they are factual and low risk:
     Attendees and their roles
     Next steps with owners and dates
     A call summary on the timeline
     Whether a call happened at all, which fixes activity data
   
   Proposed and confirmed, never written directly:
     Competitor present
     Close date or timeline
     Forecast category or deal risk
     Budget or authority signals
   
   Send the proposals to the rep in one message with one click to accept or
   correct. A close date set by inference is a forecast nobody can defend in
   a board meeting, and the first time it is wrong the whole integration
   loses credibility.

5. 5

   Measure whether it changed anything

   Otherwise it accumulates fields nobody reads.

   Headless skills tray-patterns

   Copy

   For each field, record whether it was subsequently viewed, edited or
   acted on.
   
   Report monthly: acceptance rate on proposed fields, correction rate,
   fields never opened, and deals where an extracted signal preceded a stage
   change.
   
   Delete any field with a low acceptance rate or no usage. An extraction
   nobody trusts is worse than none, because it makes people distrust the
   ones that are good.
   
   Also report calls with no match, trended. A rising number means the
   matching is broken, not that fewer calls happened.

6. 6

   Validate, then hand the extraction set over

   Because what is worth extracting changes with the sales motion.

   Copy

   Run the per-step schema checks and the whole-workflow audit before this
   touches production.
   
   Then open the same workflow in Tray Build so revenue operations can
   change which fields are extracted and which are gated in the visual
   canvas. The right set changes as the sales motion does, and a field that
   stopped being useful should be removable in an afternoon.

What it connects to

The conversation happens in one tool and has to reach the record without overwhelming it.

Named systems are the ones most teams run, not the only ones that work. Each is an authentication in your Tray workspace, referenced by name, so the workflow never holds a credential. Where we have a connector page, the name links to it.

Questions people ask

Further reading

Background on the same subject, for the case rather than the build.

• Edge out competitors by automating conversation intelligence

Related guides