
How to choose the first process to apply AI to
Value, feasibility, risk and adoption: the criteria by which we choose the first process to entrust to data and AI, with an illustrative example.
JoinX · Last updated
When a management team decides to invest in data and artificial intelligence, there are usually many proposals on the table already: the email assistant requested by sales, the complaints classification that customer service would like, a demand forecast that logistics has been talking about for some time. Choosing which to start with matters more than it seems, because the first project also decides how the organisation will judge the ones that follow, and if it stalls halfway the second will face an uphill start.
In Mapping we look at each candidate from four sides: value, feasibility, risk and adoption. None of the four is enough on its own, and the right process to start with is rarely the one with the highest value on paper.
Value is measured on today's process
The value of an initiative is estimated from how the work is done now: how many files come in each month, how long each one takes and how many come back because of an error or a missing document. If nobody can answer these questions, the first job is to measure, because without an agreed baseline it will not be possible to say what has changed after release.
The calculation also includes the time people will spend checking an agent's work, an item that is often forgotten in estimates. That is why in the business case for each initiative we write a prudent estimate next to a potential one, so that whoever decides also sees the less favourable scenario.
Feasibility depends on data more than on models
A process can be worth a great deal and not be ready. The data needed may sit in an archive nobody knows how to query, or two systems may give two values for the same item. The data does not have to be in order already, because putting it in order is part of the work, but it must be accessible, and someone must be able to say what it means. With Data Foundation we prepare data process by process, so the question to ask concerns the data of that process and the time needed to prepare it.
Systems carry weight too. An agent that has to read from and write to the management system needs access and permissions that IT must be in a position to grant, and it is worth knowing this before promising a date.
Risk is looked at before starting
Some processes touch on decisions that the law treats as sensitive. The AI Act classifies as high-risk, for example, systems that assess the creditworthiness of natural persons, and from 2 December 2027 those who use them will have precise oversight and documentation obligations. These processes can be tackled, with the time and controls they require, and for that very reason they are rarely the place to start.
It also matters how much the agent has to do on its own. A first project in which the agent gathers the information or prepares a proposal, while the decision stays with a person, teaches the organisation a lot and exposes it little. Whatever cannot be undone, or reaches the end customer, is decided by a person regardless, as we explain in relation to the levels of autonomy.
It needs someone accountable, and a team willing to change
The fourth criterion is the most neglected. A process is a good candidate when one person has the authority to change it and is accountable for the result, and when those who run it every day are willing to work in a different way. It is the team doing the work that uses a solution or works around it, and no technical test can replace its willingness. That is why with Adoption we prepare people while the project is being built, instead of waiting for release.
An example
To give an example, which is illustrative and does not describe a real project, imagine that in a mid-sized mechanical engineering company there are two candidates. The first is preparing quotes: customer requests arrive every day, many are alike, the feasibility criteria are written down but some exceptions are known only to the engineering department, and the head of sales wants to change the way of working. The second is the automatic adjustment of a production line, which is worth a great deal but acts directly on the machinery, and requires data from several systems and a specialist safety assessment that lengthens every release.
In a case like this we would put quotes among the initiatives to launch and the line adjustment among those not to take on for now, writing the reasons alongside. Quotes, beyond the result on the process, leave something to the next project, for example shared item master data, written feasibility criteria and a team that has learned to work with an agent. It is the mechanism we describe on the Compounding Intelligence page.
When the choice is still open
If there are many proposals and no order among them, Mapping serves to line them up. With Compass we gather the work as described by the people who do it, compare it with what can be read in the data and give management a portfolio of initiatives to launch, to prepare or not to take on for now. The list of initiatives to set aside is as much a part of the result as the others, and if an area is not worth investing in we say so in writing. When the process has already been chosen, the starting point is Deployment, which takes it into production inside the company's systems.