Preparing your data for AI starts with the use case. This guide explains how access, quality, freshness and permissions affect document workflows, knowledge retrieval and predictive models.
Where the received wisdom goes wrong
The standard advice is that data readiness precedes AI, and organisations act on it by commissioning a data programme that runs for two years and delivers a warehouse before anything useful reaches a user. Sometimes that is correct. Often it postpones all value for the sake of use cases that were never the priority.
The distinction is straightforward. Predictive work, forecasting, scoring, analytics, genuinely requires clean historical data, because it learns patterns from that history. Language-driven work, reading documents, drafting, classifying, answering questions, mostly does not, because the model reads the same messy input a person would.
Readiness by use case
| Use case | Data needed | Readiness required |
|---|---|---|
| Document processing | The documents themselves | Low |
| Support assistant (RAG) | Accurate current documentation | Medium |
| Drafting and summarising | Context from existing systems | Low |
| Classification and routing | Historical examples with labels | Medium |
| Demand forecasting | Clean history, 2+ years | High |
| Lead or churn scoring | Labelled outcomes at volume | High |
| Recommendations | Behavioural data at volume | High |
The top of that table is where most operational value sits for mid-market businesses, and it is reachable without a data warehouse programme.
What actually needs to be true
Even for low-readiness use cases, four things matter:
- Accessible. Reachable by API, database connection or an automated export. Manual spreadsheet exports may be a starting point, but a repeatable production workflow needs an appropriate ingestion process.
- Current. A knowledge base with a documented owner and update process. Stale content produces confidently wrong answers, which is worse than no system.
- Non-contradictory. Three versions of the same policy means the system picks one, effectively at random. Deduplication matters far more than schema perfection.
- Permissioned. You need to know who is allowed to see what, expressed in a way the system can enforce at retrieval time.
Notice that none of these require a warehouse, a lake, or a modelling exercise. They require ownership and hygiene.
A sequence that delivers value while you improve
-
Start with a use case that tolerates mess
Document processing or a drafting assistant. It delivers value in weeks and, importantly, reveals exactly which data problems actually matter to you.
-
Fix the specific problems it surfaces
You will learn that three fields matter and forty do not. That is a far better basis for a data project than a general quality initiative.
-
Build the shared foundation second
Once two or three automations want the same data, consolidate that data specifically. The requirement is now evidenced rather than theoretical.
-
Take on predictive work last
By this point you have clean, structured, current data for the domains that matter, and forecasting becomes feasible without a dedicated multi-year programme.
Frequently asked questions
Do we need a data warehouse before starting with AI?
How much historical data do predictive models need?
Our data is spread across ten systems. Is that a blocker?
What about unstructured data like emails and PDFs?
For implementation support, explore our AI data engineering services or discuss your workflow in a free consultation.
A 30-minute call. Bring one process that costs you real time and leave with an honest answer on whether automating it is worth the money.
