Drafts Ready Before The Rep Gets Back to the Car
The single biggest source of pipeline leakage in field sales is the follow-up that never gets sent. A rep finishes a visit, writes a quick note, drives to the next stop, and the follow-up email gets bumped to "tonight" — which becomes "tomorrow" which becomes "I forgot". Velocity X's AI Brain solves this by drafting the follow-up the moment the rep checks out of the visit, so the rep just reviews and sends.
What AI Brain Does, Exactly
The moment a rep marks a visit complete, Brain runs three jobs in parallel:
1. Drafts a follow-up email. Pulls the visit notes, the deal history, the prospect's industry/role from the CRM, and the brand voice rules from brand.json. Outputs a 4-8 sentence email written in the rep's voice (Brain learns from the rep's previous outgoing emails), referencing specific details from the visit. The rep gets a notification: "Brain has a draft ready — review and send?"
2. Updates the deal summary. Adds a one-paragraph summary of the visit to the deal record. What was discussed, what's next, who promised what. The summary syncs back to the CRM via the bidirectional sync layer.
3. Queues the next best action. Brain reads the deal state and proposes the next action — "send the quote", "book a follow-up call", "wait 2 weeks then check in". The action lands in the rep's task list, scheduled appropriately.
The Prompt Design
Brain's draft quality comes from prompt engineering. The system prompt includes:
- The rep's name, role, and a few example outgoing emails (learned over time).
- The prospect's company, role, industry, and any notes from the CRM.
- The voice rules from brand.json (phrases to lean on, phrases to avoid).
- The visit notes the rep entered.
- The deal history (previous emails, previous visits, current stage).
- A few-shot example of a good follow-up email for this deal type.
The output is constrained — Brain is prompted to return JSON with subject, greeting, body, signature, and a confidence score. If confidence drops below 0.7, the rep sees a "Brain has a low-confidence draft — review carefully" warning. Above 0.9, Brain offers a one-tap send option.
Which Model
Velocity X uses Claude Sonnet for the default Brain model. Sonnet hits the sweet spot: fast enough to deliver a draft in 4-8 seconds, capable enough to produce on-brand prose that matches the rep's voice, and cheap enough at scale that the per-draft cost is rounding error (~$0.003 per follow-up).
For high-stakes deals (deal value above a configurable threshold, e.g. $50k+), Brain auto-upgrades to Claude Opus for the draft. Slower, more expensive, materially better at nuanced messaging. The cost of an Opus draft is roughly $0.04 — trivial vs the deal value.
How It Stays On-Brand
This is the brand.json connection. Brain reads brand.json on every invocation. The voice section there defines:
{`"voice": {
"tone": "professional, helpful, customer-first",
"phrasesToAvoid": ["leverage", "synergy", "circle back"],
"phrasesToUse": ["sorted", "happy to help", "let me know"],
"signOff": "Cheers, [rep name]"
}`}The phrases-to-avoid list goes into the prompt as negative instruction. The phrases-to-use list goes in as positive examples. The sign-off becomes the default closing. After a session or two of rep edits, Brain learns the rep's personal variations on top of the brand defaults.
Privacy & Data Handling
Brain sends the visit notes + CRM context to Anthropic's API. The Anthropic API doesn't train on customer data (this is contractually guaranteed under their commercial terms). Drafts are stored in your Supabase database under your control. If you delete a deal, the associated drafts get cascaded.
For customers in regulated industries (healthcare, legal, financial advice), Brain can be configured to run against an on-premises model instead — typically Llama 3 70B locally via Ollama. The trade-off is draft quality (Llama-class models produce roughly 70% as good drafts as Claude Sonnet) for full data residency.
Frequently Asked Questions
Can the rep edit the draft before sending?
Yes. Drafts open in an editable interface. Most reps make minor edits and send. Power users sometimes use the draft as a starting structure and rewrite half of it. Either is fine — Brain learns from the edits over time.
What if Brain gets the tone wrong?
Update brand.json's voice section. If a specific phrase shouldn't appear, add it to phrasesToAvoid. If a phrase should appear more often, add it to phrasesToUse. Brain reads brand.json on the next invocation. No retraining needed.
How much does AI Brain cost to run?
For a 10-person field team doing 30 visits/day each (300 visits/day, 6,000/month), at Sonnet's ~$0.003 per draft, the API cost is roughly $18/month. With occasional Opus upgrades for high-value deals, maybe $25-$30/month total. Cheap relative to the value of the follow-ups it actually gets sent.
Can it draft in languages other than English?
Yes. Set the brand.json voice section's language hint and Brain drafts in that language. Tested in English, Spanish, German, and French. Other languages work in theory; quality may vary.
Does it work for cold outreach too?
Yes. Brain has a separate "cold" mode that uses a different prompt template emphasising first-contact framing. Still pulls voice rules from brand.json. Quality is generally lower than follow-up drafts because there's less context, but still useful as a starting point.
What about other AI models — GPT-4o, Gemini?
Brain is model-agnostic at the architecture level. The default is Claude because Anthropic's combination of speed, cost, and tone-matching is the best fit for sales prose in our testing. Customers can swap to GPT-4o or Gemini by changing one config value if they prefer.
The Bottom Line
AI Brain follow-ups turn "I'll send the email tonight" into "send" before the rep gets back to the car. Drafts are on-brand because Brain reads brand.json. Drafts are personal because Brain learns from the rep's outgoing emails. Drafts are accurate because Brain has access to the deal history and visit notes. Bundled in Velocity X's buy-once price — no separate AI subscription.