Get your data ready before you hand anything to AI

Before you hand a task to AI, ask yourself a simple question: where does the information it needs actually live? In many companies, the answer is a worrying list: accounts at three or four banks in different currencies, an accounting system, a half-maintained CRM, personal inboxes, WhatsApp groups and a dozen spreadsheets holding it all together.

AI that reads wrong data produces wrong answers, with confidence. That is why the first layer of the method I apply is called "your data, connected". Nobody sees it. Yet it decides everything else.

It comes down to three pieces of work: connect, clean, and decide who approves.

Connect: banks, CRM, invoices, email

Connecting means two things. First, each source is read automatically, with no re-keying. Second, you can move from one source to another without doubt: this CRM customer is the one on the invoice, and this transfer settles that invoice.

Start with an inventory: where each piece of information lives, who updates it, how often. Then connect the four sources that matter everywhere.

• Banks. Statements from every account and every entity, sometimes in several currencies, collected automatically through a bank connection or a regular export, in one place.
• Invoices. Each invoice with its number, customer, amount, due date and status, read from your invoicing or accounting system.
• CRM. Each customer with a single identifier, the same one used in invoicing. The VAT number or the company registration number is a natural candidate.
• Email and messages. The shared inboxes first (enquiries, accounts, customer service), then the WhatsApp numbers your customers actually use, each message attached to the right customer or file.

At Oliverlist, the automation we started in 2020 runs from contact import to invoice. It holds because each step reads the data of the step before. When AI entered the product in May 2023, it landed on data that was already connected. Sorting a prospect reply, which AI does in production for us today, is only useful if that reply is attached to the right contact and the right client.

Clean what matters, not everything

Cleaning is daunting because people picture it as total. It does not need to be. Clean what the first AI task needs, not ten years of history.

Five things to check:

• Duplicates. The same customer under three different spellings; the same invoice entered twice.
• Identifiers. One number per customer, present in every system.
• Empty fields. An invoice without a due date cannot be chased.
• Formats. Dates, amounts and names written the same way everywhere, including day-month order and currency codes.
• Parallel files. The "tracker_v3_final" spreadsheet that three people update on their own.

Then set one source of truth per type of information. The invoiced amount comes from the invoicing system, not from the CRM. Cash comes from the bank, not from a spreadsheet kept by hand. When two sources disagree, everyone knows which one to trust.

Finally, stop the mess from coming back: mandatory fields at entry, drop-down lists rather than open text, and an automatic check that flags duplicates and gaps every week. Without these safeguards, the clean-up has to be redone a few months later.

Decide who approves

Once the data is connected and clean, AI can prepare a great deal: a payment reminder, a reply to a customer, a bank reconciliation, an alert. One question is often forgotten: who approves?

For each flow, write down three things.

• An owner. A name, not a department. The overdue-invoice reminder is approved by one named person, the customer reply by another.
• Thresholds. What goes out without approval, what needs one click, what goes up to management. For example: a first reminder approved in one click, a reminder above an amount you set approved by the finance director. No payment ever leaves without a person.
• Access. The tool sees only the data it needs, read-only to begin with. Accounts and subscriptions stay in the company's name. And the data protection rules that apply to you, in Europe or elsewhere, apply here as anywhere else.

This work looks administrative. It is what lets you hand over a task without losing control.

Start with a single flow

Do not connect everything at once. Pick one task, then connect only what it needs.

Take overdue-invoice reminders. You need the invoices (amount, due date, status), the payments received (bank) and the billing contact (CRM). Three sources, one flow, one owner. Once it runs, the next task reuses what is already connected, and each new task takes less work than the one before.

Write down the starting point as well: how many invoices are overdue, how long a reminder takes to go out today. It is the only way to know, a few months later, whether the effort paid off.

Key takeaways

• AI does not fix wrong data: it repeats it faster.
• Connect first: banks, invoices, CRM and messages read automatically, with one identifier per customer.
• Then clean what the first task needs, and stop the mess from coming back.
• Decide who approves: an owner, thresholds and limited access for each flow.
• This is the "your data, connected" layer of the method: nobody sees it, but everything else rests on it.