Sales Prep Agent for Client Calls

The problem

Reps were spending 20–45 minutes preparing for discovery calls. The prep was valuable — knowing the company, understanding recent context, having a hypothesis about pain — but the time cost was unsustainable as call volume grew.

The goal: compress prep time to under 5 minutes without reducing quality.

What the agent produces

Given a prospect’s company name and contact name, the agent returns a structured brief:

• Company overview (size, funding, products, key customers)
• Recent news and signals (last 90 days)
• Tech stack indicators (if inferable)
• Likely pain points based on company stage and category
• 3 suggested opening questions
• Known context from HubSpot (prior touches, notes)

Total generation time: 90 seconds to 3 minutes depending on available data.

Architecture

Sample output

CALL BRIEF: Thornfield Analytics
Contact: Marcus Webb, VP of Product
Date: November 12, 2024

─────────────────────────────────

COMPANY
Series B ($22M), 80 employees. B2B SaaS — product analytics
for e-commerce brands. Founded 2020. Customers include
Gymshark, Allbirds, several mid-market DTC brands.

RECENT SIGNALS (last 90 days)
• Announced partnership with Shopify Plus (Oct 14)
• 3 open roles in data engineering — scaling infrastructure
• CEO published a post on "the measurement problem in DTC"

TECH SIGNALS
React frontend. Likely AWS. GitHub activity suggests active
product development. No obvious analytics incumbents visible
on job descriptions (may be evaluating alternatives).

LIKELY PAIN
Growing data infrastructure complexity post-Series B.
Engineering bottleneck on analytics requests from product
and go-to-market teams. Shopify partnership suggests
they're expanding surface area — more data sources to unify.

CRM CONTEXT
1 prior touch (cold email, no response, Aug 2024).
Tagged as "warm inbound" — Marcus downloaded the benchmark
report via LinkedIn ad on Oct 28.

─────────────────────────────────

SUGGESTED OPENING QUESTIONS
1. "You published something recently about the measurement
   problem in DTC — what's driving that thinking right now
   at Thornfield?"
2. "The Shopify Plus integration is interesting timing —
   is that expanding the data sources your team is managing,
   or simplifying them?"
3. "What does analytics infrastructure look like on your
   team right now — are product and data working from the
   same stack?"

Implementation notes

The parallel data gathering step runs three requests simultaneously:

• Exa search for company background (3–5 results, contents enabled)
• Exa search for news (time-bounded to last 90 days)
• HubSpot API call for contact and company record

The synthesis step runs after all three complete. The pain hypothesis and talk tracks are generated in a single Claude call using the assembled context, which keeps latency low and maintains coherence across the output.

Reliability

The brief is reliable when the company has a meaningful web presence and the contact has a LinkedIn profile. For smaller companies or private contacts, Exa returns less — the brief is shorter and the pain hypothesis is more generic.

One signal worth noting: the “tech signals” section is consistently the least accurate. Inferring tech stack from job postings and public pages is inherently noisy. Reps are instructed to treat this as a hypothesis to probe, not a fact to state.

Adoption

Reps use this before every discovery call and about 60% of demos. Feedback after 60 days: the “recent signals” and “suggested opening questions” sections are the most used. The CRM context section is redundant for reps who already reviewed the HubSpot record — candidates for removal in a future iteration.