Data Foundations for AI
Every AI system is only as good as the data it can reach. For most organisations, the gap between an impressive demo and real value is not the model. It is data quality, integration, ownership and access.
AI learns from and works with your information, so if that information is messy, missing or locked away, the AI will get things wrong.
The short version, in a few minutes.
A plain-English walkthrough of this page. Prefer reading? Everything in the video is covered below.
What it is
Data foundations are the systems, processes and responsibilities that make your data accurate, connected, understood and available to the right people and tools. They include data quality, integration between systems, shared definitions, clear ownership, access control and the ability to trace where data came from.
Capable AI models are now widely available, and they keep improving. Your data is the part nobody else has. Two organisations using the same model can get very different results, depending on how complete, current and connected their data is. That makes data foundations a lasting source of advantage, and the most common reason AI projects stall.
How it works
In most organisations, the data that matters lives in operational systems: ERP for finance and stock, CRM for customers, WMS for the warehouse, EPOS for sales and EDI for orders and invoices exchanged with trading partners. Around them sits a long tail of spreadsheets, shared drives and email. Integration brings this together, through APIs, event streams or regular loads into a data platform such as a data warehouse, so AI can work from a consistent, current view.
Around that sit the practices that make data trustworthy: shared definitions, so “customer” or “on-time delivery” mean the same thing everywhere; quality checks where data is entered and as it moves; a single, managed record for each product, customer and supplier; named owners for each data set; and access rules that apply equally to people, reports and AI tools. Documents and emails need the same discipline, now that generative AI can use them.
Where it creates real value
- AI projects that reach production. Pilots stop stalling at the point where they meet real data.
- One version of the truth. Forecasts, reports and AI assistants give the same answer because they use the same definitions.
- Faster new use cases. Once data is connected and governed, each new AI project starts from a much stronger position.
- Lower risk. Knowing where personal and sensitive data lives, and who can reach it, makes privacy and security manageable.
- Freedom to switch. Well-organised data lets you change AI tools or vendors without starting again.
- Better operations anyway. Cleaner, connected data improves everyday reporting and decisions, even before any AI is added.
Where it's commonly misapplied
- Model first, data later. Choosing a model or platform before checking the data it needs leads to expensive pilots that cannot scale.
- Waiting for perfect data. No organisation has it, so improve the data that matters for the next use case rather than everything at once.
- A data lake with no owners. Copying everything into one place without definitions, quality checks or owners just moves the mess.
- Treating it as an IT project. Data quality is created by the people and processes that enter the data, so the business has to own it.
- Forgetting access. Connecting AI to data without matching permissions turns a data problem into a security incident.
The model is the easy part.
The architectural choices are about integration and trust. Decide how data moves from systems such as ERP, CRM, WMS and EDI into a place AI can use, and how fresh it needs to be: real time, hourly or daily. Prefer documented APIs and events over fragile manual exports. Keep lineage, so you can trace any answer back to its source, and build quality checks into the data flows so problems are caught before they reach a model or a decision.
Security and cost follow from the same design. Classify data by sensitivity, apply access control consistently across systems, reports and AI tools, and know where data is stored and processed, including by third-party AI services. Storage is cheap, but moving, copying and processing data at scale is not, so avoid copying everything into every new tool. Give each data set an owner who is responsible for its quality, and the time to do it.
What this means for your organisation.
Do not start a multi-year data programme before any AI work. Pick the first two or three use cases, identify exactly which data they need, and fix that data properly. Each project then leaves the foundations a little stronger for the next one.
Make ownership real. Every important data set needs a named business owner who agrees its definition, is accountable for its quality and decides who may use it. This is an organisational change more than a technical one, and it is usually the part that makes the biggest difference.
Questions to ask your team or vendor
- Which data does this AI use case need, and where does each piece live today?
- Who owns that data, and how good is its quality?
- Do our core systems agree on basic definitions such as customer, product and order?
- How will the AI get the data: through APIs, a data platform or manual exports?
- Will the AI respect the same access permissions as the source systems?
- Where will our data be stored and processed, including by AI vendors?
How it connects to the rest.
AI Strategy & Governance
Strategy decides which data to fix first, based on the use cases that matter most.
Explore →LLMs & Retrieval (RAG)
Retrieval is only as accurate as the documents and data it can search.
Explore →Predictive Analytics
Forecasts depend on complete, consistent history from your operational systems.
Explore →MLOps & Pipelines
Data pipelines and monitoring keep the data feeding your models fresh and correct.
Explore →I have worked with ERP, WMS, EDI, CRM, EPOS and data platforms across retail, healthcare, finance and logistics. I can help you assess your data readiness, or join your team to build the integrations AI depends on.
Do we need a data warehouse before we can use AI?
Not always. Some uses, such as searching documents, can start with what you already have. For AI that works across several systems, though, a governed central data platform usually becomes worthwhile quickly.
How clean does our data need to be?
Clean enough for the specific use. A forecast needs complete, consistent history. A document assistant needs current, authoritative documents. Assess the data for the use case you have in mind rather than trying to perfect everything.
Who should own data quality, IT or the business?
Both, with different roles. The business owns what the data means and how good it must be. IT owns the systems, data flows and controls that keep it that way.