Clean Your Data Before AI: The Four Registers Every Business Needs

I Asked a ₹14 Cr Surat Trader to Show Me His Customer List. He Opened Four Files. All Four Disagreed. That Moment Explains Why Most AI Fails Before Day One.

Last month, in a Surat textile trader's office, I asked one simple question. "Show me your customer list."

He is a sharp man. ₹14 crore turnover, second generation, ready to put serious money into AI. He smiled and opened his laptop.

He opened an Excel file. 1,340 customers. Then he paused. "Ek minute, sir." He opened Tally. 1,712 ledgers. Then his phone contacts. Then a diary his accountant keeps for "the important parties."

Four lists. Four different totals. Same business.

His manager finally said the honest sentence that most offices never say out loud. "Sir, list toh hai. Par kaunsi wali?"

We laughed. But that moment is the exact reason most AI projects are dead before day one. Not the tools. Not the team. Not the budget. The AI he was about to buy was going to read those four lists and confidently answer questions from all of them. Wrong totals, dead numbers, duplicate parties, missed dues. At machine speed, with full confidence, in perfect English.

Here is the sentence I want you to keep from this article.

AI does not add intelligence to your business. It multiplies the intelligence that is already written down. If what is written down is confusion, you get confusion, multiplied, delivered politely.

The Era of Confident Garbage

In the software era, bad data gave you error messages. The report crashed. The field stayed blank. Bad inputs announced themselves.

In the AI era, bad data gives you answers. Fluent, well-formatted, reasonable-sounding answers built on garbage. The report does not crash anymore. It lies with confidence.

This is new, and most owners have not adjusted to it. Your team will catch a blank field. They will not always catch a confident summary that quietly counted 372 duplicate customers, or quoted a price from the 2023 rate list that still lives in the same folder as the current one.

So the question that actually decides your AI outcome is not "which tool" or "which vendor." It is asked and answered before any of that.

What will the AI read on day one, and is it true?

In June I gave you the 90-Day Sprint, the week-by-week map. Many of you started. And the most common message I received was some version of this: "Sir, day one pe hi atak gaye. AI ko data diya, output galat aaya."

The Sprint assumed something I never said out loud. It assumed your registers are clean. This article is the missing week. Call it Week Zero.

The Four Registers

Every business, whatever its size or sector, runs on four registers. I use the word register deliberately. Your grandfather ran his business on physical registers, and he could open any one of them and trust what he saw. That trust is exactly what we are rebuilding, in digital form.

Before you spend one rupee on AI, these four need to pass a test I will give you at the end.

Register 1: The Money Register

Who owes you, whom do you owe, and is Tally telling the truth?

What dirty looks like: ledger names that do not match reality. "Sharma Traders", "Sharma Traders New", "Sharma T." as three separate parties. Advances sitting unadjusted for two years. Receivables that everyone knows are dead but nobody has written off. A "Suspense" ledger that has quietly become a warehouse.

Why it matters for AI: the first thing every owner asks AI to do is chase payments and summarize dues. If your receivables list contains dead balances and triplicate parties, your AI will politely send a reminder to a customer who paid you in 2024. You will lose more trust in that one WhatsApp message than the AI saves you in a month.

The one-week fix: sit with your accountant for three evenings. Merge duplicate ledgers. Write off what is genuinely dead. Adjust the old advances. You do not need perfect books. You need books where every open balance is real.

Register 2: The Customer Register

One list. Current numbers. One truth per customer.

What dirty looks like: the Surat situation. Excel says 1,340, Tally says 1,712, the phone says something else, and the real knowledge of who matters lives in a diary. Mobile numbers that changed three years ago. No record of which customer buys what, or when they last bought.

Why it matters for AI: customer follow-up, reorder reminders, and WhatsApp campaigns are the highest-value early AI wins for most Indian businesses. Every one of them reads this register. A duplicate-riddled list means the same customer gets two messages. A stale list means messages to dead numbers. Both are worse than sending nothing.

The one-week fix: pick the most complete list as the base, usually Tally. Merge the others into it. One row per customer: name, one verified mobile number, city, what they buy, last purchase date. If your business has under 2,000 customers, this is genuinely two or three days of one focused person's work. It is boring. It is also worth more than most software you will buy this year.

Register 3: The Product Register

One name per item, one price that is current, one place where both live.

What dirty looks like: the same item entered six ways. "Cotton 40s", "COTTON-40", "Cotton 40 count", "C-40". Rate lists as Excel files named final, final2, and FINAL-new-latest, all in the same folder. Nobody fully sure which one is live.

Why it matters for AI: ask AI "what did we sell most last quarter" and it will treat your six spellings as six products and give you an answer that is precisely, confidently wrong. Ask it to draft a quotation and it may pull the 2023 rate file, because to a machine, final2 and FINAL-new-latest carry equal authority.

The one-week fix: one master item list with one canonical name per SKU. Rename or map the variants in Tally. Then the rule that matters more than the cleanup: one current rate list, in one location, and every outdated version moved to a folder literally named OLD. Machines respect folders, not intentions.

Register 4: The Knowledge Register

The rules of your business, written down. This is the one nobody has, and the one that decides everything.

What dirty looks like: nothing is dirty, because nothing exists. Your pricing logic, credit rules, discount slabs by customer type, vendor payment terms, seasonal patterns, the reason you never supply a certain party without advance. All of it lives in your head and your senior staff's heads. When I ask for it, every office gives me the same answer. "Yeh toh humein pata hai." Yes. You know it. Your AI does not.

Why it matters for AI: this register is the difference between AI that answers like an outsider and AI that answers like your best employee. In the Sprint article I called it the context file and told you to build it in month two. I now believe the seed of it belongs in Week Zero, because even ten written rules change the quality of everything AI does from day one.

The one-week fix: open one document. Write ten rules of your business as if explaining to a new joinee. Pricing logic. Credit policy. Top twenty customers and their quirks. What needs your personal approval. Ten rules, one evening, one chai. This document will be worth more than any subscription you buy, and unlike the subscription, it appreciates.

The New-Joinee Test

You do not need an audit or a consultant to know where you stand. Run this test tonight. It takes ten minutes.

Imagine a smart new employee joined Monday morning. No handholding. For each register, one question:

1. Money: could she produce a true list of outstanding receivables in ten minutes?
2. Customer: could she find your top fifty customers with working mobile numbers in ten minutes?
3. Product: could she quote today's correct price for your top ten items in ten minutes?
4. Knowledge: could she learn your credit and discount rules from anything written, in ten minutes?

Whatever she cannot do, your AI cannot do either. This is the part most owners miss. AI is exactly that new joinee. Intelligent, fast, tireless, and completely dependent on what your business has written down. The new joinee at least walks over and asks Ramesh bhai when confused. AI does not know that it should be confused.

Score yourself honestly. Most businesses I meet pass one register out of four. The good news is that the gap is not a technology gap. It is two weeks of unglamorous housekeeping, and it costs almost nothing.

What Happened in Surat

Back to the trader with four customer lists.

We did not touch AI for two weeks. His team merged the four lists into one register of 1,417 real customers. About 300 entries died in the merge, duplicates and defunct parties. His accountant cleaned the ledgers, and eleven lakh of "receivables" turned out to be dead balances that had been inflating his working-capital picture for two years. Painful, and worth knowing. His son wrote the knowledge register, twenty-two rules, four pages.

Then we switched on the same AI he had almost bought a month earlier.

Payment follow-ups went out to real parties with real dues. The morning summary read one customer list, not four. When he asked "which customers who bought regularly last year have gone quiet," the answer came back correct, and it came back in seconds, because for the first time the question had one truthful place to look.

He said something afterwards that I have repeated to every client since. "Sir, AI toh badhiya hai. Par asli kamaal woh do hafte ki safai ka tha." The AI is great. But the real magic was the two weeks of cleaning.

Two more, briefly. An Ahmedabad pharma distributor discovered the same molecule entered nine ways across Tally and Excel. Week Zero cut his item master from 3,100 entries to 2,140 real SKUs. Every stock answer AI has given him since rests on that cleanup. And a Pune services firm learned the expensive version of the lesson. They skipped Week Zero, pointed AI at their proposals folder, and it drafted a client quote using their 2023 rate card. The client accepted the price with delight, and the firm honored it at a loss. The AI worked flawlessly. It just read the wrong truth, because two truths were lying in the same folder.

The Week Zero Plan

Here is the whole plan, small enough to remember.

Days 1 to 3, Money. Merge duplicate ledgers, write off dead balances, adjust old advances. Your accountant leads, you decide the write-offs.

Days 4 to 6, Customer. One merged list, one row per customer, one verified number. One focused person can do this.

Days 7 to 9, Product. One master item list, one live rate list, everything outdated into a folder named OLD.

Days 10 to 12, Knowledge. Ten rules, written by you, one evening. Not a manual. A seed.

Days 13 to 14, the test. Run the New-Joinee Test again. Pass three of four and you are ready for the Sprint's day one, and your first boring task will work on the first attempt instead of the fifth.

Two warnings from the field. First, do not turn Week Zero into a six-month ERP migration. Eighty percent clean this week beats one hundred percent clean never. Clean enough to be true is the bar, not perfect. Second, do not delegate the knowledge register. The other three, your team can clean. The rules of the business have to come from you, because you are the only one who knows which exceptions are policy and which are just habit.

The Question to Sit With This Saturday

My grandfather's generation had a phrase for how you treat a shop before you install anything new in it. Pehle jhadu, phir machine. First the broom, then the machine.

They understood something we forgot in the excitement of this decade. A machine placed in a clean room multiplies order. The same machine in a messy godown multiplies the mess, faster than any human could.

So this Saturday's question is not about AI at all.

Take the New-Joinee Test tonight. Four questions, ten minutes each, honest scoring. Which register failed worst?

That register is your Monday. Not a tool, not a demo, not a vendor call. Three evenings of cleaning, this week, on the register that failed. It is the least glamorous advice I have ever written and it will do more for your eventual AI results than anything else in this series, because every intelligent thing you switch on later will stand on what you clean now.

The broom first. The machine after. The order, as always, is the strategy.

Read next

• The One Rupee Test: How to Know If Your AI Project Is Real
• Sales Are Up 20%, Profit Is Down: What Should Your AI Tell You?
• AI Employee vs Chatbot: What Happened When 52 Indian Businesses Tried It