How I work with you: a private AI partner, not an agency

What partner means, in practice

A partner grows when you grow, and is on the hook when something breaks. That's the entire idea. Everything below follows from it.

An agency sells hours or campaigns to as many clients as it can serve. A SaaS vendor sells the same product to everyone and lets you figure out the rest. I do neither. I take on a small number of companies, six to eight at most, and for each one I build and then run the operating layer its people work through every day: the agents, the system of record behind them, the connections into the tools the company already uses, the lessons that make the team good at it.

Two rules keep that honest. First, the split of ownership is fixed before anything is built: your business, your data, your documents, your outputs and your configuration are yours; the factory that makes them, the runtime, the skills, the connectors, the review pipeline, the private models, is mine or licensed from its makers. Second, the money follows the outcome. A fixed build, a monthly fee for owning the system in production, raises only at verifiable gates, and a share only on ventures I build and run end to end. No hourly rate exists anywhere in the arrangement.

The companies this is written for are private by nature: family firms, owner-run e-commerce groups, offices where deadlines and documents are the work. Their reason for choosing this shape over an agency is the same one every time: nobody wants people to find out what they're doing with AI, or with whom. Confidentiality isn't a clause here. It's the product.

Step one: measure what actually runs

Before I propose anything, I count what is already in use, over a window, from the system's own records. Configured is not used. Installed is not adopted.

For a company that already has agents from me, that means reading the scheduler and the run logs, not the list of what was set up: how many groups did real work in the last 30 days, which people have an operator and use it, which scheduled jobs finished and which have been silently failing for two weeks. For a company starting from nothing it means the harder version: where does the truth live today? Usually in one spreadsheet on one PC, a folder of paper in the office, and an ERP three people can open. I read the formulas and every distinct value anyone ever typed, because those are the places where the work quietly went around the system.

The numbers from that pass go into the proposal as the baseline, dated, with the window stated. When I later say a gate has been met, it is measured against that baseline, in the company's own numbers.

Scope is decided on paper, once

Nothing is negotiated on a call. The scope, the price, the acceptance criteria and the timeline sit on one page you say ok to. If the page and the conversation ever differ, the page applies.

The page has the same shape every time. Where we are: the measured baseline. What gets built in the sprint, listed, and what explicitly does not, listed with the reason. What "done" means: acceptance criteria you can check without me in the room. What the monthly fee covers from the day the sprint is accepted, and what it doesn't. The growth ladder, with each step tied to a gate. The rules before the first byte: private infrastructure, what I won't build, how people are treated. A date table.

I write it in your words. A proposal for a company that already runs on my agents opens with the services its team already uses, by name, because that's what the owner recognises, and it names a preview as a preview, because one slide that overclaims costs more trust than ten that are exact.

The 30-day sprint

The first build is one system, delivered in 30 days from your ok and your payment, against acceptance criteria written before work starts. A single one-off price covers it, paid in advance, and nothing else is charged in that window.

For a company with scattered tools the sprint is centralisation: one address, passkey sign-in with no passwords, access per person and per team, the task manager that replaces whatever the team was tracking work in, the tools they already use moved inside so there's no second login, and the agent embedded as one more door into the same operator the team already talks to in Slack or Telegram. For a company with no system of record, the sprint starts one layer lower and builds the company its own HQ; that shape is documented in how a company gets its own HQ.

Acceptance is a list, not a feeling. A typical sprint is done when: the team signs in with their own passkey; the approved workflows run in the new system and the old tool is off with a recoverable export kept; the existing tools work inside it; every person has their lessons and their prompt maker; the operator answers inside the system with the same capabilities as in a direct message, within each person's permissions; the owner can see usage per person; and the onboarding call has happened. Seven checks, each one verifiable by you.

The first three weeks look quiet from outside, because centralisation is plumbing. I send a short daily note on what moved and we talk once a week. Then everything shows up at once.

Every person gets an operator, and lessons

The goal is never to replace people. It is for every person on the team to do more through the system, measurably, and for the know-how to live in the system rather than in one head.

Each person gets their own operator, bound to their own access: it can read what they can read, propose what they could propose, and record what they are allowed to record, with a named human deciding anything that touches money, official records, stock or another person. Each person also gets their own lessons on the Scalably Learning Hub, live today and rolling out to client teams by invitation, which they have to pass, and a prompt maker for their role. A new hire passes the same lessons, gets the same operator, and works from day one. Nobody is the only person who knows how something works.

The owner sees usage per person. That is all the owner sees from me. If an owner asks me whom to let go, the answer is no, and I'll say it here so nobody has to ask: I provide the metrics, never the lists. Decisions about people stay with the person who employs them.

Ownership: what the monthly fee buys

From the first of the month after the sprint, one monthly fee pays for owning everything in production: monitoring, fixes, incidents, integration and API changes, model migrations, evaluation and replay, workflow changes, team training, usage metrics, and the monthly report. No contractual hours. Measured by outcome: what runs, who uses it, what was freed.

It also buys the part most people underestimate: what appears in AI reaches your company in days, not months. A new model, a new protocol version, a new tool: I evaluate it against your own replayed decisions and bring it in when it wins. That is the head-of-AI role, taken seriously, without the salary line.

What the fee does not cover is written just as plainly. New modules and new systems are defined and priced separately, always before building, and a team member can request one from inside the system: the request is recorded, scoped and priced, and its delivery and usage stay on record. And the fee never covers manual execution of the team's own tasks, a spreadsheet, a one-off document, hand data work, even when the request arrives through the agent; the retainer maintains the system and its agreed workflows, and a new automation for such a task is scoped on its own. That line is the one owners push back on first and thank me for by month three, because it's the difference between an operating layer and an outsourced employee.

Consumption above the capacity the signed scope includes is separate and never a surprise: model inference, GPU capacity and third-party services run at cost, on your accounts wherever the provider allows it, with a written estimate before anything is switched on and an actual figure every month. In a typical engagement nothing changes on day one: the system runs on the model subscription the company already pays.

If we stop, everything built stays yours, with an export of your data; the notice period after the initial term is in the agreement, not in this article.

Gates, not dates

The fee goes up only when something verifiable has happened, and each step is a one-page addendum, never a renegotiation of the whole agreement.

The ladder has the same rungs for every company, with the thresholds set from that company's baseline. Adoption: 30 days after onboarding, most of the team does real work through the system, the old tool is off, and the organisation map is live, every person mapped to every recurring task, what the operator does, what is still manual. No raise at this step; the sprint bought it. Second wave: two more systems moved in, and a value ledger, kept in your numbers, showing freed hours and retired tools worth a multiple of the fee. Then the steps that belong to that company's own ambitions, whatever they are, each with a gate you can check.

Shared upside is the top rung and it is deliberately separate. On ventures I build and run end to end for a company, from the proposal through the build to the daily operation, a profit share applies, with the costs that come off the top named, in its own one-page agreement signed before the first one goes live. The monthly fee is never charged to any venture's P&L, and AI added to the company's existing business is covered by the fee and carries no share. The percentage is not in the proposal on day one, and that is on purpose: a number attached to something unbuilt is the one thing in a package a client can argue with.

Liability first: the contract protects you before you ask

The agreement puts unlimited liability on my side for misuse or disclosure of your data, and it was drafted that way before any client asked for it. Every client gets the same paper.

The shape is simple to state. My liability is uncapped for the harms that would actually hurt you: misuse or disclosure of your data, a negligent data incident, a careless choice of sub-processor, gross negligence. Ordinary operating events are capped at twelve months of fees and named one by one, so there is no grey zone. No indirect loss either way. Every sub-processor is listed by name, region and who pays for it, and if one of them fails despite a compliant choice on my part, the cap applies and the pass-through is yours.

Data, IP, outputs and configuration are yours, with export on request. The platform is mine or its licensors'. Model usage runs on your own account where the provider allows it, which means the provider's training setting is yours to switch off; I check it with you before we sign, because the default differs by plan and by provider.

Why write it this way? Because if I ever get it wrong, it could cost me everything. That's the point. I want to walk in with that trust already on the table, before anyone has to negotiate for it.

What I don't build

Fake reviews, fake users, anything your brand wouldn't sign. Refused once in writing, it stays refused.

The rest of the list is shorter than you'd expect and just as firm. I don't replace people with agents and I won't write a proposal that sells it that way; the numbers on the page are usage per person, and there is no headcount column. I don't bill hours. I don't hide a margin in consumption. I don't promise a Slack replacement or any other migration I haven't verified can be done on one server with the agent as a first-class participant; where I'm not sure, the proposal says "researched, not guaranteed", and a pilot is offered instead. And I don't put a client's name, numbers or documents anywhere public, which is why this article names none.

What it costs

A build quoted by scope and fixed before work; production ownership from $5,000 a month; consumption above the signed capacity at cost; raises only at gates; a share only on ventures built end to end.

The build is quoted separately because a read-only reporting workflow and a multi-system agent with approval-bound writes carry different work and different risk; the AI agent development cost guide walks through what moves that number. The monthly figure is the one I publish, because the expensive part of an agent is owning it after launch, and that is the part most proposals leave blank.

For a company where every person on the team has their own operator, the team channels run through it daily and a few dozen scheduled jobs finish overnight, the value ledger at gate two has to show a multiple of the fee in freed hours and retired tools, in the company's numbers, or the fee doesn't move. That is the deal in one sentence: I get paid more only when you can measure that you got more.

Evidence you can check yourself

Since February 2026, the agent platform I run has recorded 10,908 scheduled runs across 5 production hosts and 169 registered agent workspaces; 98.0% of them finished without error, and the last 30 days ran at 96.3%. Counted from the scheduler's own status column on 12 September 2026; a finished run is a runtime status, not a human acceptance judgement, and the count is a lower bound because four of the five hosts only started keeping run logs in July.

The last 30 days of that window are 3,821 runs. A separate 30-day measurement across two of those systems recorded model cost of $0.97 and $1.01 per task (14 July to 13 August 2026); different workloads, so not a price list, but proof that cost per task can be measured, and that in those two systems the model was not the expensive part. The way I scope and run business agents is in AI agents for business; the privacy-safe version of one company's whole operating layer is in AI agents in agriculture.

The parts of the factory that can be public are public and installable: the skills library and the MCP server gallery, each release checksum-verifiable without asking me. The lessons live at learn.scalably.io, live today, invitation only, rolling out to client teams. None of that proves my judgement; it proves the machinery exists and runs, which is the part an agency deck can't show.

How to start

Send one recurring job your company would genuinely hand off, and where the work starts today.

Email [email protected]. I'll measure what already runs, put the scope, the price and the acceptance criteria on one page, and the sprint starts on your ok. If the page is wrong for you, you've lost a week and learned what you actually need; if it's right, the quiet three weeks start the day you pay.

Reading this in Serbian? The same page, adapted, is at scalably.rs/partner-za-ai.

Common questions

Is Scalably an agency?

No. An agency sells hours or campaigns to many clients at once. I work as the external AI partner of a few companies at a time, embedded in how each one runs, on infrastructure the company owns, with a monthly fee that pays for owning the system in production and a share only on ventures I build and run end to end. The contract is written to protect the company first, before it asks.

How do I choose an AI consulting partner?

Ask four things. Who owns the system after launch, and what does that ownership cover in writing? What is measured, and can you read the measurement yourself? What happens to your data, named per provider and per region? And what will the partner refuse to build? A provider who cannot answer all four in a page is selling a project, not a partnership.

What do you own and what do I own?

You own your business, your data, every document and output the system produces, and the configuration that makes it yours, with an export on request. I own the factory: the agent runtime, the skills, the connectors, the review pipeline and the private models, or license them from their makers. That split is in the agreement, and it is the reason the monthly fee is a fee and not a licence.

What happens to my data?

It stays on infrastructure you own or that runs for you alone, with every external provider and region named in the agreement before the first byte moves. Private models run on my own hardware where a workload needs it. Model usage runs on your own account where the provider allows it, so a provider's training setting is yours to switch off, and I check it with you before signing.

How much does an AI implementation partner cost?

With me: a build quoted by scope and fixed before work starts, then production ownership from $5,000 a month, then gates that raise the fee only when a verifiable event has happened, and a profit share only on ventures built and run end to end under a separate agreement. Consumption above the capacity the signed scope includes is separate: model inference, GPU capacity and third-party services run at cost on your accounts, estimated in writing before anything is switched on and reported monthly. Hourly billing does not appear anywhere.

How do we start?

Send one recurring job your company would genuinely hand off, and where the work starts today. I measure what already runs, put the scope, the price and the acceptance criteria on one page, and the sprint starts on your ok. The first three weeks are quiet; the fourth is when everything shows up at once.