Why you need an Agent Management Platform
Thibault Martin
Want just the recap? Take a look at this:
After two years deploying AI agents with customers ranging from industrial firms to leading AI companies, we've developed a clear understanding of what separates AI-first companies from those that stall. The difference is almost always about having the right approach to managing AI at scale.
The truth is, switching to an AI-first routine is a journey with distinct stages. Most companies are discovering this the hard way, moving from initial excitement about ChatGPT to the sobering reality that 95% of enterprise AI pilots fail to deliver tangible value. We've seen this pattern repeatedly: organizations that treat AI as a single deployment decision inevitably hit a wall. Those that approach it as a systematic transformation climb steadily toward transformative outcomes.
From traditional to AI-First: Four stages
There's no shortcut to becoming an AI-first company, it's a gradual climb. After working with over 2,000 companies, we've seen the same four stages play out. Understanding where your organization stands today will clarify what you need to do next.
Stage 1: Secure LLMs
What it is: A secure, private version of Large Language Models for your organization.
Why it matters: ChatGPT already proved AI's business value. Secure LLMs is about control. Your employees are already using AI; the question is whether they're copying sensitive data into consumer products where "if the product is free, you are the product." A secure GPT gives you a controlled environment where AI adoption can happen safely.
Reality check: If you think your employees aren't using AI yet, you're wrong. Research shows that while only 40% of companies have purchased LLM subscriptions, over 90% of workers use personal AI tools like ChatGPT for work, multiple times daily. This is your Shadow AI problem.
Stage 2: Knowledge Assistants
What it is: AI agents that can access and leverage your company's internal context and knowledge.
Why it matters: This is where ROI becomes undeniable. When properly implemented, knowledge assistants transform how employees access expertise, turning scattered information into instant, accurate answers that previously required hours of searching or waiting for colleagues.
Example from GTM at Clay (see story): Clay achieved 100% adoption of their GTME knowledge assistant, saving 58 hours monthly for their growing 20-person team. With 10 new hires joining monthly, each asking an average of 10 questions daily, the time savings compound quickly. As Caren Duane from their Operations team explains: "Dust democratizes information access, so it's no longer about who joined the company earliest or who sits closest to the Head of Sales. Now there's just a single source of truth for everyone."
Example from engineering at Persona (see story): Engineers were constantly interrupted by technical questions flooding the #ask-engineers Slack channel from sales, customer success, and solutions teams. To solve this, they built PersonaEngineer—a Dust agent with specialized sub-agents for code, infrastructure, databases, and documentation—that automatically answers technical questions in Slack by routing queries to the right knowledge source.
Stage 3: Personal Agents
What it is: Agents that don't just answer questions but actually do things on your behalf.
Why it matters: This removes manual steps from processes and lets people focus on where they bring the most value: thinking, not executing.
Instead of asking an AI "what should I do?" and then doing it yourself, agents complete entire workflows autonomously.
Our own example at Dust (see story): Sales teams using Dust operate in a radically different way: our Frames functionality combined with research lets you build hyper-personalized presentations by pulling data from narratives, support tickets and usage analytics. After a call, you can upload a transcript and automatically update the presentation with new insights from battle cards. The entire sales workflow transforms when agents handle the heavy lifting.
Example from productivity at Wakam: Wakam, one of Dust’s earliest customers, has built a personal productivity agent for all of their employees that integrates with their core tools (Outlook email/calendar, Slack, Notion, and file management) to handle cross-platform tasks like scheduling meetings, managing emails, searching company knowledge, and creating or updating documents.
Example from sales at Patch (see story): Patch, a climate tech startup, needed to scale their scarce climate expertise across their sales team. They built three specialized agents: a Corporate Sustainability Decoder that analyzes prospects' sustainability strategies with expert-level insights, a Business Intelligence agent that queries proprietary carbon credit transaction data, and a Project Recommendation Engine that filters through 25,000+ carbon projects based on natural language criteria. Within 3 months, they achieved 70% weekly active usage and increased data usage in sales calls from 10% to 70%.
Stage 4: Autonomous Agents
What it is: Autonomous AI teammates that collaborate with humans without constant oversight. This is Dust's core investment area: features and triggers designed for company-wide intelligence, not just individual productivity.
Why it matters: This is where you transcend time savings and accomplish things that weren't feasible before. It's an infinite extension of your workforce at low marginal cost. Importantly, this doesn't replace people, it augments them. Everyone can have mini-colleagues that work exceptionally fast. It's the ultimate acceleration.
Below are examples of in-production use cases at some of our most advanced US customers.
Our own example at Dust: After each sales call at Dust, transcripts are automatically converted into product updates in Notion, CRM enrichment happens automatically, and follow-up emails are pre-drafted. What would have taken hours of manual work (or would have been prohibitive) now happens in the background.
Example from operations at Vanta (see story): Vanta built a three-layer autonomous system for their 200-person GTM team. Domain experts created specialized agents (GRC for compliance, Finance for usage metrics, Product for Voice of Customer feedback), which were then orchestrated into automated workflows. Their flagship use case: automated Quarterly Business Review prep. Instead of reps spending hours manually gathering data from dashboards, Dust calls multiple agents, generates pre-built decks with speaker notes, and surfaces insights they would have missed. This single workflow saves ~400 hours per week across 200 reps—thousands of hours reclaimed annually. As Danny Baralt, Business Systems AI Solutions Lead, notes: "We used to do the work. Now we build the agents that do it."
Example from operations at a $10B+ company: The IT team created a three-tier automated ticket classification system. High-confidence requests are automatically deflected and closed. Medium-confidence requests match against documentation with self-service options. Only low-confidence requests reach specialized teams—and even then, they arrive with context and routing information that gives agents a head start.
The challenge is that moving from Stage 3 to Stage 4 requires rethinking fundamental paradigms like permissions in the age of AI. If your agent works autonomously, how do you handle access to information that employees can't directly see? This is where many organizations realize they need an Agent Management Platform.
Wait… why not a workflow tool? This is a question we often get. The short answer is: Agents let you describe what you want in plain language and figure out the steps on their own, while workflows force you to map out every single scenario upfront—which gets messy when unexpected situations pop up. As AI models get better, agents automatically improve without you changing anything, but workflows need to be completely redesigned to take advantage of improvements. Check the long answer here.
What is an Agent Management Platform and why you need one
The companies climbing this ladder share something in common: they recognized early that AI agents aren't just another SaaS tool. They're a fundamentally new category of technology that requires new infrastructure to manage them: Agent Management Platforms.
When we talk about an Agent Management Platform, we mean a central hub where companies build and manage AI agents for their teams. Instead of everyone using AI tools separately, it gives you one place to create custom AI agents that can access your company's information and tools. Anyone in the company can build these agents without knowing how to code, and employees can use them wherever they work—in third-party tools, on the web, or through custom-built user interfaces. While this positioning was quite unique when we launched Dust two years ago, the industry is now validating our approach, with major players pivoting toward similar infrastructure.
Here's why we believe Agent Management Platforms will become a dominant product category.
1. It has robust foundations built for scale
The platforms built specifically for agent management were designed from the ground up for the challenges raised with Stage 3 and 4. They're not chat interfaces retrofitted with agent capabilities. They're not search tools trying to become action platforms.
When Doctolib, a $6B-valuation scale-up evaluated whether to continue building their internal AI platform or partner with Dust, their VP of Data put it bluntly: "I was the first one to ask to stop internal development. We wanted to be free from the 'burden' of having to be the product owner, rather than the customer."
Building an agent platform does NOT mean building a chatbot. It means building permission systems that work at agent-scale, conversation management, multi-model orchestration, semantic search infrastructure, evaluation frameworks, and user interfaces. The market moves faster than most internal teams can keep pace.
The reality: According to MIT's research on enterprise AI, strategic partnerships have approximately a 67% deployment success rate compared to just 33% for internal builds. Internal builds fail twice as often as partnerships.
2. Your employees get better over time (not left behind)
If you believe your company needs to use AI (and you should), you can't afford to leave people behind. The beauty of a unified platform is that it helps all your teams climb the transformation ladder together.
Consider Kyriba's deployment. Within months, they saw:
- 7% of users saving 5+ hours weekly
- 14% saving 3-5 hours weekly
- 42% saving 1-3 hours weekly
- 32% saving 30 minutes to 1 hour weekly
But more importantly, they doubled the number of people saving 3+ hours per week in just one quarter. The platform was helping the early adopters on top of systematically elevating everyone's AI capabilities.
This is what agent management platforms enable: a systematic approach to upskilling your entire organization, not just the technical teams or early adopters.
3. It helps your IT and Procurement teams
For your CISO: If agents proliferate without central management, you face a nightmare scenario. Who maintains which agent? What data does each one access? Which models are being used, and how are we compliant with regulations?
One Dust customer, a French industrial group concerned about FISA-702 and Cloud Act compliance, made it clear: "We can't move forward without this level of control." When you have agents everywhere, it becomes exponentially difficult to handle permissions, identify who knows what, maintain audit logs, and ensure compliance.
Agent Management Platforms provide:
- Centralized governance frameworks
- Unified permission systems designed for agents
- Audit capabilities that track what agents do
- Compliance controls that scale with your agent fleet
For your CFO: Your employees have one tool instead of a dozen fragmented AI subscriptions. Spendesk's CEO explained their pre-Dust challenge: "We tried individual contracts with different AI assistants, but this proved inefficient and unsustainable." Consolidation is dramatically more cost-effective.
For your Business Teams: Adoption comes from clarity. When AlertMedia faced their "honeymoon to reality" transition, they realized: "Everyone has ChatGPT licenses and we don't know how they're using it... everyone's using ChatGPT in a different way and generating different content and creating different agents. We need control before we go down a rabbit hole."
A unified platform solves the discovery problem. Instead of wondering if an agent exists for your use case, you can search, find ratings, see usage patterns, and deploy proven solutions. Wakam deployed 136 agents across their organization—50 company-built and 90 personal agents—because their platform made discovery and adoption systematic rather than chaotic.
4. It avoids vendor lock-in
What if your CRM's AI capabilities fall behind? What if a competitor disrupts the space with better technology? Would you rather bet everything on one vendor's ability to keep pace with AI innovation, or maintain the flexibility to swap underlying technologies while preserving your agent investments?
Agent Management Platforms are typically model-agnostic. At Dust, we give you access to OpenAI, Anthropic, Mistral, Google, and others—switching models is a configuration change, not a rebuild. When Anthropic released Claude Sonnet 3.5, 70% of user queries shifted to Claude Sonnet 3.5 within weeks from other models because the superior capabilities were immediately available. No need to switch vendors or rollout new tools.
Compare this to being locked into a single vendor's model strategy. Or worse, having agents built into dozens of different tools across your organization, each with its own AI strategy, creating a fragmentation nightmare that's nearly impossible to govern or improve systematically.
Questions to ask yourself
Beware the 'easy' option
The most dangerous decision in this AI journey is choosing what feels safe today but creates a glass ceiling tomorrow.
Consider permissions. For decades, your security model has been straightforward: users have access or they don't. But what happens when an HR agent needs to access documents that employees can't see directly? Or when a sales agent needs to pull competitive intelligence from multiple confidential sources?
One CISO told us about a major paradigm shift they faced: "We have public SharePoint sites that AI now accesses. Before, they were effectively hidden because no one knew to look there. Now we need to rethink what 'public but hidden' actually means in the age of AI."
The conservative path—adding AI capabilities to your existing tools without rethinking underlying assumptions—will hit a ceiling. You'll get to Stage 2 (Knowledge Assistants) and struggle to progress further because your infrastructure wasn't designed for autonomous agents.
Think of this as a long-term investment
Moving to AI-first takes time. Industry research shows:
- 6-12 months for initial results
- 2-3 years for full transformation
- Multiple phases of capability building required
Too many companies choose the tool that makes sense for 2025 without considering that 2022 was when most people discovered ChatGPT. In just three years, we've moved from basic chatbots to autonomous agent systems. Where will we be in another three years?
McKinsey's research is clear: only 5-6% of organizations are achieving transformative value from AI. The ones succeeding share common traits—they're 3x more likely to:
- Pursue transformative change (not just efficiency)
- Fundamentally redesign workflows
- Scale agents across functions
- Have deeply engaged C-suite leadership
These aren't companies that chose the easy option. They're companies that made long-term investments in platforms that can scale with AI's rapid evolution.
It's all about “adoption,” so involve your employees!
Doctolib's approach to their company-wide AI rollout is instructive. They positioned AI as a "national cause"—an imperative business transformation, not an optional nice-to-have. They trained all leaders first (7-hour sessions), built a champion network from usage analytics, and provided always-on support through hotlines, office hours, and peer mentoring.
The result? 30% daily usage across 3,000 employees, 70% weekly usage, and 20% of employees built their own agents. Their Chief People Officer's guidance: "Start with adoption, not ROI. The first 6 months should be about behavior change."
If you move too fast without this cultural foundation, you risk losing trust entirely. Employees can't graduate directly from nothing to working with autonomous agents. This staged progression is at the core of why Agent Management Platforms matter—they're built to help you navigate the journey, not just deploy the technology.
Do you have the right skills internally? Should you partner with experts who can bridge the gap between AI knowledge and your industry expertise? The companies climbing the ladder fastest aren't going it alone.
How to evaluate the Agents space
When we started building Dust in 2023, we were the first to focus on deploying a fleet of AI assistants across the enterprise. Today, we're reassured to see the market catching up—existing players are repurposing their platforms and converging toward AI Agent solutions.
This shift is happening across different types of companies:
- Workspace solutions adding agent capabilities to collaboration environments
- AI search engines evolving into platforms that can take action, not just find information
- Collaboration tools building out agent orchestration features
This convergence validates what we've believed from the start: enterprises need a unified approach to AI agents, not a collection of disconnected tools.
What matters for your decision:
1. Proven track record
- How long has the platform been in production?
- Are AI-native companies using it?
- Do they have customers who've reached Level 4?
2. Architectural approach
- Was this built for agents from day one, or retrofitted?
- Does it work across your entire tech stack?
- How do autonomous agents handle permissions?
3. Path to value
- Can you start small and scale?
- Is there a clear progression from simple to complex?
- How quickly can you see ROI?
Conclusion
We started Dust with a conviction: enterprises would need a new category of platform to harness AI's full potential. Today, that conviction has become reality.
What we've learned from 2,000+ companies:
- AI-first is a journey, not a destination. Companies that succeed embrace progressive complexity, starting simple and scaling sophistication.
- Platforms beat point solutions. The companies reaching Stage 4 all use unified platforms. The companies stuck at Stage 1 have scattered tools.
- Adoption is everything. The best technology is worthless if employees don't use it. Success requires treating AI as organizational change.
- The window is now. The gap between companies that started 12 months ago and companies starting today is already significant. In another 12 months, it will be a chasm.
Ask yourself:
- Where are we on the ladder today?
- Where do we need to be in 12 months?
- What's preventing us from climbing?
- Do we have the right infrastructure to scale?
Agent Management Platforms are the operating system that makes your entire AI strategy coherent, governable, and scalable. The companies making this platform decision today are building competitive advantages that will compound for years to come.
Want to learn more about how companies like Assembled, Vanta, Cursor, Clay, Doctolib, Qonto, and Wakam are scaling AI with Dust? Visit dust.tt to explore case studies and see how an Agent Management Platform can accelerate your transformation.