Without reliable data, AI is useless
Several studies suggest that around 80% of corporate AI projects do not meet their objectives — and in the majority of cases, the cause is not the algorithm but the data. Duplicates in the CRM, empty fields in the ERP, inconsistent formats across spreadsheets: these problems, often invisible day to day, become fatal when you plug an AI model into them. The good news: cleaning and structuring your data is a project within reach of any SMB, with a ROI that goes well beyond the AI scope alone.
The problem: unusable data that sabotages AI projects
In a typical French SMB, data are scattered across 5 to 15 different tools: CRM, ERP, Excel spreadsheets, mailboxes, project management tools, shared drives. Each tool has its own entry conventions, its own formats and its own gaps. The result is a patchwork of inconsistent data that neither a human nor a machine can use effectively.
The most frequent problems we observe with our clients:
Duplicates and inconsistencies
The same customer exists 3 times in the CRM under slightly different names ("Dupont SA", "DUPONT S.A.", "Ets Dupont"). Addresses are not normalized. Phone numbers mix formats. An AI customer scoring model that trains on these data produces absurd predictions.
Empty fields and missing data
40% of contact records have no business sector filled in. 60% of sales opportunities have no expected closing date. Without this information, it is impossible to segment, forecast or automate anything. AI needs complete data to function.
Obsolete data
Contacts who left the company 2 years ago, 2022 prices still in the catalog, outdated internal procedures in the document base. An AI chatbot that relies on these data gives wrong answers — and loses the trust of users within a week.
The cost of poor data quality is estimated at 15 to 25% of revenue for companies that have not put governance in place. This cost existed before AI, but AI amplifies it: a bad spreadsheet produces occasional errors, an AI model trained on bad data produces systematic errors at scale.
The quality framework in 5 dimensions
To assess and improve the quality of your data, we use a framework in 5 dimensions. Each dimension corresponds to a specific type of problem and to concrete corrective actions.
Completeness — is the data there?
Measure the rate of filled-in fields for each critical entity (customers, products, transactions). A minimum threshold of 90% completeness is necessary for most AI models. Below that, the model compensates for the gaps with assumptions — often wrong. Action: identify the 10 most important fields per entity and make them mandatory in your tools.
Accuracy — is the data correct?
Check that the values match reality: valid email addresses, consistent amounts, dates in the right format. Accuracy is measured by sampling: take 100 records at random and check them manually. An error rate above 5% justifies a systematic cleanup. Tools: validation scripts, address verification APIs (Google Geocoding, SIRENE for French companies).
Consistency — does the data agree across systems?
A customer's revenue in the CRM must match the revenue in the ERP. The address in the customer base must be the same as the one on the invoices. Cross-system inconsistencies are the number 1 source of teams' distrust of AI tools. Action: create a single repository (master data) for the entities shared across systems.
Uniqueness — no duplicates
Duplicates distort all analyses: a customer counted 3 times artificially triples the number of customers, biases the segments and corrupts the predictive models. Automatic deduplication (fuzzy matching on name + address + SIRET) eliminates 80% of duplicates. The remaining 20% require human validation.
Freshness — is the data up to date?
Define an expiry date by type of data: customer contact details (6 months), product prices (1 month), internal procedures (12 months). Set up automatic alerts when a data point has not been updated for too long. A CRM where 30% of contacts have not been verified for 2 years is a dangerous CRM for an AI project.
Implementation: cleaning your data in 6 weeks
Data cleaning does not need to be a monumental project. Here is a pragmatic approach in three phases, designed for SMBs that do not have a dedicated data team.
Phase 1: express audit (week 1-2)
Export your 3 main data sources. For each, measure the 5 dimensions of the framework: completeness, accuracy, consistency, uniqueness, freshness. Use a tool like Great Expectations (open source) or Soda Core to automate the checks. Produce a "quality score" per source and identify the 3 most critical problems.
Phase 2: targeted cleanup (week 3-5)
Focus on the data that will feed your first AI project. Deduplicate the CRM contacts (tool: Dedupe.io or a Python script with the fuzzywuzzy library). Fill in the critical missing fields. Normalize the formats (ISO dates, structured addresses, amounts in cents). Do not try to clean everything: target the 20% of data that serves 80% of the use cases.
Phase 3: prevention and governance (week 6+)
One-off cleaning is not enough: without entry rules, data degrades within a few months. Set up validations at entry in your tools (mandatory fields, enforced formats, drop-down lists). Define a "data owner" per business domain. Schedule an automated quarterly quality check. This lightweight governance prevents 90% of future quality problems.
Metrics and tools
To steer quality over time, track these key indicators and use the right tools.
Recommended tools
Automated validation (open source)
Python framework to define and run "expectations" on your data: checking types, value ranges, completeness, uniqueness. Integrates into your existing data pipelines. Free and widely adopted in production.
Continuous quality monitoring
A monitoring tool that automatically checks the quality of your data at each update. Real-time alerts in the event of degradation. Cloud version available for teams without data infrastructure. Ideal for SMBs that want to automate without coding.
Smart deduplication
Deduplication service that uses machine learning to identify duplicates even with spelling variations. Visual interface to validate the matches. Particularly effective on contact and customer bases.
Frequently asked questions
Do you need perfect data to use AI?
No. The goal is not perfection but a level of quality sufficient for the targeted use case. An internal chatbot tolerates 5 to 10% of inaccuracies in the document base. A cash flow forecasting model requires less than 1% of errors in the financial data. Define an acceptable quality threshold per use case, not an absolute standard.
How can we assess the quality of our data in one day?
Export your 3 main data sources (CRM, ERP, spreadsheets) and check: the rate of empty fields (completeness), the number of duplicates (uniqueness), the consistency of formats (dates, addresses, amounts). A 50-line Python script or a tool like Great Expectations is enough to obtain a quality score in a few hours.
Who should be responsible for data quality in an SMB?
Ideally, a "data owner" per business domain: the sales director for CRM data, the CFO for financial data, the HR director for HR data. In an SMB, this role is often combined with the business function. The essential thing is that each data source has an identified owner who validates the quality rules.