Skip to content

How to prepare the data of a process for AI

The steps by which we prepare the data of a process before entrusting it to AI, from the inventory of sources to permissions, with a checklist.

JoinX · Last updated

When a company decides to apply AI to a process, the first question is almost always which model to use. The question that decides whether the project will reach production concerns the data that process rests on, and it usually comes later. In this guide we describe the steps by which we prepare that data, process by process, and close with a checklist to use on your own. The criteria by which we judge a single data item are explained in the article When data is ready for AI, while here we concentrate on the work to be done.

Start from a decision in the process

Preparing the company's data as a whole is a job that never ends, whereas preparing the data of one process has a scope that can be described. That is why we first choose the process and, within it, the step AI will have to help with, for example checking that a claim file is complete or classifying incoming complaints. From that step we work back to the data that is really needed, and leave the rest out until another process needs it. If the process has not yet been chosen, the criteria we use are in the article How to choose the first process.

Take an inventory of the sources people use

With those who run the process we reconstruct which information they look at, in which system and at which point in the work. An important part often sits in no official system, because it lives in a spreadsheet kept by one office, in a shared mailbox or in the notes a claims handler writes in the margin of the file. We put these sources in the inventory too; they are often the ones most trusted by the people doing the work. For each data item we note the source system, who updates it and how often, and whether copies exist in other systems.

Fix the definitions with those who use the data

The next step is to check that every item has a single meaning. When two functions mean two different things by the same name, we do not ask either to give up its own reading, because usually both are needed. Instead we give the two concepts two different names and write the definitions in the glossary, where anyone can find them. Each definition is approved by whoever uses that data to make decisions, who from then on becomes its owner, with a name and a role.

Measure quality before building

Before writing a line of the system we measure completeness, accuracy and timeliness on the real data of the process, over a period long enough to include the difficult moments such as the quarter-end close. The measurement shows whether the project can be done with today's data or whether a source needs fixing first, and it sets the baseline against which we will measure the improvement. For each data item we then write a quality rule with its threshold, and decide what happens when a check fails. Usually the data is flagged before it reaches an agent, and its owner receives the alert.

Decide who can read what

An agent does not need all the data that the system it works in has access to. For every data item in the inventory we establish who can read it, person or agent, and for what purpose, and which fields must be masked before they reach a model. It is worth deciding this from the start together with those responsible for data protection, because a choice made once the project is finished forces part of the work to be redone. With Security these permissions become rules applied to every access, and each access stays in a log that those who carry out controls can consult.

Document what has been prepared

At the end of the work the data of the process becomes a documented data product, with sources, definitions, owners, quality rules and permissions written in a place that whoever comes next can find without having to ask. This is the work we do with Data Foundation, the module that prepares data for people and for agents, and the result stays with the company, in its systems and in open formats. When a second process arrives that uses the same data, the definition and the checks are already there and nobody has to redo them.

There is a limit worth allowing for. The data of a process will never be completely in order, and waiting until it is is the surest way never to start. What is needed is to know where it is weak and who is dealing with it, so that the system passes uncertain cases to a person.

The checklist

Before entrusting a process to AI, we check with whoever is accountable for it that each of these statements is true.

  • It is clear which step of the process AI will help with, and who carries it out today.
  • The inventory includes all the sources people really use, including spreadsheets and shared mailboxes.
  • For every data item the source system, who updates it and how often are known.
  • Every item has a definition written in the glossary, approved by whoever uses it to make decisions.
  • Every data item has an owner with a name and a role.
  • Completeness, accuracy and timeliness have been measured on the real data, and every quality rule has a threshold.
  • It has been decided what happens when a check fails, and who receives the alert.
  • For every data item it is established who can read it, person or agent, and for what purpose.
  • Confidential data is masked before it reaches a model, and those responsible for data protection have seen the choices.
  • Sources, definitions, owners, rules and permissions are documented in a place the company controls.

When some answers are still negative, there is usually no need to stop the project. What is needed is to decide who deals with it and by when, and to write that in the work plan. If you want to know what state the data of an entire area is in before choosing the process, that is one of the assessments we carry out with Mapping.

The checklist as a PDF

The ten checks in this guide on one page, to print and fill in with whoever is accountable for the process. We ask for your contact details, then the download starts straight away.

Your details reach JoinX by email, only to respond to your request, and are not kept on the site.

Prepare the data for your first process

If you have already chosen the process, we can start from the inventory of the sources people really use.