My records are a mess across spreadsheets and email. Can I still use AI?

By Ranjit Bhinge, founder of Blur Studio · Updated · 7 min read

Short answer: Yes, in most cases. You do not need clean, perfect data to start with AI. You need data that is good enough for one specific process. Take 30 real records from that process and run five checks: can you find them, are they complete, are they consistent, are they current, and are you allowed to use them. Modern AI tools such as Claude, a general-purpose AI assistant, can read messy text, emails and scanned pages, so the mess is often less of a barrier than you fear.

Ranjit Bhinge

Your customer and job records are scattered across spreadsheets, email and paper, and you want to know if that rules out AI.

I'm Ranjit, and I run Blur Studio: 15+ years building software, 120 apps taken live, and clients who have gone on to raise $10M+. I build AI systems and processes that help businesses do a lot more with a lot less time. I also work differently from most of what else is out there, more on that further down this page. Say hello.

If you want to know which process in your business is worth testing first, the AI Strategy Sprint helps you pick it in 3 days.

The short version: test one process, not your whole business

"Data readiness" sounds like a company-wide project. For a small business it should not be. Most of the guides written about it are aimed at large companies with data teams, and they tell you to audit everything before you start. That advice would put off a ten-person firm for a year.

A better question is smaller: is the data for this one process good enough for an AI to help with it? Pick a process (answering price questions, chasing late invoices, sorting enquiries). Look only at the data that process touches. That makes the test a morning's work, not a project. If you want a wider look at your whole setup first, my AI readiness checklist covers people, tools and goals as well.

What AI can read, and what it cannot

Here is the good news. Language AI is built for untidy text. Claude's documentation says it supports PDFs, and that because this relies on its vision abilities, each page is processed as an image. In plain words, it can look at a page the way you do, including a scan, a table or a form. It accepts plain text and CSV files directly too. Spreadsheet and Word files need converting to text or PDF first, according to the same docs.

So messy does not always mean blocked. Here is what usually matters:

  • Usually fine: emails, PDFs, scanned letters, free-text notes, spreadsheets with odd layouts, supplier invoices in different formats.
  • Usually a problem: information that is in someone's head, records that contradict each other, missing fields the process depends on, and anything you do not have permission to use.

The difference is simple. AI can read an ugly page. It cannot invent a fact that was never written down, and it cannot know which of two conflicting records is right.

The five checks, on a sample of 30 records

Pull 30 real records from the process. Not your best 30 and not the worst. Take the last 30. If it is invoices, take the last 30 invoices. If it is enquiries, the last 30 enquiries. Put them in a list and go through each check.

Check 1: Can you find them?

For each record, how long did it take to locate? Count how many lived in more than one place, such as part in email and part in a spreadsheet.

What breaks: an AI cannot use what it cannot reach. If half your quotes are in someone's personal inbox, a quote assistant never sees them.

Cheapest repair: do not move everything. Name one shared place (a shared mailbox, a folder, a spreadsheet) for new records from today, and leave the old ones where they are until you know they matter. My piece on an AI knowledge base shows how scattered documents get pulled into one searchable place.

Check 2: Are they complete?

For each record, list the three or four fields the process needs (customer name, amount, date, status). Count how many are blank.

What breaks: if 8 of 30 records have no date, anything that depends on dates is guesswork. The AI will not tell you it is guessing unless you make it.

Cheapest repair: make the key fields required for new records, and fill the gaps in old ones only if the process needs them. Often the old gaps do not matter.

Check 3: Are they consistent?

Look for the same thing written in different ways: "Acme Ltd", "ACME Limited", "Acme". Dates in different formats. Statuses like "done", "complete" and "finished".

What breaks: counting, matching and reporting. A person sees that these are the same customer. A formula does not.

Cheapest repair: a short list of allowed values for each field, used from now on. For old records, AI can help tidy names and formats when you ask it, and you check the result.

Check 4: Are they current, and free of duplicates?

Mark records that are out of date (old prices, former staff, closed jobs) and records that appear twice.

What breaks: an assistant that answers from an old price list will quote old prices with full confidence. This is the most expensive failure on the list, because the answer looks right.

Cheapest repair: one named owner and one review date for each source document. Delete or archive what is dead. Archive is safer than delete.

Check 5: Are you allowed to use them?

Check whether the records hold personal details (names, contact information, health or financial details) and whether your customers or contracts limit how that data is used or where it is processed.

What breaks: not the AI, but you. The UK data protection regulator, the ICO, publishes guidance on AI and data protection that covers topics such as accuracy, fairness and data minimisation, which means using only the personal data you need. If you handle regulated or sensitive data, ask a lawyer or your regulator's guidance before pasting records into any tool.

Cheapest repair: remove what the task does not need. For a first test, replace names with codes.

How to score it

Result on your 30What it meansNext step
Mostly findable, complete and consistentReadyTry an AI pilot on this process now
Findable but patchy or inconsistentReady with cleanupFix the key fields, then pilot. Use AI for tidying
Scattered across places, in people's headsNot yetGather first. One shared place, simple rules for new records
Cannot be used for privacy or contract reasonsStop and checkGet advice first, or use anonymised data

A worked example, with made-up figures

Say a small plumbing firm wants AI to draft replies to price enquiries. It takes its last 30 enquiries and the price list. Findability: 24 are in the shared inbox, 6 are in a manager's own email. Completeness: 27 name the job type; 3 do not. Consistency: the same service appears under four names. Currency: the price list on the shared drive is two years old, but a newer one sits on a manager's laptop. Permission: enquiries include customer names and addresses.

The verdict: not hopeless, but the stale price list is the real danger. The fixes cost almost nothing: move the current price list to the shared drive with an owner and a review date, agree four standard service names, and anonymise addresses for the test. Total effort, about half a day. The mess in the inbox is not the blocker. The old price list is.

When you can skip the cleaning and start anyway

Do not clean first if the task is reading, not counting. These work on untidy data:

  • Summarising long email threads.
  • Pulling dates, amounts and names out of scanned invoices or contracts (see AI document processing).
  • Drafting replies from a few example answers.
  • Answering questions from a set of policies or manuals.

Claude's own guidance on long documents suggests asking it to quote the relevant parts of the text first before it does the task, which helps it stay on the source. That habit also lets you check its work against the page. The same idea works for you: have the AI cite where each answer came from.

Anthropic's guide to building with AI says that for many applications, a single AI call with retrieval and in-context examples is usually enough. In plain words, try a small version with a few good examples before you start a big data project.

Do clean first if the task is counting, matching or deciding across records: monthly reports, finding duplicate customers, working out who is overdue. A wrong answer here looks right, because it is a number.

What not to do

  • Do not start a six-month data clean-up before you have tested one use. You will clean things that never matter.
  • Do not trust an AI answer on data you have not spot-checked. Check ten.
  • Do not feed everything in. Start with the least data that works. NIST, the US standards agency, publishes a voluntary AI Risk Management Framework if you later want a structured way to think about risk.
  • Do not buy a data tool before you know what your process needs.

The picture across AI for small business is the same: small, specific pilots beat big preparation. And once you have picked the process, my pages on business process automation and AI implementation show what comes next.

How I'd approach this

I would run the five checks this week on one process, and I would give it half a day. If the answer is "ready with cleanup", I would fix only the fields that process needs and start a small pilot. The pilot teaches you more about your data than any audit, because the AI shows you exactly where it trips.

The question that matters most is not "is my data clean". It is "what would a wrong answer cost, and who would catch it". If a person reviews every output at first, messy data is a manageable problem. If a customer sees the output directly, it is not.

If you are not sure which process to test, the AI Strategy Sprint is a paid 3-day first step where you and I find where AI could transform your business, and you get the exact steps written down, starting with a free 15-minute AI brainstorm call.

Questions people ask

Do I need clean data before I use AI?

No. You need data that is good enough for one process. AI can read untidy emails, scanned pages and notes. It cannot fill in facts that were never written down, so check completeness for the fields that matter. My readiness checklist covers the wider picture.

How many records do I need to test AI readiness?

Thirty recent, real records from one process is enough for a first check. Use the last 30, not a hand-picked set, so the result reflects normal work. The aim is to spot the pattern of problems, not to prove anything statistically.

Can AI read scanned documents and PDFs?

Yes. Claude's documentation says its PDF support works through its vision abilities, so it processes each page as an image. Quality drops with very poor scans, so spot-check results. See AI document processing for worked uses.

Is it safe to put customer data into an AI tool?

It depends on the tool's data terms, your contracts and the rules where you operate. Remove what the task does not need, replace names with codes for a first test, and ask a lawyer if you hold sensitive or regulated data. This is information, not legal advice.

What is the cheapest first fix for messy data?

Name one shared place for new records, make the key fields required, and set a short list of allowed values. Then use AI to tidy old records, and check the results. See business process automation for what to connect next.

More in this topic: Problem solving: AI strategy and adoption (24 guides) →

Sources

Ready for the next step?

A free 15-minute AI brainstorm. We find 3 broad areas where AI could help your business.

Ranjit Bhinge You talk to me, Ranjit. I reply within a few hours.
See the AI Strategy Sprint

Not sure what you need? Tell me what is going on. I will point you the right way, even if it is not me.

Blur Studio · 30 minute call

Book a 30 minute call

You talk to me, Ranjit. If you do not need me, I will tell you.

  • We talk about your business and what is going on.
  • You leave with a clear next step, even if it is not me.
  • I reply to every message within a few hours.
Loading the calendar…