[Home](/)
/ [The Operator Is the Spec](/essays/the-operator-is-the-spec)
[hi@parallelpartners.ai](mailto:hi@parallelpartners.ai)
_Martijn van der Does · 2026-05-28 · 7 min read_
The working spec of a mid-market business is the operator, not a process map. Horizontal AI configures around the process and scales to nothing useful. A vertical build is shaped around the operator and compounds quarter over quarter on their P&L. The model underneath is fungible. The operator is not.
"We can know more than we can tell." Michael Polanyi
01
Rebuild the real economy, one operator at a time, into something that compounds while its owner sleeps.
The labs are building the gods. We are building the machines that put them to work inside the businesses that actually move the world. Not the frontier. The factory floor of the mid-market, where the real economy lives and where almost none of it has been touched.
02
Every founder asks the same thing now, usually with a little fear in it. Is there any AI application layer left to build, or do OpenAI and Anthropic eventually take all of it.
Most of it, they take. They own the model, the distribution, the cadence. Anyone laying a thin sheet of cleverness over the top is already dead and has not noticed. That is the part they take.
The part they cannot take is the only part worth building on. And it is not a feature. It is a person.
03
In the battle against entropy, a business survives on the truths it knows that nobody else does. Not the truths in the manual. The truths in the operator's head.
Consider a €5M mechanical installer. It runs on one estimator's knowledge. He knows which subcontractors pad their quotes in winter. Knows which inspectors take digital sign-off and which still want paper. Knows which customers stretch to 90 days if you do not chase them on day 31, and which municipality wants a second site visit before it issues a completion certificate, because one inspector had a bad experience three years ago and quietly changed how he works.
None of that is in any training set. None of it is on the public web. It is not even fully in the estimator's mouth. Sit him in front of a blank page and he could not write it all down, because he knows more than he can tell. That is the spec of the business. Not the process. The Operator.
04
Software automated the world of bits. Language and math fell first because they were already symbols, already legible, already on the page.
The physical, local, tacit world resisted. The knowledge that runs a real business was never written down, so software could never reach it. That is the last untouched territory, and it is where the next era of value is hiding in plain sight. Look around your own company. The most important things you know are the things you have never typed into anything.
05
The critical early decision in applied AI is the same one the roboticists face. Generalist or specialist. To humanoid or not to humanoid.
A horizontal tool is the generalist. It can attempt almost anything and master almost nothing, built to look capable across a thousand tasks in a room designed for humans. It is a wonder in a demo. The room applauds. Then the demo ends.
Real work does not happen in demos. It happens on the line, the same operation a thousand times a day, where a tenth of a second and a single misread input compound into money won or lost. There the generalist is the wrong machine and everyone on the floor knows it. What wins is the thing engineered for one job and nothing else, shaped around the task until it disappears into it. Not a clever arm reaching for everything. A purpose-built rig that does the one thing perfectly and never tires of it.
A vertical build is the specialized machine. It is not built to do everything. It is built to do this operator's work, at this operator's scale, inside this operator's idiosyncrasy, and to earn its keep doing it. We make Gainfully Employed Agents. Agents with a real job, a real owner, and a P&L they visibly improve.
The labs proved the point this month. OpenAI and Anthropic both stood up forward-deployed joint ventures, whole companies whose only job is to sit inside a customer and bend the model to its work, one account at a time. If a generic coworker were enough, they would not be sending their own engineers into the building. The hard work lives below the model. They are pointed at the enterprise, with Wall Street capital behind them. The mid-market is a different country, and it is ours.
06
The destination is not in doubt. Only the timing is.
Two years ago the agents could not do the work. Today they cross from demo to production on narrow, well-scoped tasks, and they still fail often on anything longer. That failure rate is the whole opportunity. For a horizontal product it is a thing to hide in the demo. For a vertical build it is the job, the thing someone has to absorb on every model upgrade so the operator never sees it.
The cost of intelligence falls and keeps falling. The capability climbs and keeps climbing. Put those two curves together and the conclusion is not a maybe. Every business that runs on a person's head will run on a system built around that person's head. The only question any owner faces is whether they get there before the competitor across town does.
The buyer already feels it. The old question was, what is the risk of changing. The new question is, what is the risk of standing still while everyone else compounds. Standing still became the expensive option, and nobody announced it.
07
Here is the thing the clone cannot copy.
Imagine a perfect copy of our team and our codebase, handed the same frontier models. For a horizontal product, the clone wins the day the next model ships, because the model was the whole product. For a vertical build, the clone has nothing, because the product was never the model. The product is the accumulated escalations, the operator's edge cases learned one override at a time, the rules baked into code, the migrations already absorbed. Months of within-customer learning that exists nowhere else and cannot be downloaded.
The model underneath is fungible. The Operator is not. We route across the whole market, open source for the bulk work, frontier models from whichever lab is ahead this quarter, small fine-tunes where the domain is narrow enough. As the top models start to look more alike, that freedom of movement gets more valuable, not less. The customer never enters the lock-in conversation, because we kept them out of it.
That is the Loop. It compounds quarter over quarter, and it is the only moat in this business that gets deeper with time.
08
We could not go horizontal if we tried, and the reason has nothing to do with taste. It is the deal.
We never take a fixed fee. We take equity, or revenue-share, and we are paid only when the operator wins. That single fact forces everything. A thin layer that works in five sectors and shines in none cannot move one operator's P&L this quarter, so we cannot afford to build it. A lab's deployment team is paid when the contract ends. We are paid when the business is worth more, year after year, which means our only incentive is to keep absorbing the work.
So we do not pick industries. We pick Operators, and we pick hard ones. Complexity is the point. The messier the knowledge, the worse a horizontal tool performs and the faster the Loop compounds. The simple, standardized work is exactly where a configured tool wins, and we hand it over gladly.
09
The fair challenge to a model like ours is the one we ask ourselves. How many operator builds can a small team carry before the maintenance load caps the portfolio.
We think in systems, and that changes the ceiling. The fear assumes every build starts from zero. Ours never do.
Only one layer is ever bespoke, this Operator's escalations and edge cases. Everything around it we built once and run forever, the harness, the evals, the migration machinery, the regulatory checks. Each new Operator plugs into a system that already exists and got stronger from the last one. The system carries the weight, not the team, so the portfolio expands as far as we choose to push it. The maintenance a horizontal tool dumps back on the owner is the exact work we are paid in upside to take off their hands. That is not overhead. That is the deliverable.
Behind the portfolio question hides the runway question, which is the better one. Equity and revenue-share carried for years means our cash is backloaded and concentrated. The modality mix answers it. Transform brings co-investor capital that funds the rebuild years. Replace and Co-build pay revenue-share within months of deployment. We do not run on patience alone.
There is one ceiling and we name it instead of outrunning it. The day we start picking deals to grow the portfolio rather than to pick the right Operator, the Loop stops compounding and the whole thing breaks.
10
If you own a business and you are deciding between a tool and a build, three questions settle it.
Pull last month's exceptions, the overrides where the standard process failed. If the top three are unique to your business, a tool will keep hitting them and a build can absorb them.
Ask who answers to your customer when the system gets a decision wrong. A vendor who caps liability at last year's fees is selling you a tool. A partner with their own capital in the outcome is selling you a system.
Ask who absorbs the next model migration when it ships. You, pulling your team off the road to retest everything, or a partner who finished before you saw the release. If there is no partner, the answer is you.
Three yeses and the question is no longer whether to go vertical. It is who you go with.
11
We are Parallel Partners. We rebuild small and mid-sized businesses into agentic-first operations with our own capital and equity at risk. We do not take fixed fees, and we do not sell tools.
Three ways in. We acquire a business with investors and rebuild it agentic-first, which we call Transform. We rebuild one process inside an existing business for equity or revenue-share, which we call Replace. Or we build a new venture with a founder inside a joint entity, which we call Co-build.
The mid-market is small enough to move fast in and large enough to build a portfolio inside. The window is open now, and windows close.
12
People ask me what to build next. My answer is always the same. Become deeply self-aware, and when the right thing arrives you will know it.
Twenty years in this work, a company built and sold, and the right thing finally arrived. It is not the model. It is the operator the model could never reach, the human knowledge that runs the real economy and has been waiting this whole time for something that could finally put it to work.
If you are an owner-operator and the three tests land yes, the first step is sixty to ninety minutes with one of us. Not a deck. An intake.
Back to building.
**Horizontal AI**
A product built once and configured by the customer, sold as 'works for any business'. The labs ship it, the customer configures it. Codex, Cowork, ChatGPT.
**Vertical build**
An agent system built around one operator's actual workflow, at that operator's scale, with that operator's rules baked into the system as tools the agent can call.
**The operator is the spec**
In a mid-market business, the working specification lives in the operator's head, not in documents or systems. Any agent that wants to do real work has to learn this specific operator, not the sector in general.
**Gainfully Employed Agents**
Agents built for one operator's job, with a real owner and a P&L they visibly improve. The opposite of a generalist tool that looks capable across a thousand tasks and masters none.
**The Loop**
The accumulated, within-customer learning a vertical build compounds over time: escalations, edge cases learned one override at a time, rules baked into code, migrations already absorbed. It exists nowhere else, cannot be downloaded, and gets deeper with each quarter.
**Transform**
PP acquires an existing business with investors and rebuilds it agentic-first.
**Replace**
PP rebuilds a single work process inside an existing business, paid in equity or revenue-share.
**Co-build**
PP builds a new venture with a founder inside a joint entity.
**Horizontal AI versus vertical AI. What is the real difference?**
Horizontal is the humanoid. One product built once, configured by the customer, sold as works for anyone. Codex, Cowork, ChatGPT. Vertical is the specialized machine, an agent system built around one operator's workflow with that operator's rules baked in as tools the agent calls. Horizontal scales like software for the seller. Vertical compounds on the operator's P&L.
**Why doesn't horizontal AI work for the mid-market?**
Because a real business runs on knowledge that was never written down. Supplier quirks, margin policy, favors owed, the escalation rules nobody typed into anything. None of it is on the public web or in the CRM. A horizontal tool the customer configures cannot reach it. A vertical system built inside the workflow learns it, one override at a time.
**What does the operator is the spec actually mean?**
The working specification of a mid-market business does not live in documents. It lives in the operator's head. Any agent that wants to do real work has to learn this specific person, not the sector in general. We build around the operator. Horizontal AI tries to configure around them and fails.
**How do the engagements work?**
Three ways. Transform, where we acquire the business with investors and rebuild it agentic-first. Replace, where we rebuild one process inside an existing business for equity or revenue-share. Co-build, where we build a new venture with a founder inside a joint entity. Every deal carries equity or revenue-share. No fixed fees, ever.
**How does an owner choose between a tool and a build?**
Three tests. Are your top exceptions unique to your business? Yes means a tool keeps failing them. When the system gets a decision wrong, does the vendor share the consequence or cap their liability? Capped means they sold you a tool. When the next model ships, who absorbs the retest? If it is you, you are paying twice.
**A deep, equity-only model with a small team. How does that scale?**
It scales by staying small on purpose. Each build compounds for a decade instead of paying once. A handful of operators carried long-term reaches portfolio economics a many-engagement consultancy never touches. Runway through the build years comes from the modality mix. Transform brings co-investor capital, Replace and Co-build pay revenue-share within months.