What data does artificial intelligence need?

An AI agent that answers customers with outdated information, a forecasting system that ignores cancelled orders, a dashboard that combines different product codes for the same item: these are not artificial intelligence problems. They are data problems. Understanding what data artificial intelligence needs is the first step in turning AI from an interesting demo into an operational lever with a measurable impact.

For an SME, the most useful information assets are rarely hidden in a single perfect database. They are usually spread across ERP systems, CRMs, e-commerce platforms, spreadsheets, email inboxes, PDF documents and customer support conversations. Value does not come from indiscriminately accumulating everything, but from selecting, connecting and making usable the information needed to make a decision or automate a specific task.

What data does artificial intelligence need in a business?

The right answer is: it depends on the objective. An internal assistant for the sales department does not need the same data as an algorithm that estimates future demand or a system that automatically classifies support tickets. Before choosing the technology, define the process to improve, the decision to speed up and the expected outcome.

In general, an AI project uses four categories of information. Structured data, such as customer records, orders, prices, margins, dates and workflow statuses, are ideal for analytics, forecasting and automations that combine rules with statistical models. Unstructured text data, such as emails, quotes, contracts, manuals and CRM notes, power assistants that can search, summarise and answer in natural language.

Contextual data are also needed: up-to-date catalogues, price lists, availability, commercial policies, internal procedures and operational constraints. Without context, even a highly advanced language model can produce a plausible but unusable answer. Finally, outcome data are essential: a sale won or lost, a ticket resolved or reopened, a delivery on time or late, a quote accepted or rejected. These make it possible to tell whether automation is genuinely improving the process.

Large volumes are not always necessary. To classify incoming requests or extract fields from repetitive documents, a few hundred consistent examples can be more effective than thousands of disorganised records. By contrast, a predictive model for seasonality, churn or demand needs enough historical data to detect variations, exceptions and recurring behaviours.

Operational data: the most practical starting point

The most valuable data are often already generated by day-to-day work. A sales business can start with contacts, deals, products, orders, reasons for lost sales and sales activities. A manufacturer can use cycle times, production orders, scrap, machine downtime, suppliers and delivery delays. A professional services firm can make use of case files, deadlines, documentation and communication histories.

The criterion is not quantity, but relevance to the process. If the goal is to reduce the time needed to prepare a quote, you need a catalogue, configurations, price lists, discount rules, availability and previous quotes. If the goal is to improve customer care, you need historical conversations, a knowledge base, order status, return conditions and the outcomes of handled requests.

A common mistake is to start by uploading entire document folders into AI without a taxonomy, confirmed versions or any distinction between official and obsolete sources. The result is an assistant that finds lots of information but cannot determine which information is valid.

Quality matters more than volume

Useful data must be accurate, up to date, complete and consistent with other systems. If the CRM contains five variations of the same customer, product codes do not match between e-commerce and the ERP, or dates are recorded using different conventions, AI amplifies confusion rather than reducing it.

Quality must also be assessed in terms of representativeness. An archive of customer requests made up almost entirely of simple cases will not prepare an assistant to handle exceptions, complaints or technical questions. A sales history that excludes returns and cancellations can produce overly optimistic forecasts. Every dataset tells only part of the story: the project team's task is to understand what is missing and whether that absence could distort a decision.

In many cases, it is useful to introduce a normalisation layer before using the data. This means standardising formats, deduplicating records, defining mandatory fields, aligning process statuses and establishing a primary source for every critical piece of information. This is not an ancillary task: it is what makes the system reliable over time.

Labelled data and human feedback

Some use cases require examples with an outcome already identified. For instance, to automatically recognise the category of a request, emails or tickets must be assigned to the correct category. To estimate churn risk, you need to know which customers have actually reduced or stopped purchasing.

Labelling must not turn into an endless project. It can start with historical data that is already available and improve gradually through human review. An operator who corrects a misclassification or assesses a suggested response provides useful feedback, as long as it is recorded in a structured way. This creates a practical improvement loop: the system supports the team, and the team makes the system more accurate.

Access and integration: data must be available when needed

Having valid data is not enough if it remains isolated. A support AI agent must be able to check order status before responding. A lead-scoring system must receive updates from the website, CRM and campaigns. A forecasting dashboard loses value if it is updated manually at the end of each month.

That is why the design must include integrations between existing systems, APIs, synchronisation rules and shared identifiers. The point is not to centralise everything at any cost. In some scenarios, it is safer and more efficient to keep the data in its source ERP and let AI consult it only when needed. In others, a data warehouse or a centralised database improves analytics and reporting.

The choice depends on update frequency, the number of sources, security requirements and operational criticality. A quote may be fine with data updated every hour; online stock availability may require almost immediate synchronisation.

Data privacy, security and governance

Business AI is not a shortcut for pouring confidential information into public tools. Personal, commercial and strategic data require a design that complies with GDPR, access roles and internal policies. You need to know who can access what, for what purpose, for how long and with which audit logs.

The principle to apply is data minimisation: use only the information necessary for the use case. Analysing on-time delivery rates does not require exposing customers' personal data. When creating an HR assistant, it is often preferable to separate or pseudonymise identifying information if it is not essential to the answer.

Source governance must also be defined. Who updates procedures? Which document takes precedence in the event of a conflict? When a policy changes, how soon must the assistant stop using the previous version? Without clear accountability, the problem is organisational, not technical.

From source to outcome: a method for getting started

An effective project starts with a narrow scope and a verifiable KPI. Not with the generic question "how do we use AI?", but with a priority such as reducing the time spent handling standard requests by 30%, lowering order-entry errors or increasing the speed of lead qualification.

The process can follow five essential steps:

1. Map the current process, identifying manual tasks, waiting times, errors and handoffs between tools.
2. Identify the required data, available sources and gaps that prevent a reliable decision.
3. Clean and connect the minimum useful information, without embarking on an endless clean-up of the entire data estate.
4. Run a pilot on a controlled workflow, with human oversight and escalation rules for uncertain cases.
5. Measure time saved, error rate, output quality, team adoption and financial impact.

This is where a tailored solution makes the difference. A generic model can produce convincing text; a system integrated with business processes, roles and sources can reduce manual work and make operations more predictable. Graffico designs this step by connecting automations, software and data around measurable business objectives.

Artificial intelligence does not require perfect data to get started, but it does require data to be governed well enough not to carry existing errors into a faster process. The best place to start is often the process that currently takes the most time, generates the most checks or leaves the most information scattered. If that data can lead to a faster, more controlled decision, AI has already found its place.