Your Agents Don't Need More Intelligence. They Need a Map.
For the past year, my team at Attercop has been building something we call Heimdall, a decision intelligence platform for professional services firms. It started as an internal tool to solve our own problems. It's become something bigger: an argument about what AI-native actually means for businesses like ours.
I want to share what we've learned, because I think it matters for every services firm wondering how to survive the next five years.
The speed problem
Y Combinator's Spring 2026 Request for Startups includes a category that should make every agency and consultancy founder pay attention: AI-native agencies. The thesis, as YC Group Partner Aaron Epstein frames it, is that agencies have always been impossible to scale. Low margins, slow manual work, and the only way to grow is to add more people. AI changes this. Instead of selling software to clients, you use the software yourself and sell them the finished product at a fraction of the time and many times the price. The result is agencies that look more like software companies, with software margins.
This isn't a future prediction. It's happening now. And the firms that can't match this pace will find themselves competing against teams half their size delivering twice the output.
But here's the part that YC's framing glosses over: the hard problem isn't getting AI to do clever things. The hard problem is getting AI to understand your business well enough to do useful things across it.
The map problem
When we started building Heimdall, we had a straightforward ambition: we wanted to ask questions that spanned our business. Simple things. "Show me all our work with this PE firm's portfolio." "Who in our team has delivered AI governance projects before?" "What did we charge for a similar engagement last year?"
These questions sound trivial. They're not. Each one requires information from multiple systems: our CRM, our project management tool, our time tracking, our document storage, our accounting platform. And each system has its own way of representing the same underlying reality. The client called "ACME Ltd" in Pipedrive is "ACME" in Forecast, "ACME Limited" in Xero, and a folder path in Google Drive. A deal becomes a project becomes an engagement becomes a set of invoices, and no system knows these are all the same thing.
This is what we came to call the "accidental ontology," a term I later found in Seth Earley's excellent work on enterprise AI. Every firm has one. It's the fragmented, ungoverned collection of naming conventions, folder structures, spreadsheet taxonomies, and CRM categories that have accumulated over years of uncoordinated decisions. It sort of works when everything runs through human brains that can make the mental connections. It completely breaks down when you try to put AI agents in the loop.
And this is the point that I think the market is missing: agents can only work effectively across a business if they have a map of how to do it. The ontology is that map.
What we actually built
An ontology, in practical terms, is a canonical model of what exists in your business and how things relate. At Heimdall, ours defines 17 entity types (organisations, deals, engagements, people, services, deliverables) and the relationships between them. It's the single reference point that says "this CRM deal, this project plan, this document folder, and these invoices are all manifestations of the same underlying engagement."
This isn't a schema diagram that sits in a wiki. It's a living, enforced layer in the platform. Every source system maps into canonical entities through identity resolution. Every query, every agent action, every AI-generated insight operates against the canonical model, not against raw source data.
The result is two things we're now using daily.
First, a chat-queryable MI system. Our team can ask natural language questions about the business and get answers that draw from every connected system. Not a dashboard. Not a report someone has to build. A conversation with the firm's institutional memory.
Second, and this is where it gets interesting, a platform for semi-autonomous agents. We've designed a portfolio of 16 agents that will progressively take on operational tasks: preparing meeting briefs by pulling together everything we know about a client from every system, drafting time entries from calendar data, monitoring pipeline health, surfacing precedents from past engagements when scoping new work. Each agent operates within a governance framework that controls what it can see, what it can do, and how much autonomy it earns through demonstrated reliability.
The critical insight is that none of these agents would be possible without the ontology. An agent that prepares a meeting brief needs to know that the person you're meeting is a contact at an organisation, that organisation has active engagements, those engagements have assigned team members and outstanding deliverables, and there's a deal in the pipeline for a follow-on project. Without the ontology connecting all of this, the agent is just a language model with access to disconnected data sources, capable of impressive-sounding summaries of individual documents but unable to reason across the business.
Why stovepipe AI won't get you there
The enterprise software incumbents understand that AI agents are the future. Salesforce has Agentforce. ServiceNow has AI Agents. HubSpot has Breeze. Every major platform is embedding agentic capabilities.
But they all share a fundamental limitation: they can only see their own data. A Salesforce agent is brilliant at reasoning about your CRM. It knows nothing about your project delivery, your time tracking, your document production, or your finances. A ServiceNow agent can orchestrate IT workflows but is blind to your client relationships.
For a professional services firm, this is a crippling constraint. The questions that actually matter (is this client relationship healthy? are we profitable on this engagement? who should we staff on this opportunity?) require information that lives across five or six systems. An agent imprisoned within one of them can only ever give you a partial answer.
Why plug-and-play won't get you there either
On the other end of the spectrum, a wave of startups are building integrated agentic platforms that promise to connect everything. The pitch is compelling: plug in your tools, and AI agents work across them.
The challenge is that these platforms shy away from the hardest part of the problem. Building a customised ontology for each client (understanding their specific entity model, their naming conventions, their business processes, the relationships between their systems) is genuinely difficult, time-consuming work. It requires domain expertise and close collaboration with the client. It doesn't lend itself to the "sign up and go" model that SaaS economics demand.
So these platforms optimise for ease of integration at the expense of depth of understanding. They can move data between systems. They can trigger actions based on events. But they lack the canonical model that would let their agents truly reason about the business as an interconnected whole. They're building plumbing when what's needed is cartography.
The ontology is the moat
I've come to believe that the ontology is the most strategically important, and most undervalued, asset in the AI-native services firm.
It's undervalued because it's not exciting. Nobody's demo day pitch is "we built a really good canonical entity model." There's no viral moment in identity resolution. But it's the foundation that makes everything else work. Without it, your chatbot is a novelty. Your agents are party tricks. Your "AI-native" operations are a thin wrapper around the same disconnected tools everyone else uses.
With it, you have something qualitatively different: a firm that can think at the speed of its data rather than the speed of its people. Where institutional knowledge doesn't walk out the door when someone leaves. Where an agent can prepare for a client meeting by drawing on every interaction across every system, not just the notes one person remembered to write up.
Seth Earley calls this "information metabolism," the rate at which an organisation can ingest, process, and act on what it knows. The ontology is what accelerates it. And in a world where YC is funding AI-native agencies designed to operate at software speed with software margins, the firms that can't accelerate their information metabolism are the firms that get left behind.
This isn't just about consultancies
It's worth noting that the problem we've described isn't confined to consulting firms and agencies. Private equity firms are, in many respects, services businesses themselves. They manage portfolios of relationships, track deal flow, coordinate across operating partners, and need to reason about performance across dozens of portfolio companies using data that lives in multiple systems. The questions a PE firm needs to answer ("what's the aggregate revenue trend across our mid-market portfolio?" or "which operating partners have experience with this type of transformation?") are structurally identical to the cross-system, relationship-rich queries that Heimdall was designed for. We see strong applications of our work into PE, and more broadly into any organisation where the value lies in relationships, knowledge, and the ability to connect information across organisational boundaries.
What's next for us
We're genuinely excited about what we've built at Attercop. Heimdall started as an internal experiment and has become both a chat-queryable intelligence system that our team uses daily and a powerful platform for the delivery of semi-autonomous agents that will progressively take on the operational load that currently consumes our senior people's time.
But the thing that excites us most is this: we're not unique in having these problems. Every professional services firm we work with (every consultancy, every specialist agency, every advisory practice, every PE house) is wrestling with exactly the same fragmented data landscape, the same inability to reason across their own business, the same hunger for AI that actually understands their operations rather than just their documents.
We're now planning how we can use the skills we've developed, and the platform we've built, to help our clients who are struggling with exactly these needs. The ontology work is hard. It requires domain expertise, careful integration, and a willingness to do the unglamorous foundational work before the impressive agentic capabilities can follow. But we've done it once. We know where the pitfalls are. And we know it works.
The firms that will thrive in the AI-native era aren't the ones with the best language models or the slickest chatbots. They're the ones with the best maps.
Graeme Cox is CEO and co-founder of Attercop. Heimdall is Attercop's Decision Intelligence platform.
