Insights / Guide / Data foundations and advanced analytics
GuideData foundations and advanced analytics: the complete guide
AI and analytics are only as good as the data behind them. How to set ownership, fix master data, agree definitions and build dashboards and forecasts people trust.
Every organisation wants better reporting, sharper forecasts and, increasingly, AI that gives reliable answers. Most already have the data to support those ambitions. The problem is that it sits in many systems, is recorded inconsistently and is hard to trust. Finance and sales arrive at the monthly meeting with different revenue figures. Operations keeps its own spreadsheets because the system numbers look wrong. An AI assistant gives a confident answer drawn from an outdated document.
Data foundations are the unglamorous work that fixes this: clear ownership, clean master data, shared definitions and a governed place where data comes together. Advanced analytics, from role-based dashboards to forecasting models, is what you build on top. This guide covers both, in the order that tends to work, with practical checks at each stage.
Why data foundations come first
It is tempting to start with the visible part: a new dashboard tool, a forecasting project or an AI assistant. These projects often stall for the same reason. The underlying data cannot support them. Models grounded in poor data give poor answers, and dashboards built on disputed numbers are ignored. Fixing data first is usually the cheapest AI and analytics investment an organisation can make.
The root cause is rarely technical. It is that nobody owns the data, nobody agreed what the terms mean and nobody stops bad records being created. Those are management problems with management solutions, and they can be fixed in weeks rather than years if the work is focused on the data that matters most.
Start with the decisions
Do not try to fix all your data. List the ten decisions that matter most to your goals, identify the data each relies on and start there.
Set ownership and agree definitions
Begin with a short data strategy: which data matters most, who owns it and how it will be used. Each important data domain, such as customers, products, suppliers, accounts and employees, needs a named business owner. The owner is not the IT team. It is the senior person who is accountable for the data being complete, accurate and fit for use, and who decides when there is disagreement.
Definitions come next, and they matter more than any tool. If sales count revenue when an order is signed and finance counts it when an invoice is issued, both are right and the numbers will never match. Write down the definitions of the twenty or thirty terms that appear in management reports, agree them with the owners and apply them everywhere: in systems, reports and AI assistants.
- Agree definitions for revenue, margin, customer, active customer, order, stock and the other terms leaders use most.
- Record each definition, its owner and how it is calculated in a simple shared glossary.
- Apply the definitions in every report and data model, and retire reports that use old ones.
- Review definitions when the business changes, for example after an acquisition or a new pricing model.
Governance should enable, not block. Set access rules and quality standards that let people get to the data they need quickly, with sensible protection for personal and sensitive information. Privacy and data residency rules apply to analytics as much as to operational systems, so involve whoever handles data protection early.
Fix master data at the source
Master data is the shared reference data that every transaction relies on: customers, suppliers, items, accounts and locations. Duplicate or inconsistent master data is the root cause of most reporting disputes. When the same customer exists three times under slightly different names, sales by customer are wrong, credit limits are wrong and AI tools give inconsistent answers.
The fix is to correct master data in the system where it is created, not in reports downstream. Cleaning data in a spreadsheet each month treats the symptom and guarantees the work will be repeated.
- Name an owner for each master data domain, with authority over standards and approvals.
- Prevent bad entries: required fields, validation rules, duplicate checks and a simple approval step for new records.
- Clean the backlog: merge duplicates, fill gaps and archive inactive records, starting with those used most.
- Measure quality monthly, such as duplicate rates and incomplete records, and report it to the owners.
Consider a typical distribution business with several thousand customer accounts. A review found that around one in eight were duplicates, created by different branches using different naming conventions. Merging them and adding a duplicate check at account creation resolved a long-running dispute between sales and finance about customer profitability within one quarter.
Build the platform and put insight in front of people
With ownership, definitions and master data in hand, bring data from finance, commerce, point of sale, operations and other systems into one governed place that analysts and AI can use. That might be a modern data warehouse, a lakehouse or the analytics layer of your main business platform. The right choice depends on data volumes, existing systems and skills, but the principles are the same: automated loading, consistent definitions, controlled access and clear lineage from source to report.
Dashboards are where most people experience the work. Effective dashboards start with the decisions a role makes, not with the data available. A warehouse manager needs today’s picking backlog and late orders; a finance director needs cash, margin and forecast against budget. Show few measures, explain each one and make the numbers trustworthy by showing when data was last refreshed and where it came from.
- Design one dashboard per role or decision, not one dashboard for everyone.
- Limit each view to the measures that change what someone does.
- Show data freshness and link each measure to its definition.
- Add conversational reporting carefully, grounded in the same governed data and definitions.
Faster reporting often follows quickly. Once data loads automatically and definitions are agreed, month-end and management packs that took days to assemble can be produced in hours, freeing finance teams for analysis rather than reconciliation.
Predict and optimise
Forecasting and machine-learning models for demand, cost, cash, risk and churn can add real value once the foundations are in place. Before starting a predictive project, ask three questions. Is there enough history, usually at least two years for anything seasonal? Is the data consistent over that period, or did a system change or reorganisation break the series? And will anyone act on the result, with a clear process for using the forecast in planning or operations?
If the answer to any question is no, fix that first. A well-built model that nobody uses delivers nothing, and a model trained on inconsistent history will mislead. Where the answers are yes, start with one forecast, compare it honestly with the current method and put it into the daily workflow, such as the replenishment screen or the weekly cash meeting, rather than a separate report.
Data readiness for AI
- Each data source an assistant or model uses has a named owner.
- Key fields are complete and duplicates are below an agreed level.
- Access rules reflect who may see what, and personal data is handled lawfully.
- Documents used by assistants are current, approved and free of conflicting versions.
- Answers are tested against known questions before go-live and monitored afterwards.
What you will achieve
Done well, data foundations give finance, sales and operations one trusted set of numbers, so meetings focus on decisions rather than reconciliation. Reporting becomes faster and cheaper to produce. Forecasts for demand, cost and cash carry more confidence, and AI assistants give answers you can defend because they draw on governed, current data. Over time the cost to serve falls as manual workarounds and duplicate spreadsheets disappear.
A practical first step is a short data readiness review: list the decisions that matter most, trace the data behind them, name owners and score quality. Within a few weeks you will know where the biggest gaps are and which fixes will pay back first.
Source: Intelliversal analysis, drawing on the briefings in this series.
