No AI without clean data: the project SMEs keep postponing

In small and mid-sized companies, the same scenario keeps repeating. Leadership wants an assistant for customer service, a more reliable demand forecast, or help drafting proposals. A prototype is built, the demonstration is convincing, and then the project slows. Customer records exist in three versions, product references follow different logic from one site to the next, and nobody can say who is allowed to correct what.

In our experience, the model is rarely the limiting factor. The limiting factor is the job that has been postponed for years: the data. This guide offers a six-step method sized for organisations that have no data science team and no unlimited budget.

Why this work keeps being postponed

Three reasons come back often. The topic has no natural sponsor: it cuts across sales, operations, finance and IT, so it belongs to no one. The benefit is indirect: cleaning a file produces no visible result next quarter. And the scale is intimidating: people imagine a multi-year programme, when a well-chosen scope can be handled in a few weeks.

Yet AI does not repair faulty data, it amplifies it. A file full of duplicates will yield skewed analyses, delivered with confidence. This is a management risk before it is a technical one.

Step 1: start from a use case, not a grand clean-up

The first mistake is trying to "fix all the data". Pick a use case whose value is identifiable, for example cutting the handling time of customer requests or making an inventory budget more reliable. Then list the data strictly required: which fields, which sources, how much history. That scope sets the effort. Everything else waits.

Step 2: map the sources and name owners

Draw a simple map: where each piece of data lives (ERP, CRM, spreadsheets, mailboxes, shared folders), who enters it, who reads it, who can change it. A one-page table is enough.

For each dataset in scope, name a business owner. This is not an IT role: it is the person who can decide when two versions disagree. Without that name, no quality rule survives over time. This is often where the topic meets how work is organised, and we link it readily to our work in Operations & processes.

Step 3: measure quality on a sample

There is no need to audit millions of rows. Take a sample of a few hundred records and assess five criteria:

• Completeness: are the useful fields filled in?
• Uniqueness: how many duplicates are there?
• Consistency: are formats and codes the same across sources?
• Timeliness: when was the record last updated?
• Accuracy: does a manual check on a few cases confirm the value?

A worked example, with its assumptions

Take a fictional industrial company of 120 people with a file of 8,000 customers. Assume a sample of 200 records shows roughly one record in six duplicated, a third with no identified contact, and addresses entered under four conventions. These figures are illustrative. They are still enough to decide: before connecting a sales assistant to this file, it must be deduplicated and a format agreed. The effort is then counted in person-days, not months, because it is confined to one file.

Step 4: fix causes at the source

Correcting once without changing how data is entered is like emptying a bath without turning off the tap. For each measured defect, look for its origin: a free-text field where a closed list would work, double entry between two tools, no check when a record is created. A few simple rules (mandatory fields, value lists, duplicate search at creation) prevent most future degradation. The tool matters less than entry discipline and clear accountability.

Step 5: set the compliance frame from the start

Since 1 September 2023, the revised Swiss Federal Act on Data Protection (FADP, or nLPD in French) has applied. For an AI project, three practical points follow:

• keep a record of processing activities and know which personal data will feed the project;
• check purpose and proportionality: data collected to invoice a customer is not meant to feed just any model;
• control subcontractors and cross-border transfers, which bears directly on the use of AI tools hosted outside Switzerland.

Companies that also serve European customers remain subject to the GDPR, and the European AI Act, now moving toward formal adoption, is expected to add obligations phased according to the risk level of each use. Waiting for it to apply before getting organised would be unwise: a clean inventory of data serves all these frameworks at once. Our case on client data privacy compliance shows how such an inventory can be built without burdening teams. Following the texts themselves can also be lightened, as shown in automated regulatory monitoring, a case where AI genuinely helps because the scope is bounded.

Step 6: document, track, maintain

A data effort never ends. Three habits make it sustainable:

1. A short data dictionary: definition of each field, owner, reference source.
2. Two or three quality indicators, reviewed every quarter (duplicate rate, completeness rate, average age of updates).
3. A mandatory checkpoint: every new AI or analytics project starts by checking these indicators on its own scope.

This is the natural extension of the logic behind our Data, AI & compliance approach: strategy sets the objective, deployment installs the rules, steering confirms that they hold.

What executives can decide this week

• Appoint a sponsor at leadership level, with a cross-functional mandate.
• Choose a single use case and derive the data scope from it.
• Ask for a quality sample check before any tool purchase.
• Require that each dataset in scope has a named owner.
• Build the FADP review into project kick-off, not delivery.

This is not a prerequisite that delays AI. It is what allows AI to be useful, or to be set aside when a simpler lever fits the need better. Sometimes an entry rule and a dashboard are enough, and we say so.

If you want to know where your data stands against a specific use case, our diagnostic provides, within a few weeks, a clear baseline, a realistic scope and a costed roadmap.