In 1515, a fifteen-year-old Benvenuto Cellini was placed by his father in the workshop of a Florentine goldsmith. What followed were years of watching, imitating, failing, and being corrected — an apprenticeship that transferred knowledge at the speed of human attention. A master might train fifteen craftsmen in a lifetime. The alloy compositions. The feel for which commissions to accept. The trick of handling a volatile patron without losing the work. All of it lived in his hands and his memory, and when he died, most of it died with him.
For eight hundred years, every serious attempt to escape that constraint ran aground in roughly the same place. It culminated in modernity in three approaches: management consulting, enterprise software, and offshoring. Consulting is knowledge arbitrage — smart people rotate through clients and carry the knowledge out the door with them. Enterprise software captures the formal layer, the rules that can be written down, and leaves everything else trapped in people. QuickBooks doesn't know which categorisation will survive an audit. Salesforce can't tell you the deal is dead despite what the pipeline says. Offshoring moved the work somewhere cheaper without changing its underlying economics, and more often than not made the organisation dumber in the bargain.
Now the constraint that has shaped eight centuries of expert work has broken. For the first time, a technology exists that can hold and structure domain knowledge — and act on it — independent of the person who originated it. A services economy measured in tens of trillions of dollars, almost entirely built on people applying judgment to specific contexts, is suddenly addressable in a new way.
But knowing the constraint has broken is not the same as knowing how to capture what's on the other side of it.
Intelligence is commoditising. Knowledge is not.
Foundation models are getting more capable every quarter, the gap between the leading ones is narrowing, and the per-token cost of inference keeps falling. Knowledge moves the other way. The contextual understanding of how a particular business actually works is scarce, and it accumulates slowly. The formal layer of a domain — statutes, standards, textbooks — can be scraped off the public internet. The layer underneath cannot. That layer is locked inside organisations, sprawled across systems and people, often only partially legible even to the practitioners living inside it.
As intelligence commoditises, knowledge becomes what binds the production of value. A more capable model pointed at the same knowledge base produces better outcomes. Every improvement in the cheap input makes the scarce asset more productive. This is the inverse of what is happening to traditional software, where every AI step-change erodes the value of tools built on static business logic.
The knowledge layer
Sequoia calls the opportunity “services-as-software.” General Catalyst writes about vertically-integrated AI-native firms. NFX has declared that the AI workforce has arrived. Each framing is right about the shape of the shift and right that the value is moving from software to services. What they understate is the hard part underneath.
The abstraction that matters isn't another workflow or another agent. It's a living, structured representation of how a domain actually works — the formal rules, the scattered institutional context buried in email threads and case files, and the tacit judgment practitioners carry but have never written down. Get this in place first, and workflows and agents become expressions of it rather than bespoke constructions stacked on guesswork.
The economics invert once this layer exists. The first application built on a shared knowledge foundation is expensive, because you're building the foundation and the application at the same time. The fifth is much cheaper. The fiftieth is almost free. Every interaction deepens the knowledge, and every correction sharpens the model. When the rules change, you update once and every downstream process inherits it — and when the next business in the domain is onboarded, the hard work is already done.
Why this is hard
Take a mid-sized accounting firm. The formal rules — tax law, reporting standards — are public; anyone can read them. The knowledge that actually runs the firm is elsewhere. It's in twenty years of client files. It's in an email thread between the senior partner and an ATO case officer explaining why a particular structure was accepted and what would happen if it were pushed further. It's in a veteran accountant's instinct for which categorisation will hold up under scrutiny, which client's numbers don't quite add up, which new regulation will bite in practice and which can be worked around.
Extracting a knowledge layer like this requires sustained access to the operation. Not a two-week consulting engagement, a vendor integration, or a customer success call. It requires being inside the business, with the authority to dig, extract, listen, and be trusted. Every standard approach to the industry sits on the wrong side of that access problem. Consultants don't stay long enough to see the edges and aren't aligned to deliver transformation. SaaS vendors see only what their product touches, which is a narrow slice of the real work. AI-native startups selling tools into the industry hit the same wall from a different angle: the customer controls the data, the context, and which questions get asked. Each is a bet that this layer can be assembled from the outside. It can't.
What about the foundation model companies?
Will OpenAI, Anthropic, and Google absorb this themselves? Will enough raw intelligence eventually dissolve the knowledge problem from above?
No. More intelligence doesn't solve the knowledge problem. A smarter model pointed at no knowledge produces smarter-sounding guesses, not better decisions. The scaling curve the labs are riding is real, but it doesn't touch the part of the problem that matters here. You cannot scale your way into knowing how the tacit knowledge lives through an organisation. That knowledge has to be extracted, and extracting it is a ground game the labs cannot play. It lives in private systems, private files, private conversations — inside businesses that have no reason or ability to hand it over to a general-purpose model vendor.
A generation of enterprise data initiatives — data warehouses, data lakes, master data management — showed how hard this is even for the structured layer, where the data at least exists in systems. The tacit layer is harder by an order of magnitude: it doesn't live in any system at all.
And even if the firms could extract and hand that data over, they won't want to. Sitting inside 40,000 mid-market services businesses is the opposite shape of a frontier lab. It is slow, operational, regulated, and unglamorous. The labs will build the layer underneath, rent it to everyone, and watch the operational layer above get built by people willing to do the unsexy work. That's the right strategic call for them.
You cannot scale your way into knowing how tacit knowledge lives through an organisation.
The opportunity
Foundation models are the enabling layer. Without the progress the labs are making in commoditising intelligence and enabling the knowledge layer, none of this is possible. Every organisation building on top of them is riding their curve and benefits directly from each generation of capability. The question isn't which model wins. It's who is building the knowledge layer — vertical by vertical, across the services economy — and turning it into a foundation that actually does the work, at a consistency and scale no human workforce alone can match.
That organisation compounds. Every client deepens the knowledge, every correction sharpens the model, every new process is cheaper than the last. Everyone else is renting.
Why Dragonfly
Dragonfly is built around this thesis. We acquire people-intensive services businesses and build the knowledge layer underneath — not as consultants, not as a software vendor, but as owners. Three things make this structure work where others can't.
- 01Permanent skin in the game
The knowledge layer isn't a deliverable; it's an ongoing capability that has to be built and maintained every day, across every edge case, for as long as the business exists. Ownership is what makes that economically rational. Every dollar we invest in extracting, structuring, and activating knowledge compounds on a balance sheet we own outright.
- 02A multi-decade horizon
Knowledge compounds, and compounding looks unimpressive for a long time before it looks inevitable. Year one of this investment looks worse than buying workflow automation off the shelf. Year ten looks like a different category of business entirely. Dragonfly is built without an exit clock — no fund return window, no turnaround mandate, no timeline optimising against the compounding.
- 03Operational control
A knowledge layer can't be extracted through access. It has to be built from inside the operation — owning the processes, the systems, the relationships, and the practitioners. We decide what gets captured, how it's structured, how it's activated, and what happens when it's wrong. Anything short of this is a vendor relationship, and vendors don't get to rebuild the architecture of the business they're selling into.
This is why we don't operate like a consulting firm, a SaaS company, or a PE roll-up. Each of those models has the wrong geometry for what we're trying to build.
The prize
The services economy is going to be rebuilt in the next decade. The consultancies that have always run it can't. The software vendors that tooled it won't. The foundation models that will power it shouldn't. The work falls to a smaller group of operators willing to buy these businesses and build the knowledge layer underneath — patiently, across a horizon longer than anyone else is willing to wait.
Every acquisition deepens the model of the domain. Every deeper model makes the next acquisition more valuable. Some time in the 2030s, it will be obvious that the largest category of enterprise created in the AI era wasn't a model lab or a SaaS vendor, but something older and more patient that most people weren't watching.
Dragonfly is building for that moment.
