Data inventory

A data inventory is not a data lake or an 80-page dictionary. It is a list of the places information lives for the flow you want to touch: system, fields, owner, whether there is personal data, whether you can export.

Without it, the pilot starts at the prompt and ends at “where did this tax ID come from?”. With it, you pick a source of truth and a perimeter in an afternoon.

We do this inventory the same day as the priority matrix. They are cousins: one says which flow; the other, with which data.

What goes on the page

• Flow name (supplier invoices, tickets, orders).
• Systems (ERP, mail, Drive, WhatsApp, María’s spreadsheet).
• Fields the flow needs — not every field in the universe.
• Owner of each source. No owner: the data is folklore.
• Personal or special data? Yes / no / don’t know (don’t know is a task, not a stable state).
• Can you export? CSV, API, screenshot, “only by hand”.
• Current retention: how long you keep it and why.

What you do not need to catalogue

The dead 2012 archive. Picnic photos. The CRM nobody opens. The “just in case” of copying the whole Drive into the model. That is not inventory. That is hoarding.

You also do not need to classify every column of every sheet on day one. Only the winning flow. The rest is invented later, if the second flow is earned.

Source of truth

Mark one. If the ERP says 12 and the warehouse sheet says 9, the model must not “decide”. It must flag the conflict. An inventory that does not name the source of truth produces a third, worse number.

WhatsApp is almost never a source of truth. It is a channel. If the order only lives there, the inventory just found your maturity level 1.

How to do it in a morning

Interview the flow owner. Screenshot the three tools. One test export. Field list on a sheet. Three real examples (easy, weird, one that failed). Close: what the pilot may use and what stays out.

If you cannot export anything that morning, the pilot is not AI. It is getting a file. Write that down. It saves a month of theatre.

GDPR and programme

This inventory is the embryo of the records of processing for that treatment. It fits Kit Consulting (diagnosis, analytics, processes). Do not dress it up as “big data”. It is a page a reviewer understands.

Page template (to copy)

Flow: ___. Source of truth: ___. Systems (name, owner, export yes/no): 1) ___ 2) ___ 3) ___. Pilot fields: ___. Personal: yes/no. Special: yes/no. Out of perimeter: ___. Retention today: ___. Blocker (no export / no owner / no basis): ___.

One page. If you have four, you are inventorying the company. Cut back to the flow. The rest goes under “later”.

Sign it (owner + leadership) with a date. An inventory without a date is a rumour. Last quarter’s is contrasted, not reused blindly.

WhatsApp, mail and “the truth is in the thread”

If the order is only confirmed in a chat, the inventory just found the blocker. The pilot does not read the group: first you agree that confirmation lands in a mailbox or a sheet. That is process work, and it is what unlocks AI.

Writing “WhatsApp (no owner, no export)” on the page is a finding, not a failure. It saves three months of “the model could read the group” — it could, and it would be an incident.

Update the page when the owner or the export changes, not “when there is time”. A nine-month-old inventory with María who has left is worse than none: it gives false safety. The date at the foot is not decoration.

If the test export contains more fields than the pilot needs, cut before you store the sample. A “full export just in case” is how the perimeter breaks on the same day the inventory is born. Minimising is an inventory act too.

Bring the one-pager to the assessment even if three cells say “don’t know”. Those cells are the work. A polished page with no doubts is usually a page that was not walked against last week’s folders.

Frequently asked questions

Do we inventory the whole company?

No. The pilot flow. The full map is another project and usually dies as a sheet nobody updates.

Is the 2021 consultancy inventory enough?

As a starting point, sometimes. Contrast it with last week. Spreadsheets change owner without notice.

Can the model “discover” the data?

It can list files. It cannot tell you which is the truth or the legal basis. That is your work, or ours, with names.

Next: digital maturity and GDPR. Let’s write the page for the first flow.