AI Agent Workflow Automation Examples

An AI agent workflow is a multi-step automated process where at least one step involves an AI agent making a contextual decision — not just executing a rule. This distinction matters. A rule-based automation says "if the ticket says 'urgent,' set priority to High." An AI agent workflow says "read this ticket, assess the actual urgency based on content, customer tier, and history, then set the appropriate priority."

The following seven examples are operational AI agent workflows drawn from B2B and B2C production environments. Each one is broken down step by step, with the tools operating at each stage and the outcomes measured after deployment.

For background on how individual AI agents operate, see What Are AI Agents?. For examples of multi-agent collaboration rather than single-agent workflows, see Multi-Agent System Examples in Production. The full business context is at AI Agent Examples in Business.

---

Workflow 1: New Lead → Enrich → Score → Assign → Notify#

Context: A B2B SaaS company generating 800-1,200 inbound leads per month from paid ads, content, and events. Lead response time was averaging 5.2 hours. Leads were assigned manually by the marketing ops manager.

The workflow:

Step 1 — Capture (Typeform/HubSpot) A new lead submits a form on the website. HubSpot creates a contact record and fires a webhook to the workflow orchestration layer (n8n).

Step 2 — Enrich (Clearbit + LinkedIn via Proxycurl) The enrichment agent receives the email address and company name. It calls:

• Clearbit API: returns company size, industry, tech stack, annual revenue estimate
• Proxycurl API: returns the contact's LinkedIn title, seniority level, department
• ZoomInfo: supplements with direct dial if available

Output: A fully enriched contact record pushed back to HubSpot.

Step 3 — Score (Custom ICP Scoring Agent) An LLM-powered scoring agent evaluates the enriched record against the company's ICP matrix:

• Company size fit (employees 50-500 = high score; fewer than 10 or more than 5,000 = low)
• Industry fit (SaaS, fintech, healthcare = high; government, nonprofit = low)
• Role fit (VP+ in Operations, Revenue, or IT = high; intern, student = disqualify)
• Technology stack signals (uses Salesforce + Slack = high fit; uses no CRM = low)

Output: ICP score 1-100, tier assignment (A/B/C/Disqualify), and a one-sentence rationale.

Step 4 — Assign (Routing Agent) The routing agent checks the assigned ICP tier and maps to the correct owner:

• Tier A (score 75+): Assigned to a senior AE based on territory and current workload (pulled from Salesforce)
• Tier B (50-74): Assigned to an SDR based on capacity
• Tier C (25-49): Enrolled in a long-cycle nurture sequence in HubSpot
• Disqualify: Marked as unqualified in HubSpot, no human assignment

Step 5 — Notify (Slack + Email) For Tier A and B assignments, the agent sends:

• A Slack DM to the assigned rep: "New [Tier A] lead: [Name] at [Company]. Score: 84/100. [Company] is a 200-person fintech firm using Salesforce. They downloaded the ROI calculator. Suggested opener: [2-sentence AI-drafted opener based on their tech stack and role]"
• A task in Salesforce to contact within [X hours based on tier]

Tools at each step: HubSpot (CRM), n8n (orchestration), Clearbit (enrichment), Proxycurl (LinkedIn data), ZoomInfo (contact data), Salesforce (AE assignment and task creation), OpenAI GPT-4o (scoring and message drafting), Slack (rep notification).

Outcomes:

• Average lead response time: reduced from 5.2 hours to 18 minutes
• ICP scoring accuracy: 89% alignment with human sales judgment (measured by sampling 100 leads and comparing agent score to AE assessment)
• Rep satisfaction: improved — reps received context with every lead, not just a name and email
• Tier A conversion rate (lead to opportunity): increased 24% — attributed to faster response and better-prepared first contact

---

Workflow 2: Support Ticket → Triage → Attempt Resolution → Escalate if Needed#

Context: A cloud software company receiving 2,800 support tickets per week. 60% of tickets were tier-1 (resolvable from documentation). Human agents were handling all tickets, including the straightforward ones.

The workflow:

Step 1 — Receive (Zendesk) A ticket is submitted via email or the support portal. Zendesk fires a webhook to the workflow engine (LangGraph) within 30 seconds.

Step 2 — Triage (Classification Agent) The triage agent reads the ticket subject, body, customer account tier, and the last 3 previous tickets from this customer. It classifies:

• Ticket type (billing, technical bug, feature request, account access, general)
• Estimated complexity (tier 1, 2, or 3)
• Urgency (routine, elevated, critical)
• Sentiment (neutral, frustrated, highly frustrated)

Output: A structured classification object, written back to the Zendesk ticket as internal metadata tags.

Step 3 — Resolution Attempt (RAG Agent) For tier-1 classified tickets, the retrieval-augmented generation resolution agent activates:

1. Embeds the ticket text and runs a similarity search against the 1,400-article knowledge base (Pinecone vector store)
2. Retrieves the top 5 relevant knowledge base chunks
3. GPT-4o synthesizes a resolution response, grounding it strictly in the retrieved documentation
4. A confidence score is computed (0-1) based on retrieval relevance and response completeness

Step 4 — Quality Gate If confidence ≥ 0.78 and ticket is tier-1: the response is sent automatically, the ticket is marked "resolved by agent," and the customer receives a 4-hour follow-up survey.

If confidence < 0.78 or ticket is tier-2/3: proceed to Step 5.

Step 5 — Escalation Prep (Context Builder Agent) For escalated tickets, the context builder agent prepares a human agent brief:

• Customer account summary (plan, tenure, spend)
• Full ticket history on this specific issue type
• Agent's attempted resolution (if any) and why it fell short
• Recommended starting diagnostic step

The brief is attached to the Zendesk ticket. The ticket is routed to the appropriate human agent queue based on ticket type.

Step 6 — Notification (Slack) Human agents are notified via Slack with a direct link to the ticket and a one-line summary. High-urgency tickets trigger a @channel mention in the support team Slack channel.

Tools: Zendesk (ticketing), LangGraph (orchestration), Pinecone (vector store), OpenAI GPT-4o (classification and generation), Slack (escalation alerts), custom Python (confidence scoring).

Outcomes:

• 47% of tier-1 tickets resolved automatically
• Median resolution time for auto-resolved tickets: 4 minutes
• CSAT on auto-resolved tickets: 4.1/5.0
• Human agent escalation context prep time eliminated: 12 minutes per escalated ticket → agents start immediately from a prepared brief
• Support team capacity for complex tickets increased 38%

---

Workflow 3: Job Posting → Screen → Schedule → Notify#

Context: A regional hospital network hiring 80-120 clinical and administrative staff per quarter. The recruiting team of 6 was overwhelmed by application volume — some roles receiving 600+ applications in the first week.

The workflow:

Step 1 — Job Posted (Greenhouse) A new job requisition is published in Greenhouse. The workflow agent receives the job ID via webhook and reads the required qualifications, preferred credentials, and hiring manager preferences from the requisition.

Step 2 — Application Screening (Screening Agent) As applications arrive (polling the Greenhouse API every 15 minutes), the screening agent:

1. Parses the resume using Affinda (structured extraction: education, certifications, years of experience, prior employers)
2. Scores the candidate against the specific job rubric (required RN license = pass/fail gate; years of relevant experience weighted; location proximity for on-site roles)
3. Generates a recruiter brief: 3-4 sentence summary of why the candidate does or does not meet requirements
4. Assigns a tier: Advance, Review, Hold, or Decline

Step 3 — Candidate Communication (Email Agent) Within 90 minutes of application submission:

• Advance candidates: Receive a personalized "your application is advancing" email with a scheduling link (Cronofy-powered availability selector)
• Decline candidates: Receive a respectful decline with an estimated notification timeline (within 24 hours, not weeks)
• Review candidates: Receive an acknowledgment that their application is under review

Step 4 — Interview Scheduling (Scheduling Agent) When an Advance candidate selects their preferred time slot:

1. The scheduling agent checks interviewer availability against Google Calendar
2. Books the slot, generates a Zoom link, and sends calendar invites to the candidate and interviewer
3. Populates a candidate dossier in Greenhouse for the interviewer: screening score, recruiter brief, and 3-5 suggested interview questions based on the job requirements

Step 5 — Reminder and Follow-up

• 24-hour reminder to candidate and interviewer
• 1-hour reminder to both parties
• Post-interview: interviewer receives a Greenhouse prompt to submit feedback within 24 hours
• Candidate receives a follow-up communication within 4 hours of interview: "Thanks for your time — we'll be in touch within X business days"

Tools: Greenhouse (ATS), Affinda (resume parsing), OpenAI GPT-4o (brief generation and email personalization), Cronofy (calendar availability), Google Calendar (interviewer schedules), Zoom (meeting links), SendGrid (email delivery), Zapier (orchestration glue).

Outcomes:

• Recruiter application review time: reduced 62%
• Advance candidate response rate to scheduling link: 81% within 24 hours (vs. 54% with manual scheduling emails)
• Interview scheduling time: reduced from 3.2 days to 6 hours
• Candidate satisfaction score (post-process survey): increased from 3.4 to 4.6/5.0

---

Workflow 4: Invoice Received → Extract → Match PO → Approve or Flag#

Context: A 1,500-person manufacturing company processing 2,200 invoices per month through a manual AP workflow. AP staff were spending 60% of their time on data entry and matching. Invoice processing cycle time was 18 days on average, causing late payment penalties and strained vendor relationships.

The workflow:

Step 1 — Invoice Received (Email / AP Inbox) Invoices arrive as PDF email attachments to a dedicated AP inbox. A Microsoft Power Automate trigger fires when a new email arrives with a PDF attachment. The PDF is extracted and passed to the workflow.

Step 2 — Data Extraction (OCR + Extraction Agent) The extraction agent processes the PDF:

1. AWS Textract performs OCR (handles scanned PDFs, structured forms, and handwritten values)
2. GPT-4o reads the extracted text and populates a structured invoice object: vendor name, invoice number, invoice date, line items with descriptions and amounts, subtotal, tax, total due, payment terms, remit-to information

Output: A validated JSON invoice object. Missing fields are flagged for human review before proceeding.

Step 3 — PO Matching (Matching Agent) The matching agent queries NetSuite (ERP) to find the corresponding purchase order:

1. Searches by vendor name + approximate amount range
2. Retrieves the PO's approved amount, line items, and receiving confirmation status
3. Runs a three-way match: invoice line items vs. PO line items vs. goods receipt confirmation
4. Computes a match confidence score and identifies specific discrepancies (e.g., "Invoice shows $12,400 for Product X; PO shows $11,800 — $600 variance on line 3")

Step 4 — Approval Routing (Decision Agent) Based on match outcome:

• Full match + amount ≤ $25,000: Auto-approve, schedule payment per vendor terms in NetSuite
• Full match + amount > $25,000: Route to finance manager for approval in NetSuite with the matched PO and match summary attached
• Partial match (variance ≤ 3%): Route to AP specialist with discrepancy highlighted and suggested resolution
• Significant discrepancy or missing PO: Flag to AP manager, notify the vendor contact, pause payment

Step 5 — Notification and Audit Log

• Vendor receives payment confirmation email (auto-approved) or "under review" email (flagged)
• AP team receives a daily digest of all invoices processed, approved, and flagged
• All decisions are written to a structured audit log in Google BigQuery for compliance reporting

Tools: Microsoft Power Automate (trigger), AWS Textract (OCR), OpenAI GPT-4o (extraction and matching analysis), NetSuite (ERP for PO data and payment), SendGrid (vendor notifications), Google BigQuery (audit log), n8n (orchestration).

Outcomes:

• AP staff data entry time: reduced 74%
• Invoice processing cycle time: reduced from 18 days to 5 days on average
• Auto-approval rate for clean invoices: 61% of total volume
• Late payment penalties: eliminated entirely in the 6 months following deployment
• Three-way match accuracy: 97.2% (3 errors per 100 invoices required human correction vs. 11 per 100 manual)

---

Workflow 5: Content Request → Research → Draft → Review → Publish#

Context: A B2B technology media company publishing 25-30 articles per week. Their content pipeline had five stages, each done manually by different team members, with content sitting in handoff queues for 1-3 days between stages.

The workflow:

Step 1 — Content Request (Notion) An editor adds a new topic to the Notion content calendar with: target keyword, target persona, publication date, and any specific angle notes. A Notion automation fires a webhook to the content pipeline agent.

Step 2 — Research (Research Agent) The research agent activates immediately:

1. Queries Semrush API for the primary keyword: volume, difficulty, SERP composition, top-ranking URLs
2. Scrapes and summarizes the top 5 ranking articles using Firecrawl
3. Queries the internal Pinecone knowledge base for any existing company content on the topic
4. Searches for relevant statistics and data points using Tavily (filtered to Tier 1 sources: academic, government, major publications)

Output: A structured research packet — 800-1,200 words of organized source material — written back to the Notion page.

Step 3 — Brief and Outline (Strategist Agent) The strategist agent reads the research packet and generates:

• A differentiated angle (what this article will argue or show that the top-ranking articles do not)
• A full H2/H3 outline with a sentence description of each section
• A target word count range
• 3-5 specific things the article must include that competitors miss

The brief is written to the Notion page alongside the research packet.

Step 4 — Draft (Writer Agent) The writer agent generates the full article draft:

• Uses Anthropic Claude 3.5 Sonnet (chosen for long-form writing quality over GPT-4o)
• Grounded strictly in the research packet — no facts introduced beyond what was retrieved
• Follows the brand voice guide (retrieved from Pinecone on each run)
• Produces the draft in Markdown, written to a Google Doc via the Google Docs API

Step 5 — Review (Editor Agent) The editor agent performs a structured review pass:

1. Fact-checks each specific claim against the source documents in the research packet
2. Evaluates the article against an editorial quality rubric: argument clarity, example specificity, SEO element presence, readability, internal link opportunities
3. Checks for any AI-typical patterns that reduce authenticity (hedge phrases, generic openers, unsupported superlatives)
4. Produces a review report: quality score, flagged issues, and line-level suggestions

The Google Doc is updated with editor comments. A Slack notification goes to the human editor with the quality score and issue summary.

Step 6 — Human Final Review and Publish The human editor reviews the editor agent's flags in the Google Doc, makes final adjustments (typically 15-25 minutes), and approves for publication. The CMS publishing step is triggered manually — publication is never automated.

Step 7 — Post-Publish (Distribution Agent) After the human marks the article as published, the distribution agent:

1. Drafts social media posts for LinkedIn, X, and the company newsletter with different angles per platform
2. Identifies 3-5 internal link opportunities in existing articles that should reference the new piece
3. Generates a link suggestion report for the editor to implement
4. Logs the article in the SEO performance tracker (Google Sheets with Semrush ranking data pulled weekly)

Tools: Notion (content calendar), n8n (orchestration), Semrush API (keyword data), Firecrawl (article scraping), Tavily (research search), Pinecone (knowledge base and brand voice), OpenAI GPT-4o (research analysis, strategist, editor), Anthropic Claude 3.5 Sonnet (writer agent), Google Docs API (draft delivery), Slack (human notifications), Google Sheets (SEO tracking).

Outcomes:

• Time-to-publish from topic assignment: reduced from 6 days to 1.5 days
• Human editor time per article: reduced from 5.5 hours to 1.2 hours (research + writing handled by agents; human focuses on judgment and final quality)
• Article output per week: increased from 28 to 44 without additional headcount
• Factual error rate in final published articles: decreased 34% (editor agent catches issues before human review)
• Internal linking coverage: improved from 2.1 links per new article to 5.4 (distribution agent surfaces opportunities systematically)

---

Workflow 6: Alert Received → Classify → Investigate → Escalate or Resolve#

Context: A fintech company running a real-time fraud detection system that generates 400-600 alerts per day. Each alert required a human analyst to investigate — reviewing transaction history, customer behavior, and flagged patterns. The team of 8 analysts was handling 50+ alerts per analyst per day.

The workflow:

Step 1 — Alert Generated (Internal Risk Engine) The company's rule-based fraud detection system fires an alert when a transaction exceeds a risk score threshold. The alert is written to a PostgreSQL database with: transaction ID, customer ID, alert type, risk score, and flagged rule.

Step 2 — Context Aggregation (Investigation Agent) The investigation agent pulls full context for the alert within 90 seconds:

1. Full transaction history for the customer (past 90 days) via internal API
2. Account behavior baseline (typical transaction sizes, frequency, geographic patterns)
3. Device fingerprint data for the triggering transaction
4. IP geolocation and velocity signals (how many transactions from this IP in 24 hours)
5. Any prior fraud flags on the account or linked accounts

Output: A structured investigation brief, written to the alert record.

Step 3 — Decision Agent The decision agent reads the investigation brief and issues a recommendation:

• Low risk (false positive likely): Approve transaction, log rationale, close alert
• Medium risk: Place transaction on 15-minute hold, send customer an authentication push notification
• High risk: Block transaction, freeze account, escalate to human analyst with full context brief
• Critical (confirmed fraud pattern): Block, freeze, notify customer via SMS and email, escalate to senior analyst immediately

Step 4 — Human Escalation (for High/Critical) High and critical alerts are routed to an analyst queue with the full investigation brief pre-populated. Analysts see the decision agent's recommendation and all supporting evidence. Their job is to confirm or override — not to reconstruct the investigation from scratch.

Step 5 — Customer Communication (Communication Agent) For holds and blocks, a personalized communication is sent via SMS and email within 60 seconds: what happened, what action to take, and how to verify their identity if the transaction is legitimate.

Tools: Internal risk engine (PostgreSQL), Python (orchestration), internal transaction API, MaxMind (IP geolocation), OpenAI GPT-4o (decision agent), Twilio (SMS), SendGrid (email), internal analyst queue system.

Outcomes:

• Alert resolution time for auto-resolved cases: reduced from 14 minutes (human) to 90 seconds
• False positive rate on auto-approved alerts: 1.8% (acceptable within regulatory guidelines)
• Analyst time shifted: from 80% investigation to 40% investigation / 60% complex case analysis
• Fraud loss prevented: 22% increase in confirmed fraud caught in the 6 months post-deployment (agents surface cases faster, reducing fraudulent transaction windows)
• Customer communication time: instant (from minutes-to-hours previously)

---

Workflow 7: Employee Expense Report → Extract → Validate → Approve or Flag#

Context: A 900-person professional services firm processing 1,800 expense reports per month. Finance staff were spending 3 hours per day on manual expense review — verifying receipts, checking policy compliance, and chasing missing documentation.

The workflow:

Step 1 — Submission (Concur) Employee submits an expense report in Concur. A webhook triggers the workflow.

Step 2 — Receipt Extraction (OCR Agent) For each receipt attached to the report, the extraction agent:

1. AWS Textract OCR processes the receipt image
2. GPT-4o extracts: merchant name, date, category, amount, and tip (if applicable)
3. Compares extracted amount against the employee-entered amount (flags discrepancies > $0.50)

Step 3 — Policy Validation (Compliance Agent) The compliance agent checks the extracted report against the company expense policy (retrieved via RAG from a Pinecone index of the 42-page policy document):

• Per-meal limits by location (NYC: $75/person; other US cities: $50/person)
• Hotel rate limits by city tier
• Receipt requirement thresholds (receipt required for any expense > $25)
• Alcohol policy (reimbursable only when entertaining clients — requires client name in notes)
• Weekend submissions (flagged for manager attention if no business justification provided)

Step 4 — Routing Decision

• Clean report (no flags, total ≤ $2,500): Auto-approve, send to payroll
• Minor flags (policy questions, small discrepancies): Route to employee's manager in Concur with flags highlighted and specific policy citations
• Significant flags (missing receipts > $100, major policy violation, total > $2,500): Route to finance controller with detailed flag report

Step 5 — Employee Communication

• Auto-approved: Employee receives confirmation with expected payment date
• Flagged: Employee receives an email listing exactly what is flagged and what information is needed, with a direct link to resubmit

Tools: Concur (expense platform), AWS Textract (OCR), OpenAI GPT-4o (extraction and policy compliance), Pinecone (policy document RAG), SendGrid (employee communication), Zapier (orchestration).

Outcomes:

• Auto-approval rate for clean reports: 54% of total submissions
• Finance staff expense review time: reduced from 3 hours to 45 minutes per day
• Policy violation catch rate: increased from 71% to 94% (agent checks every line against policy; humans were inconsistent)
• Average reimbursement time: reduced from 12 days to 5 days

---

Principles for Designing Reliable Workflows#

The examples above share five structural principles that separate workflows that work in production from those that fail:

1. Narrow each agent's responsibility. An agent that does one thing well is more reliable than an agent doing five things adequately. The invoice workflow uses separate agents for extraction, matching, and routing — not one agent doing all three.

2. Grade every decision by reversibility. Auto-approve what can be easily corrected. Route to humans what cannot. The fraud workflow auto-resolves low-risk alerts and always escalates high-risk ones.

3. Pre-built context beats post-hoc investigation. Every escalation in these workflows arrives with a pre-built context brief. Humans should review and decide — not reconstruct from scratch.

4. Confidence thresholds are required, not optional. Every generative step (resolution, matching, classification) needs a confidence score with a defined floor below which the workflow routes to a human.

5. Audit every automated decision. Each workflow logs what the agent decided, why, and what happened next. This is how teams improve agents over time and satisfy compliance requirements.

For technical implementation guidance, see Build Your First AI Agent and the AI Agent for Sales Automation tutorial. Browse AI Agent Templates for pre-built workflow configurations, and see Use Cases for additional applications by industry. For multi-agent coordination within these workflows, see Multi-Agent System Examples in Production.