Sales Intelligence · Azure AI Foundry · Claude Tool Use

Seller Intelligence Agent —
Automated Prospect Research

A curated multi-section prompt framework that transforms a single company name into a comprehensive 13-section technology diligence briefing in minutes — with live-data tool calling, seller-aligned insights, and a structured output designed for account executives.

13
Report Sections
5
Live Data Tools
2
Agent Modes
4
News Sources
~8min
Full Report Time
0
Manual Clicks

Seller research: the highest-friction part of every sales cycle

Before an enterprise account executive walks into a first call, they need deep knowledge of the prospect's technology stack, financial signals, strategic priorities, and executive landscape. Getting that right requires hours of research — which most sellers don't have time to do well.

⏱️

2–4 Hours of Manual Research per Prospect

An AE preparing for a first call with a Fortune 500 company needs to check LinkedIn, 10-K filings, recent news, job postings, technology press, and the company's own website. Doing it right takes half a day. Most sellers shortcut it.

🎲

Inconsistent Quality — Analyst-Dependent

Some sellers are excellent researchers; most aren't. The quality of pre-call intelligence varies wildly across a sales team. There's no repeatable, standardized output format that every seller uses to walk into a meeting prepared the same way.

🎯

Generic Output That Doesn't Map to What You Sell

Generic AI summaries give you a company overview. They don't tell you which of your service lines has the highest probability opportunity, what the evidence is, and what the specific entry point conversation should be. The connection from research to sales motion is left as an exercise for the reader.

📊

No Financial Signal Integration

Manual research often misses real-time signals: a stock at a 52-week low signaling cost pressure, a CFO change six months ago, a recent acquisition that just created a cloud sprawl problem. These are the conversation openers that make cold calls warm.

🗺️

No Stakeholder Mapping

Even well-researched briefings rarely include a stakeholder map: who is the economic buyer, who is the technical champion, who controls procurement, who will block the deal. This information exists in public sources — but synthesizing it takes expertise.

📋

Reports That Sit in Inboxes Unread

When research teams do produce briefings, they're often 20-page PDFs that sellers don't read before calls. The format isn't built for how sellers actually consume information — skimmable sections, prioritized opportunities, clear next actions.


Two specialized agents — same infrastructure, different jobs

The platform ships two prompt configurations deployed on the same Azure AI Foundry infrastructure. Each mode has a distinct system prompt, a different output structure, and different tool orchestration rules tuned to the research task.

PE Intelligence Mode

Corporate Intelligence Analyst

General-purpose pursuit intelligence for consulting and advisory teams. Analyzes any company for technology transformation opportunities, cloud maturity gaps, AI/ML adoption signals, data platform needs, and integration challenges. Produces a layered research brief that feeds into proposal and business case development.

  • Open-ended exploration — any company, any industry
  • Technology debt and legacy system identification
  • Cloud maturity and migration opportunity signals
  • AI/ML adoption signals and data readiness
  • Integration architecture observations
  • Conversational follow-up after initial brief
Seller Diligence Mode (RXT)

Technology Sales Intelligence

Rackspace-specific seller enablement tool. Every section ends with an explicit RACKSPACE SELLER INSIGHT that names the specific service line, the evidence for the opportunity, the recommended entry point, and the competitive risk. Output is a complete 13-section briefing with a Top 3 Focus Areas summary and full Citation Index.

  • 13 sections, produced in one pass without interruption
  • Every finding mapped to a Rackspace service area
  • Stakeholder map: buyer, champion, technical evaluator, procurement
  • Top 3 Focus Areas with opportunity, evidence, and entry point
  • For internal pre-call prep — not for sharing with prospects
  • Explicit anti-dithering rules prevent the agent from stalling

A structured prompt is a software specification for AI behavior

The agent's quality is not primarily determined by the model — it's determined by the prompt. The Seller Diligence prompt is a precision-engineered specification covering role definition, execution rules, tool orchestration, output schema, and quality constraints. Each layer solves a specific failure mode observed in unstructured AI output.

RXT Seller Diligence System Prompt v3.1 — Anatomy
Layer 1
Context Declaration
Who the agent is, who it works for, what it sells
SELLER: Rackspace Technology · Service areas explicitly enumerated: Cloud, Data, AI, Security, Infrastructure, Applications, Advisory, FinOps, Managed Services.

Declaring the seller's service portfolio upfront means every downstream finding is automatically evaluated against these specific categories — not generic cloud-speak. The agent knows what it can sell before it reads a single article.
Layer 2
Anti-Dithering Rules
Prevents the #1 LLM failure mode in structured reports
Seven numbered execution rules prevent the agent from pausing, asking clarifying questions, or stopping between sections. Drawn verbatim from the actual prompt:

DO NOT ask clarifying questions DO NOT pause between sections DO NOT summarize what you're about to do DO NOT say "Shall I continue?"

Without these rules, LLMs frequently stop after 2–3 sections and ask "Would you like me to continue?" — destroying the automated single-pass execution model.
Layer 3
Output Schema
Structural contract for every section
Every section follows an identical 4-part structure:

3-5 specific findings in plain language Citations in brackets after every claim RACKSPACE SELLER INSIGHT callout SOURCES block with full URLs and dates

This means the output is predictable and parseable. The frontend renders each section as a collapsible card. The SELLER INSIGHT is highlighted as a distinct callout box — immediately visible to the AE scanning before a call.
Layer 4
Tool Orchestration Rules
Prevents token waste and ensures speed
Each tool has explicit usage rules built into the prompt — not the tool description:

get_company_news: max=5 per call 2–3 targeted queries, do NOT repeat similar queries
fetch_company_website: at most 2 pages — IR page first, then newsroom
get_stock_quote: call once only — Section 04 (Financial Profile) only, if publicly traded
get_stock_history: skip entirely — not needed for seller diligence
generate_custom_report: call IMMEDIATELY when user says "create" — no description, just call it

These rules prevent the agent from making 15 news calls (expensive, slow) and from ignoring tool calls when document generation is requested.
Layer 5
Accuracy Contract
Governance against fabrication
"Only report facts from source data — never fabricate."

If data is unavailable for a subsection, the agent is instructed to state "Limited public data available" and continue — rather than inventing plausible-sounding content. This is the same anti-hallucination principle as the migration platform's provenance model, applied to a research agent: the instruction doesn't eliminate hallucination risk entirely, but it creates an explicit behavioral norm that the model follows significantly more reliably than an unconstrained prompt.
Layer 6
Closing Deliverable
Seller-actionable synthesis
After the 13 sections, the prompt specifies two mandatory closing deliverables:

TOP 3 RECOMMENDED FOCUS AREAS — each with: The Opportunity, The Evidence, Why Now, Entry Point, Competitive Risk, The Blocker, Sources.

CITATION INDEX — consolidated reference list with complete https:// URLs for every source cited across all 13 sections.

The Top 3 is what the AE reads before the call. The Citation Index is what the pre-sales architect reads before the proposal. Both are generated automatically.
Section 01
Strategic Profile
Company overview, mission, recent strategic moves, M&A activity
Section 02
Leadership Profile
C-suite bios + Stakeholder Map: Economic Buyer, Champion, Technical Evaluator, Procurement
Stakeholder Map
Section 03
Industry Profile
Industry dynamics, competitive landscape, regulatory environment
Section 04
Financial Profile
Revenue, growth, margins, stock performance, capex signals
Live stock data
Section 05
Technology Profile
Known tech stack, cloud providers, SaaS footprint, engineering scale
Section 06
Tech Stack & Legacy Risks
Legacy system signals, technical debt indicators, modernization triggers
Section 07
Multi-Cloud & Hybrid Infra
Data center inventory, cloud provider mix, infrastructure contracts
Section 08
Data Platform & AI Readiness
Analytics maturity, data lake/warehouse signals, AI/ML adoption stage
Section 09
Security & Compliance
Security incidents, compliance posture, regulatory exposure, zero-trust signals
Section 10
FinOps & Cost Optimization
Cloud cost pressure signals, optimization opportunities, CFO initiatives
Section 11
Managed Services Appetite
Outsourcing signals, IT org headcount trends, vendor consolidation moves
Section 12
Procurement Profile
Procurement cycle timing, contract renewal windows, buying signals
Section 13
Shift & Investment Trajectory
Strategic investment direction, budget allocation signals, technology roadmap
Closer
Top 3 Focus Areas
Opportunity · Evidence · Why Now · Entry Point · Competitive Risk · Blocker
Index
Citation Index
Full https:// URLs for every source cited in every section

Claude uses five tools during report generation. Each tool has explicit usage rules in the system prompt — preventing token waste, controlling cost, and ensuring the right data sources are used for each section.

📰
get_company_news
NewsAPI · GNews · Bing · MediaStack
Primary research tool. Aggregates from 4 news APIs, deduplicates by URL and title similarity, sorts by recency. Claude makes 2–3 targeted queries with distinct keywords (e.g., "CompanyName cloud AI", "CompanyName security breach", "CompanyName earnings CFO") — not one broad query.
max=5 per call · 2–3 calls total · no repeated queries
🌐
fetch_company_website
BeautifulSoup · 10KB text extract
First-party intelligence. Fetches and extracts text from investor relations pages, newsrooms, and technology pages. Strips nav/footer/scripts. Truncated to 10KB to prevent payload bloat. Cached in Cosmos DB for 24 hours so subsequent calls within a day don't re-fetch.
2 pages max per report · IR page first, then newsroom
📈
get_stock_quote
Alpha Vantage · 5-minute cache TTL
Financial signal. Called once, for Section 04 (Financial Profile) only. Real-time stock price, change, volume. Cached in Cosmos DB with 5-minute TTL to prevent API rate limit exhaustion on repeated research sessions for the same company.
call once only · publicly traded companies only
📄
generate_custom_report
Blob Storage · SAS URL (365-day)
Document export. When the user says "create a report", "export this", or "make a markdown file", Claude calls this tool immediately — without describing what it will do. Generates markdown, PDF, or JSON. Stored in Azure Blob Storage with a SAS URL valid for 365 days, returned to the frontend for download.
call immediately · no description first · mandatory on export keywords
Acme Corporation — Seller Diligence Brief Generated 2026-06-11 · 13 sections · RXT v3.1
Section 02 — Leadership
Section 07 — Infra
Top 3 Focus Areas
02 LEADERSHIP PROFILE & STAKEHOLDER MAP

CTO — Jane Mitchell (appointed March 2024) joined from AWS where she led the Enterprise Migration team. Her hire is a strong signal of cloud-first intent. [Acme Press Release, Mar 2024]

CFO — Robert Chen (since 2022) has made public comments on "optimizing IT infrastructure costs" in three consecutive earnings calls, citing cloud spend as a target area. [Q3 2025 Earnings Call Transcript]

Acme posted 14 open roles tagged "cloud infrastructure" and "AWS" in Q1 2026, concentrated in the new CTO's engineering org — suggesting active build vs. buy evaluation. [LinkedIn Jobs, Apr 2026]

VP Infrastructure — David Park previously ran managed services procurement at a prior employer and has a public track record of vendor consolidation. [LinkedIn Profile]

Stakeholder Map
Economic Buyer
Robert Chen, CFO
Controls IT budget. Focused on cost reduction — reachable via FinOps angle.
Champion
Jane Mitchell, CTO
Cloud-first agenda. Will benefit from managed services freeing up her team.
Technical Evaluator
David Park, VP Infra
Procurement-experienced. Will run vendor evaluation. Show differentiation early.
Procurement/Legal
Unknown — research needed
Limited public data available. Flag for discovery call.
Rackspace Seller Insight · Advisory & Managed Services

CTO background in cloud migration + CFO cost pressure + VP Infra's procurement track record = classic managed services motion. Open cloud infrastructure roles signal active investment with limited internal expertise. Lead with a FinOps assessment offer to CFO; position managed cloud to CTO as a talent multiplier, not a cost — she needs her team on innovation, not ops.

Report generation takes 6–10 minutes as Claude makes multiple tool calls. The frontend polls GET /api/v1/chat/progress/{request_id} at 500ms intervals. As each tool call fires, an event is written to _progress_store in memory, keyed by a UUID sent at request start. The UI renders live status cards: "📰 Fetching news — CompanyName cloud AI…", "🌐 Reading investor relations page…".

This transforms the experience from "blank screen for 8 minutes" to a visible, trustworthy research process that sellers watch unfold in real time.
Every completed report is stored in Cosmos DB with the company name as partition key. The sidebar lists prior research sessions — sellers can return to a previous brief without re-running the full pipeline.

API responses are also cached per-container with TTLs: stock-data (5 min), news-articles (1 hour), website-content (24 hours). A second research session on the same company within the cache window uses cached data — reducing API costs and latency.
The get_company_news endpoint fans out to four sources in parallel: NewsAPI, GNews, Bing News Search, and MediaStack. Results are deduplicated by URL (exact match) and title similarity (first 50 characters), then sorted by publishedAt descending. The resulting set is the most recent unique articles across all sources — not just what one API returned.

Claude receives a max of 5 articles per query call (enforced in the prompt), but each article from multiple sources covers different angles of the same story — giving broader signal than a single-source news query.
A single POST /api/v1/chat endpoint handles both agents. The request body includes a prompt_type field ("pe" or "rxt"), which routes to either PE_SYSTEM_PROMPT or RXT_SYSTEM_PROMPT. Both use the same tool definitions, the same Cosmos DB cache layer, and the same Azure AI Foundry model — but produce structurally different outputs because the system prompt specifies a completely different output schema and section sequence.

This architecture makes it trivial to add new agent modes (e.g., a due diligence mode for M&A targets) without new endpoints — just a new system prompt and a routing key.

What automated seller research changes

The primary ROI is time recovered. But the secondary effect — consistency and depth — is often more valuable: every seller on the team walks into every first call with the same quality intelligence, regardless of their individual research skills.

Before — Manual Pre-Call Research
Time per prospect brief2–4 hours
Sections consistently covered3–5 (informal)
Stakeholder mapRarely produced
Financial signals integratedInconsistent
Service line recommendationSeller's intuition
Citation indexNone
Quality consistency across teamHigh variance
After — Seller Intelligence Agent
Time per prospect brief~8 minutes
Sections consistently covered13 sections + Top 3 + Index
Stakeholder mapSection 02, every time
Financial signals integratedSection 04 with live data
Service line recommendationEvidence-backed, every section
Citation indexFull URLs, every source
Quality consistency across teamUniform — prompt-driven
95%
Time Reduction
8 minutes vs. 2–4 hours. An AE running 5 new prospect meetings per week recovers 10–20 hours per week — or can run 5× more first calls with the same prep time.
100%
Coverage Consistency
Every seller, every prospect, every call — 13 sections, stakeholder map, and Top 3 recommendations. Quality doesn't depend on who on the team runs the research.
4
News Sources, One Query
NewsAPI, GNews, Bing News, MediaStack — aggregated, deduplicated, and ranked by recency. A single company name covers the full public signal landscape.
0
Clicks to Generate
Type the company name. Hit enter. Eight minutes later, a complete downloadable briefing is in front of the seller, formatted for pre-call consumption.

Prompt Engineering as Product Design

The agent's intelligence isn't in the model — it's in the specification. A well-engineered system prompt is a behavioral contract that produces consistent, structured, seller-aligned output at scale.

Azure AI Foundry · Claude Sonnet 4.5 · Azure Functions · Cosmos DB · Python 3.11