Data quality is the real AI strategy

The pattern we keep seeing

We've worked on AI strategies with charities of different sizes, sectors, and levels of digital maturity. One pattern repeats: the organisations that get value from AI are not the ones with the best AI tools or the most ambitious plans. They're the ones with usable data.

Every time. Without exception.

A charity with clean CRM data and a free ChatGPT account will get more from AI than a charity with messy data and an expensive enterprise AI subscription. As we put it in the AI Playbook: AI doesn't fix messy data, it amplifies it. The data is the strategy. Everything else is secondary.

What "good enough" looks like

Here's the good news: "good enough" data for AI is a lower bar than most people think. You don't need a pristine database or a data warehouse. You don't need a dedicated data team.

You need consistent formatting: if you store phone numbers, store them the same way every time. If you track programme outcomes, use the same categories across years. AI can work with variation, but consistent data produces much better results.

You need reasonable completeness. Not every record needs every field filled in. But if 40% of your donor records are missing email addresses, AI can't help you with email segmentation. Know where your gaps are and prioritise filling the ones that affect your planned AI use cases.

You need unique records. Duplicates confuse AI analysis. If the same donor appears five times with slightly different name spellings, your donor analysis will be wrong in ways that are hard to spot. Deduplication is boring and incredibly valuable.

And you need your data in a format AI can read. Spreadsheets, CSVs, and structured databases work well. Data locked in PDFs, scanned documents, or legacy systems that only export in proprietary formats is harder to work with. Not impossible, but harder.

Common charity data problems

We've seen these in nearly every charity we've worked with.

The colour-coded spreadsheet where categories are indicated by cell colour rather than a data field. Only one person knows what the colours mean. AI can't read cell colours.

The merged-cells masterpiece. A beautiful-looking spreadsheet that is functionally useless for any kind of automated analysis because merged cells break every data processing tool.

The CRM nobody trusts. Years of inconsistent data entry by different staff members, with no validation rules and no data standards. The CRM has information in it, but nobody is confident that information is correct.

Data in email. Programme updates, beneficiary information, and outcome data shared via email rather than entered into a system. This data exists but it's not accessible to AI or anyone else.

And the legacy system: an old database, sometimes decades old, that contains valuable historical data but can't easily export it in modern formats.

Incremental improvement, not big-bang projects

The worst approach to data quality is a six-month "data transformation project" that tries to fix everything at once. These projects are expensive, disruptive, and have a poor track record of success.

The better approach: fix your data incrementally, starting with the data that supports your most valuable AI use case.

Our data readiness assessment recipe scores your data across six dimensions and identifies blockers before you start. It takes a few hours and costs nothing. Beyond that, the principle is simple: fix your data incrementally, starting with the data that supports your most valuable AI use case.

If you want to use AI for donor analysis, start by cleaning your donor data. Deduplicate records. Standardise how you record giving amounts. Fill in missing email addresses where you can. This might take a few days of focused work, not months.

If you want to use AI for programme evaluation, start by standardising how you record outcomes. Choose a consistent set of categories. Apply them to new data going forward, and backfill historical data when you have time.

Each incremental improvement makes your data more useful for AI and for everything else. You don't need to fix everything. You need to fix the bit that matters most right now.

AI can help with data quality

This is a useful irony: AI is good at helping you clean the data that makes AI work better. Upload your CRM export and ask AI to identify likely duplicates based on name, email, and address similarity. It can take a messy spreadsheet and propose standardised formats. It can scan your dataset and report which fields have low completion rates, helping you prioritise. Our AI Recipes include step-by-step guides for assessing your data readiness, detecting duplicates, and cleaning and standardising contact data.

This week, open your CRM or main data system and export your most-used dataset. Look at it honestly. How consistent is the formatting? How many duplicates do you spot in the first 100 records? How many fields are empty? That's your data quality baseline. This month, pick one data quality issue to fix: the one that would make the biggest difference to a task you're trying to do with AI. This quarter, implement one data standard for new data going forward. Write down the rules and share them with everyone who enters data. This prevents future mess while you clean up the past.

Data quality isn't exciting. Nobody puts "cleaned 2,000 CRM records" on their annual review. But it's the work that makes everything else possible. If you're writing an AI strategy, put data quality on page one.