We have a BI system. We also make Excel.

“My aspiration is that I should be doing a hundred per cent on BI. But we make Excel.”

A manager at a national ICT distributor said this to me. His company has a substantial internal business intelligence system — years old, widely used, genuinely capable. He was not complaining about it. He was describing what happens anyway.

This is what a data problem sounds like from the inside. Not an absence of software. Good software that cannot do two or three specific things the team needs, and a spreadsheet built for each of them that has quietly become the real record.

The mechanism is always the same. One thing the system cannot do forces a task into a spreadsheet. The spreadsheet works. The next person inherits the spreadsheet rather than the system. Six months later the spreadsheet is where the truth lives, and the system holds a version of events that is close but not identical.

A definition worth being precise about: a data foundation problem is not an absence of systems. It is having more than one answer to the same question, with no agreed rule for which one is right.

How do I know whether my data is actually a problem?

The short answer: ask three people to produce last month’s sales, stock position and margin, separately, and compare what comes back. If the three answers match and arrived in minutes, your foundations are sound. If they do not match, the decisions built on those numbers rested on figures nobody could fully vouch for.

If they match only after some assembly, you have a reporting problem rather than a data problem, and it is far cheaper to fix.

In the same meeting I asked how many order-loading spreadsheets the branch produces on an average day. Three people gave three different answers, and the eventual response was that they would have to check with other teams, because some teams do not make one at all. The task taking the most hours in the room was the one nobody could measure. That is the ordinary condition of work that has no owner.

Why fix this before doing anything with AI?

The short answer: because the failure is quiet. A system built on disconnected data does not break. It produces plausible answers that are wrong in ways nobody catches for months. That is worse than no system, because a broken thing gets fixed and a confidently wrong thing gets trusted.

It is also the cheapest part of the whole exercise. Finding the few things the system cannot do, and the spreadsheets built to cover them, is usually a week of work rather than a project. Closing them is often a matter of adding one view somebody needed and never got. The businesses that see returns from AI fastest are not the ones that spent the most. They are the ones that could produce a trustworthy number before they started.

What does the first month look like?

The short answer: pick one number, decide which system is its source of truth, and make every report take it from that one place. One number, one source, before anything else.

In practice:

1. Choose the number you argue about most. Last month’s sales is the usual answer.

2. Name the system that owns it. Not the best system — the one that will own it from now on.

3. Find every other place that number currently comes from, and list them.

4. Make each of those take the number from the owner, or stop producing it.

5. Settle disagreements at step two, not by arguing over the reports.

It sounds too small to matter. It is the step everything afterwards rests on.

There is more appetite for this than the market usually assumes. In a June 2026 CMR study of more than 400 Indian MSMEs, released alongside Vi Business’s programme for small firms, 78% named operational efficiency as a key priority for the year ahead, ahead of revenue growth (71%) and customer acquisition (64%). The demand is not for cleverness. It is for the ordinary things working properly, which is exactly what foundations are for.

So why does a business with good software still run on Excel? Because a system that cannot do one needed thing quietly hands that thing to a spreadsheet, and spreadsheets do not stay where you put them. Find what the system cannot do, close those gaps, and the software you already own starts being the record again.

Client details are anonymised and used with the client’s knowledge.

Frequently asked questions

Why do employees keep using Excel when the company has a BI tool?

Because the BI tool can’t do one specific thing they need, such as a particular view, a calculation or a field nobody added. Excel can, so a spreadsheet gets built to fill the gap. It works, the next person inherits it, and over time the spreadsheet becomes the real record instead of the system the company paid for.

What is a single source of truth?

A single source of truth is the one system your business agrees is correct for a given number. Every report that shows that number takes it from there, not from a spreadsheet or a copy. It doesn’t have to be one system for everything. It has to be one system per number, and everyone has to know which.

Do we need a data warehouse before using AI?

Most small and mid-sized businesses don’t. What you need is one agreed source for each number you make decisions on: sales, stock, margin, receivables. That is usually a decision and some clean-up inside software you already own, not a new platform. A warehouse makes sense later, once several systems need combining regularly.

Can AI work with messy spreadsheets?

It can read them, but its answers inherit the mess. If two spreadsheets disagree on last month’s sales, an AI tool will pick one or blend them, and give you a confident answer either way. That is the risk: the output looks finished, so nobody checks it. Agree which source is right before asking AI to work from it.

Should we replace our BI or ERP system?

Rarely. When teams work around a system, the cause is usually two or three missing views or fields, not a bad system. Replacing it takes months and often recreates the same gaps, because nobody wrote down what the spreadsheets were covering. List what each spreadsheet does that the system doesn’t, and close those gaps first.

Who should own data quality in a small business?

One named person per number, not a team and not “IT”. The owner decides which system holds the number, fixes it when it is wrong, and settles it when two reports disagree. Without a named owner, disagreements get argued meeting by meeting and nothing changes. Usually the right owner is whoever already gets asked for the figure.