NEWThe Runbaze platform is ready for early adoptersRead the announcement
Platform
Solutions
Company
Who we are
About Careers Contact
Follow along
LinkedIn ↗ Substack ↗
Newsroom
Get in touch Book a demo

The winners of this decade will not be the companies that bought the most AI.

They will be the ones that changed shape around it, deliberately and in the right order. Our working paper, The Principal Firm, sets out where this goes. This page is the argument in ten minutes, and what it means for your P&L along the way.

WORKING PAPER · SSRN / AMSTERDAM · SINGAPORE / AUGUST 2026

What is actually happening.

Strip away the noise and three facts remain. Each one is measurable inside your own company today.

Fact 1

The cost of doing work is collapsing

Drafting, coding, analysing, answering, processing: the execution layer of the firm is becoming machine work, and its price is falling toward the price of computation. Every competitor you have gets the same discount. The discount itself is not an advantage.

Fact 2

Two abilities stay scarce

Saying precisely what you want, and standing behind what comes out. Specification and answerability. Machines do not supply either one, and every hour of cheap execution makes them more valuable, because more work now depends on fewer decisions.

Fact 3

The classical firm is built the wrong way round for this

Firms were designed to organise scarce execution: layers of people doing, and a thin layer deciding. When execution is abundant, the pyramid inverts. The scarce roles become the people who decide what good looks like and answer for the result. Most companies feel this as chaos. It is actually a direction.

Why bolting AI on fails.

Most AI programmes add tools to a company shaped for the old economics: a licence per employee, a pilot per department, a bot in the corner of the architecture diagram. Eighteen months later the spend is real and the P&L has not moved. The pattern of failure is always one of three.

NO DIRECTION

Activity without a destination

Nobody chose the work on purpose, so the tools automate whatever was easiest to reach. Individual tasks speed up. The processes that carry the business do not.

NO GROUND

Built on data that cannot carry it

Agents inherit whatever is underneath them. Contested figures, unowned systems and processes that live in two people's heads produce confident automation of the wrong thing.

NO ACCOUNT

Nobody can answer for the output

The first serious question from a board, a customer or a regulator goes unanswered, and trust in the whole programme dies with it. This is where most pilots quietly stop.

The Principal Firm.

The paper's claim is simple: when machines perform the work, the firm that remains is organised around the people who answer for it. We call that a principal firm, and it has a definite shape.

PEOPLE

Principals decide what good looks like

A smaller core with more reach. They specify the work, keep the decisions that carry consequence, and are named against outcomes.

AGENTS

Agents carry the volume

Working from specifications rather than vibes, stopping at gates, at whatever scale the work demands. Capacity stops being a hiring problem.

ORCHESTRATORS

Orchestrators govern the fleet

A role most companies do not have yet: the people who own the specifications, the data paths and the agent roster. They are how one principal safely directs a hundred agents.

THE RECORD

The record proves all of it

What was asked, what was produced, who approved it. It answers auditors, survives staff turnover, and becomes the one asset a competitor cannot copy: the account of how your work goes wrong and what you did about it.

The controlled sequence.

You do not get there by proclamation, and you do not get there by buying tools. You get there in an order, and the order is the strategy. Skip a step and the next one breaks.

Step 1

See where you stand

Which work is worth putting agents on, what it is worth, and whether the data and the process can carry it. Weeks, fixed fee, a document your board can act on. This is the Agent Case and the groundwork around it.

Step 2

Prove it on one process

One consequence-heavy workflow live, with the gates where a person still decides and a figure agreed before the build. The number funds the next step and ends the internal argument. This is first agent live.

Step 3

Install the model, then scale it

Specification discipline, permissioned data paths, gates and the record become how the company works rather than how one project worked. Applications are added one at a time, each cheaper than the last, on one layer that also takes the agents you bought elsewhere.

The result

P&L now, direction underneath

Each step pays for itself on a figure you already report, and each step is also a piece of the destination. That is the difference between a transformation and a string of projects: nothing gets thrown away, and stopping at any step still leaves you better than you started.

Done in this order, responsible is what lets you go fast.

Gates where consequence lives are what let agents run everywhere else at full speed. A record you hold is what makes European and Asian regulation a formality instead of a programme, including the transparency duties that already apply and the high-risk regime arriving in 2027. And a named plan for the people whose work the agents absorb is what makes the change survivable inside your own organisation. Repurposing comes before redundancy, and we do not start without that plan.

These are commitments you can check, and they are written down. Five of them, on the About page.

Read it critically.

The paper is not marketing. It states the conditions under which the argument fails, so you can test it against your own market rather than take our word for it. If the argument holds for your company, the sequence above is how you act on it. If it does not, better to know now, for the price of an evening's reading.

Decide the direction. Then move.

Thirty minutes with the founder. Bring your situation, and we will tell you where the sequence would start in your company, and what it would cost to find out.