n8n + Claude API AI Research Agent

This workflow uses Claude API (Anthropic's Claude 3.5 Sonnet) as the primary analysis engine combined with Perplexity API for real-time web research, all orchestrated through n8n's visual workflow builder. The agentic reasoning step occurs when Claude evaluates research results against the user's stated criteria and decides which sources are credible, which data points contradict each other, and what additional research queries are needed to fill gaps. This is not a simple search-and-summarize loop: the system maintains a research state across iterations, tracks which questions are answered and which remain open, and only stops when it has enough evidence to produce a report. Output reports are written to Google Drive for internal reference and HubSpot CRM is updated with structured company intelligence. Teams using this pipeline complete research briefs in 2-3 hours that previously took 12-18 hours of manual work — a 70-80% time reduction.

BUSINESS PROBLEM

Sales and research teams spend enormous time on manual research before client engagements. A prospect evaluation that requires checking recent news, competitor positioning, financial health signals, and technology stack typically consumes 4-8 hours per company. For a team evaluating 5-10 prospects per week, that is 20-80 hours of research time — before any selling happens. HubSpot's 2025 Sales Enablement report found that sales reps spend only 34% of their time actually selling; the rest goes to research, data entry, and administrative tasks (Source: HubSpot Sales Enablement Report, 2025). The cost is direct: a sales rep billing $150/hour who spends 66% of their week on non-selling activities represents $49,500/year in lost revenue capacity per rep. For a team of 10, that approaches $500,000 annually. This pipeline automates the research layer entirely, letting reps spend their time on relationship building and deal strategy.

WHO BENEFITS

B2B sales development teams evaluating 20+ target accounts per week who currently spend 3-4 hours per account on manual research across Crunchbase, LinkedIn, news sites, and company blogs. This pipeline cuts research time to 15-25 minutes per account. Market research analysts who produce weekly competitive intelligence briefs for internal stakeholders — this pipeline cuts research time from 15 hours to 2 hours per brief. HubSpot CRM power users who want their contact and company records automatically enriched with AI-researched intelligence without manual data entry or third-party enrichment tools.

HOW IT WORKS

1. Research Trigger: n8n receives a company name and research topic via webhook, HubSpot deal creation, or Google Form submission. Input: company name, industry, specific research questions. Output: structured research brief object in n8n.
2. Document Ingestion: If the user attaches documents (PDFs, previous reports, internal notes), n8n's Google Drive node reads the files and extracts text. Input: file IDs from Google Drive. Output: concatenated document text.
3. Initial Research Query: n8n sends the company name and research questions to Perplexity API with sonar-pro model for real-time web search. Perplexity returns grounded results with citations. Input: query object with company name and max_results. Output: JSON array of results with citations, relevance scores, and source URLs.
4. Deep Analysis (Agentic Reasoning): Claude API receives the Perplexity results plus any ingested documents. It evaluates each result for credibility (source authority, recency, factual consistency), identifies contradictions, and determines which research questions remain unanswered. Output: structured JSON with findings, confidence scores, and gap analysis.
5. Follow-up Query Generation: If gaps exist (confidence below 70% on any key question), Claude generates 2-3 follow-up queries for Perplexity. n8n loops back to step 3 with the new queries. Output: iterative research refinement, max 3 loops.
6. Report Generation: Claude synthesizes all findings into a structured research report with sections: company overview, recent developments, competitive positioning, technology stack, risk signals, and strategic recommendations. Output: formatted report text.
7. CRM Update: n8n's HubSpot node updates the company record with research findings: custom properties for funding status, technology stack tags, competitor mentions, and a link to the full report in Google Drive. Output: updated HubSpot record.
8. Storage and Notification: The report is saved to Google Drive as a Google Doc in a shared research folder. n8n sends a Telegram or Slack notification with a summary and link. Output: Google Drive file ID and notification message.

TOOL INTEGRATION

n8n: The orchestrator. Connects all APIs via visual nodes. Self-hosted (Docker) or cloud (app.n8n.cloud). Required: n8n v2.14+ for native Perplexity node. Gotcha: n8n's loop functionality for iterative research (step 5) requires the Loop Over Items node, which is not visually obvious — most users try to use the Split In Batches node instead, which does not support conditional exit.

Claude API: The analysis and report generation engine. API key from console.anthropic.com. Model: Claude 3.5 Sonnet (best balance of speed and depth for research). Rate limit: 80 RPM on API, 200K context window. Gotcha: Claude's system prompt is critical for research quality — you must instruct it to cite specific sources from the Perplexity results, not from its training data, or it will hallucinate sources.

Perplexity API: The real-time research source. API key from perplexity.ai/settings/api. Model: sonar-pro for deep research with citations. Rate limit: 100 requests/hour on Pro tier. Gotcha: Perplexity's sonar-pro model has a 10-result limit per query — for comprehensive research, run 3-4 parallel queries with different angles instead of one broad query.

Google Drive: Document storage. OAuth 2.0 credentials from Google Cloud Console. Scopes: drive.file (create and read files). Gotcha: The n8n Google Drive node creates documents under the authenticated user's Drive, not a shared Drive, unless you explicitly set the driveId parameter to the shared drive ID.

HubSpot: CRM destination. API key from HubSpot Settings. Required scopes: crm.objects.companies.write and crm.objects.companies.read. Gotcha: HubSpot's custom properties for storing research data must be created in advance — the n8n node does not create them on the fly.

ROI METRICS

1. Prospect research time: 3-4 hours per company manual → 15-25 minutes automated, measurable from first run.
2. Research depth (sources consulted): 5-8 manual sources → 20-40 sources across Perplexity + internal docs.
3. CRM data entry: 20-30 min per prospect for manual enrichment → fully automated at deal creation or stage change.
4. Weekly research capacity: 5-8 companies per week per analyst → 20-30 companies with the automated pipeline.
5. Cost at $150/hr analyst rate: $600-1,200/week manual → $40-80/week in API costs.

CAVEATS

1. Source quality variance: Perplexity returns results from the open web, including low-authority blogs and unverified forums. Claude's credibility filter helps but is not perfect. Add a domain allowlist in the Perplexity query parameters.
2. Loop cost risk: The iterative research loop (step 5) can trigger 9 total API calls (3 loops x 3 queries each). Without a cost cap, a broad research topic could cost $15-20 in a single run. Set n8n's Loop Over Items to max 3 iterations.
3. CRM data overwrite: If the pipeline runs twice on the same company, it overwrites the previous research fields in HubSpot. Use n8n's IF node to check for existing data before updating.
4. This workflow does NOT handle human-curated expert interviews or proprietary paid database access (e.g., Gartner, Forrester). It synthesizes publicly available information only.