Skip to content
A spiral staircase seen from below, its steps winding up to the ceiling.

The asset one AI project leaves to the next

Process knowledge, data, models, rules and skills stay with the company after an AI project, and we measure how much of them the next project reuses.

JoinX · Last updated

A company sometimes reaches its third AI project and realises it has paid for the same things three times. Each project had its own supplier, rebuilt the data it needed and wrote its rules into a model's instructions, and at the close it left a working solution and little else. At that point the next project starts again from scratch.

This problem is the origin of the idea we call Compounding Intelligence. Models will be available to everyone, and what will set a company apart from its competitors is its data, the rules by which it decides and the experience of its people. An AI project can grow these assets, provided it is built to do so.

What accumulates

The first thing that accumulates is process knowledge, that is, how the work is really done, at what volumes, on which systems and under which constraints. Mapping gathers it and the Company Brain keeps it, updated with every new piece of work.

Then comes the data. A shared definition of “customer”, with a source and an owner, serves every process that uses it, and with Data Foundation we prepare it just once. The same goes for analyses and models, because an indicator or a score that has already been tested can be reused in another process together with the checks it has already passed.

The rules are probably the most valuable part. The decision criteria, exceptions and limits of each agent also hold for the agents to come, provided they are written in the Company Brain and recorded in Governance, and not locked inside the instructions of a single agent.

Controls accumulate too. Each agent's permissions, the run log and the evidence collected for audits remain in Governance and Security, and the next project starts from access rights already authorised and connectors to the systems already built.

Lastly, people's skills remain. A team that has learned to work with an agent, to check its work and to correct it goes into the second project already knowing what to ask for and what to verify.

An example in three projects

Imagine a bank that starts with complaints handling. In the first project it builds almost everything, from the classification rules to the single definition of customer, through to preparing the team that will work with the agent. When it moves on to credit assessments it finds the definition of customer, the system access and the approval criteria already in place, and the new project only has to build the credit assessment rules and a priority score. With the third, an assistant for the branches, it also finds the product knowledge already put in order, together with two teams that already know how to work with an agent.

This is an illustrative example, which describes the mechanism and not the measured result of a real project.

Why it often does not happen

The assets are lost in fairly predictable ways. The most common is the self-contained project, in which the rules sit in the agent's instructions and the data has been prepared in one of the supplier's environments, so at the end of the contract what remains is a solution that works as long as nobody touches it. At other times the problem is ownership, because rules and data have ended up in formats the company cannot export.

That is why in our proposal we distinguish the client's data and materials, the developments made for the client, our own components and third-party technologies, and for each we write down whose it is and under what rights it is used. Rules, data, logs and documentation remain the company's, in its systems and in open formats, and the knowledge transfer is written into the very first contract, as we explain on the About us page.

This work also has a cost: writing a rule so that it serves other agents, or agreeing a definition with every function that uses it, takes longer than solving the problem for the current project alone, and the first project pays for that time. The return shows from the second onwards.

How we measure it

We measure compounding on the process and on the project that comes next. We compare the result of each process with an agreed baseline, also counting the time people spend on checks. When a project closes we write down what it leaves that can be reused, and when the next one starts we verify what was actually reused, so comparing the two lists shows whether the second project started with an advantage and on which items.

Meanwhile models, regulations and procedures change, and assets that nobody updates lose value. With Presidio we keep the rules, data and models in production up to date, so that the next project finds what it needs in the state in which it has to use it.

The first project, built with the second in mind

We can look together at what your AI projects have left behind so far, and what they could leave.