AI agents examples with Dust: Use cases across different teams
Davis Christenhuis
AI agents are evolving from experimental projects to production systems that change how teams work. Companies are using them to solve problems across different departments.
This article walks through five AI agents examples from teams across different functions. Each one solves a different problem with measurable results.
📌 TL;DR
Before diving into the examples, here's what this article covers:
- What makes a good AI agent: Clear instructions, connection to the right data sources, a focused scope, and feedback loops to improve over time.
- Five AI agents examples used by companies: Brevo personalizes sales outreach, Back Market detects fraud, Vanta automates QBR prep and knowledge sharing, Persona answers engineering questions, and Watershed built agents across HR, sales, and engineering.
- The results range from building in one week to company-wide adoption: One agent was built in a weekend, another in a week, others saved 400 hours per week. Impact ranges from 80% faster workflows to a fraud prevention initiative projected to save €1.2M annually.
- What they all had in common: Built by domain experts, connected to company data, started small and spread organically, and delivered measurable results in weeks.
- Dust makes it possible to build agents without code: Connect your tools, write instructions in plain language, and deploy where your team already works.
💡 Want to build an AI agent for your team? Try Dust free for 14 days →
What makes a good AI agent?
Not every agent delivers results. The difference between agents that get adopted and agents that get abandoned comes down to four factors:
- Clear instructions: specificity drives accuracy. Generic prompts produce generic outputs. The best agents have detailed instructions that explain exactly what to do, when to escalate, and what format to use.
- The right data: agents need access to the sources they'll actually query, not every data source in the company. A sales agent needs CRM access and product docs, not HR policies.
- A defined scope: one job done well beats ten jobs done poorly. Teams that try to build a single agent for every use case end up with agents that do nothing particularly well.
- A feedback loop: usage analytics and user feedback help teams improve agents over time based on what actually works.
AI agents examples with Dust across different teams
The following five examples come from real companies running agents in production.
1. Brevo's sales agent that personalizes outreach
Brevo is a Paris-based customer engagement platform serving over 600,000 customers globally. Their go-to-market team faced a bottleneck: business development reps needed to send personalized outreach to hundreds of prospects weekly, but manual research was slowing them down.
The problem: Sales reps spent most of their time hunting for context across systems. CRM data sat locked in their CRM system, product information lived in Notion, and enrichment data required manual web research. At that pace, personalization meant sacrificing volume.
Brevo's revenue operations team built multiple AI agents connected to their operational database. The email generation agent pulls complete prospect history from their CRM, enriches it with web and LinkedIn data, routes to specialized sub-agents based on prospect type, and generates three personalized email variations per contact.
The agents they created include:
- Customer referral finder: Surfaces relevant customer examples for social proof before sales calls
- Personalized email generator: Creates multi-variant email sequences based on prospect context and company data
- Landing page personalization agent: Generates custom marketing plans for inbound leads in real time
Result: 80% time reduction on email personalization research, bringing BDR work down from 30+ minutes per prospect to just minutes. The team has executed 2,500+ production actions through their Supabase-connected agents with zero engineering tickets filed.
2. Back Market's fraud detection agent
Back Market is the world's leading online marketplace for refurbished electronics. They faced a logistics fraud problem: fraudsters were purchasing expensive items, requesting refunds, and either sending back empty boxes or manipulating shipping so packages never reached their destination.
The problem: Manual fraud investigation took hours per case. The team had to cross-reference addresses, analyze purchase histories, check shipping data, and scan conversations for fraud patterns. Any update to detection rules required engineering resources, creating a bottleneck that let fraud patterns evolve faster than the team could adapt.
Back Market's fraud team built a multi-agent system called the Fraud Orchestrator. A central coordinator routes each case to specialized sub-agents that handle specific checks:
- Address Check agent: Evaluates delivery address risk against known fraud locations
- Return Distance agent: Calculates geographic anomalies between delivery and return points
- Customer Search agent: Analyzes purchase history and incident frequency
- Tracking Incidents agent: Flags shipping anomalies
- Payment Incidents agent: Identifies payment-related fraud signals
- Conversation Pattern agent: Detects fraudulent message templates stored in Confluence
Result: Built in one week by the fraud team without engineering support, the system now contributes to a broader fraud prevention initiative projected to save €1.2M annually. The AI-powered claims analysis alone has prevented nearly €100K in fraud over five months, and pattern updates that used to require months now take less than a day.
3. Vanta's knowledge agent
Vanta is the leading trust platform used by companies of all sizes to unify compliance, risk management, and customer trust workflows. Preparing quarterly business reviews required GTM reps to pull data from finance, GRC, product, and marketing across disconnected tools.
The problem: Every function housed critical insights, but those insights lived in silos. Quarterly business review prep meant hours spent hunting through dashboards, assembling charts, and synthesizing context from multiple teams. Multiplied across 200 reps, the hidden cost was enormous.
Vanta built a three-layer agent architecture. Domain experts in each function first built their own specialized agents:
- Finance agent: Provides usage insights and revenue signals
- GRC agent: Handles compliance frameworks and security questions
- Voice of Customer agent: Surfaces client feedback from support tickets and calls
- QBR automation: Orchestrates the domain agents above to generate a pre-built deck, speaker notes, and a context-rich summary
The GRC agent also answers compliance questions live in Slack with quick human review before responses go out.
Result: Around 400 hours saved per week across the GTM team on QBR prep alone. Reps now go into customer meetings with better context because the agent synthesizes information from sources spread across multiple systems. Adoption expanded beyond GTM to the entire company.
4. Persona's knowledge agents across engineering and sales
Persona is an identity verification platform backed by a $200M Series D. As they scaled, engineers were drowning in Slack questions, sales reps were buried in RFPs that took days to complete, and solutions engineers were spending hours writing formal documents from call transcripts.
The problem: The #ask-engineers Slack channel was a firehose. Engineers faced constant interruptions pulling them out of focus time. Sales needed days to compile accurate RFP responses. Solutions engineers manually reviewed multiple call transcripts to write Solution Requirement Documents following internal templates.
An engineer built PersonaEngineer over a single weekend. Instead of dumping every knowledge source into one agent, he built specialized sub-agents for different domains:
- PersonaEngineer (orchestrator): Routes questions to the right sub-agent
- DDDEngineer: Searches GitHub codebases
- InfrastructureEngineer: Handles production changes and incidents
- DataQueryExpert: Queries data warehouses
- PersonaHelpCenter: Searches technical documentation
- PersonaGlossary: Resolves internal terminologies
- PeopleNerd: Searches the employee directory
The success of PersonaEngineer inspired other teams to build their own standalone agents:
- RFPNerd: Answers RFP questions using past proposals and company data
- SRDNerd: Generates Solution Requirement Documents automatically from call transcripts
Result: Within six months, Persona hit 80%+ company-wide adoption, with 85% of Sales and Post-Sales teams actively using agents.
Fraud analysts who used to spend hours on complex SQL queries now finish the same work in under 30 minutes, while sales teams that once spent days on RFPs now generate accurate answers in a fraction of the time. Nearly 300 agents have been built across the company.
5. Watershed's agents for HR, sales, and engineering
Watershed is an enterprise sustainability platform used by companies like Airbnb, Visa, and FedEx. They had grassroots AI experimentation happening across the company, but no consistent adoption or shared infrastructure.
The problem: Different teams were experimenting with AI tools in isolation. Nothing was shared, nothing was documented, and nothing could scale beyond the person who built it.
Watershed deployed Dust company-wide and empowered every department to build agents tailored to their specific workflows:
- SDR prospecting agent: Researches prospects and drafts outbound emails following the sales playbook
- Gong + Salesforce notes agent: Processes call recordings, extracts key takeaways aligned with Watershed's sales methodology, writes summaries back into Salesforce automatically
- Performance review coach: Guides employees through writing thoughtful performance reviews (used by ~1/3 of the company)
- Engineering design doc agent: Helps engineers complete design documentation with completeness checks
Result: Watershed went from 20% to 90% company adoption within months. Every major function now has dedicated agents, and employees shifted from thinking AI was useful for 1% of their job to understanding that not using AI means not working as effectively as they could be.
Curious how other companies are implementing Dust? Explore all our customer stories →
What these AI agents have in common
Look across all five examples and four patterns emerge:
- Built by domain experts, not engineers: the fraud team at Back Market, revenue ops at Brevo, an engineer at Persona who built it over a weekend, GTM teams at Vanta and Watershed. The people who understood the problem built the solution.
- Connected to the company's own data: these aren't generic AI tools. Every agent pulls from CRMs, databases, support tickets, call recordings, and internal documentation. The company's context is what makes them work.
- Started small, spread organically: Most of these companies saw adoption spread organically. Others, like Watershed, combined grassroots experimentation with structured enablement programs. In every case, the agents succeeded because they solved real problems.
- Measurable impact in weeks, not months: Back Market shipped in one week. Persona's engineer built over a weekend. Brevo went from concept to production workflows in days. Speed to value determines whether teams keep using agents or abandon them.
These patterns show up across teams in every function: sales, marketing, engineering, operations, HR, and more. When agents are built by the people who understand the work and connected to real company data, they become assets that teams actually use.
More about Dust
The companies in this article are using Dust to deploy AI agents at scale. Dust is an operating system for AI agents — a platform that lets teams deploy, orchestrate, and govern specialized agents connected to their company's knowledge and tools.
You connect the tools your team already uses: Notion, Google Drive, Slack, Salesforce, GitHub, Zendesk, and dozens more. Then you write instructions for what the agent should do.
Agents can answer questions, research across systems, generate content, or query databases in plain language. The process of building an AI agent is accessible and requires no coding experience.
Teams deploy agents where work happens: in Slack, inside support tickets through Zendesk, or via a browser extension that works anywhere on the web.
Security is built in from the start, with SOC 2 Type II certification, enterprise-grade permissions, and regional data hosting for companies with strict compliance requirements.
Dust is a platform, not a single tool. That means domain experts can build agents themselves without waiting for engineering resources to become available.
💡 Try Dust free for 14 days and see how your team can build its first agent. Start here →
Frequently asked questions (FAQs)
How do AI agents learn and improve?
AI agents improve through feedback loops and usage patterns. When teams use agents, they can rate responses, flag errors, and refine instructions based on what works. The agent's instructions and connections to data sources can be updated in real time without rebuilding the system. Some agents also learn from conversation history and user corrections, allowing them to adapt to specific team preferences and workflows over time.
What's the difference between an AI agent and automation?
Traditional automation follows fixed rules: if X happens, do Y. AI agents reason through problems and adapt to context. As edge cases accumulate, automation workflows grow increasingly complex to maintain. AI agents handle variation through reasoning rather than pre-defined paths. For example, automation can route a support ticket based on keywords. An agent can read the ticket, search past conversations, check the customer's account status, and draft a personalized response that addresses the specific situation.
What tasks are AI agents best at?
AI agents excel at tasks that combine research, reasoning, and repetitive execution. They work well for answering questions from large knowledge bases, synthesizing information across multiple systems, generating content that follows specific guidelines, querying databases in natural language, and routing work between specialized systems. The best use cases involve work that requires judgment but follows a pattern, tasks that pull from multiple data sources, or workflows where speed and consistency matter more than creativity.
Can I build AI agents with Dust without a technical team?
Yes. Dust lets people build agents without writing code. You connect your data sources, write instructions in plain language, and deploy where your team works. Teams across sales, operations, HR, and fraud detection have all built production agents without engineering support. The people who understand the problem are often the best ones to build the solution. While engineers can build more complex multi-agent architectures, most production agents have been built by non-technical teams without engineering support.