Data ready for AI: how to prepare your company

A company is not ready for AI because it has purchased licenses or chosen a model. It is ready when its data is reliable, accessible, governed, traceable, and useful for a measurable business case. Failing to prepare multiplies rework, errors, regulatory risk, and technological expense with no impact on EBITDA.

Taxonomic definitions

1. Data preparation for AI or AI Data Readiness

It is the ability of an organization to deliver relevant, reliable, accessible, protected, updated, and traceable data to models, agents, and artificial intelligence applications.

It is not limited to “clearing information”. It includes:

Quality and consistency.
Integration between systems.
Metadata and lineage.
Security and permits.
Operational availability.
Representativeness.
Data evaluation against a business case.
Continuous monitoring.

2. Data product or Data Product

It is a data set designed, documented, and managed to solve a specific business need. It has a owner, defined users, quality controls, access rules, service level agreements, and performance metrics.

For example, a data product to predict customer churn should not be limited to exporting information from the CRM. It should integrate purchase behavior, interactions, complaints, consents, service history, and a corporate definition of what “churn” means.

3. Data lineage or Data Lineage

It is the record of where a data originated, what transformations it underwent, which systems consumed it, and which version reached the AI model.

The lineage allows us to answer critical questions:

From what system did this response come from?
What document did the agent use?
What version of the policy was in effect?
Who modified the registry?
Can the decision be reproduced?
Is there evidence for an audit?

Without lineage, an AI-generated response may seem correct, but it may not be defensible in the eyes of clients, auditors, or regulators.

Comparative table: data ready for AI vs. data not prepared

Variable	Data not prepared	Data ready for AI	Business impact
Quality	Incomplete, duplicate, or contradictory records	Quality rules, validation and correction from the source	Less errors, rework and incorrect decisions
Integration	Isolated information in ERP, CRM, Excel, email and legacy systems	Governed pipelines and connected sources	Less preparation time and automation of complete processes
Accessibility	IT department to extract information	Access via roles, APIs, and reusable data products	Lower time-to-market for new AI cases
Traceability	It cannot be explained how an answer came about.	Lineage, version, source and available evidence	Greater regulatory control and audit capacity
Unstructured data	Documents loaded without context or classification	Documented, tagged, and linked to their source	Better accuracy in assistants and RAG solutions
Security	Inherited permissions or excessive access	Identities, minimum privileges, classification and usage policies	Lower exposure of sensitive information
Update	Outdated or manually uploaded information	Frequency defined according to the business process	Timely decisions and lower operational risk
Property	“The data is from IT”	Defined business owners and quality managers	Accountability and sustainable improvement
Evaluation	AI is tested with isolated examples	Set of evidence, acceptance criteria, and human review	Lower probability of errors in production
Economy	Cloud and licensing costs without measurement	Consumption, benefits and risks associated with the business case	Greater control over ROI and impact on EBITDA

What does it really mean to have data ready for AI?

Having large volumes of information does not mean being prepared to use artificial intelligence.

An organization is ready when it can provide the right system:

The right data.
At the required time.
With the correct permits.
In a usable format.
With sufficient context.
With evidence of its origin.
With a quality compatible with the risk of the decision.
With mechanisms to detect deviations.

This distinction is relevant because the adoption of AI has advanced faster than its ability to generate business results. In 2026, McKinsey reported that only 7% of the companies had fully scaled up AI within the organization, identifying data preparation as one of the main bottlenecks.

The difference between a demonstration and an entrepreneurial capacity is not usually in the model itself. It lies in the infrastructure, the data, the controls, and the operational process that surrounds the model.

Why do AI projects get stuck with data?

Proof-of-concept tests typically use selected files, small samples, or manually prepared information. Under those conditions, the model can produce convincing results.

The problem appears when connecting it to the actual operation:

The customer names do not match across systems.
There are multiple versions of the same policy.
The documents have no date, category, or owner.
The product catalog contains duplicates.
Historical records do not preserve the context of the decision.
The permissions from the originating system are not transferred to the agent.
The information comes after the decision was supposed to be made.
There is no “correct answer” against which to evaluate the model.

IBM found that 42% of the organizations surveyed believed that they did not have enough proprietary data to customize their AI models.

The problem is not solely technical. When information is not prepared, the company faces:

More hours of cleaning and reconciliation.
Delays in going into production.
Unnecessary cloud consumption and processing.
Inconsistent results between areas.
Specialist team.
Decisions that require manual verification.
Risk of using personal data without authorization.
Loss of trust from users and customers.

McKinsey reported in its 2025 global study that 88% of the participants were already using AI in at least one function, but approximately a third had started to scale it. Only 39% reported impact on EBIT at the enterprise level. Moreover, 51% of the organizations using AI had experienced at least one negative consequence and nearly a third reported problems stemming from inaccurate results.

Executive reading is straightforward: adopting AI does not equate to capturing value. Without prepared data, the company can speed up the generation of responses, but it can also speed up errors.

How to prepare data for an artificial intelligence project

Step 1. Start with the business decision, not the model

Before integrating sources or selecting platforms, it is necessary to define what decision, process, or outcome is intended to be improved.

The use case must answer:

What economic problem is being solved?
Which area will be responsible for the result?
What decision will AI make or recommend?
What is the current cost of the process?
What error is acceptable?
Which decisions require human authorization?
What financial or operational KPI will demonstrate value?
Which systems will have to run the action?

An agent that answers internal questions, a model that detects fraud, and a solution that approves credit should not use the same quality, oversight, and risk thresholds.

The greater the impact of an incorrect response, the stricter the controls on the data must be.

Step 2. Identify the critical data domains

It is not necessary to correct all the information of the company before starting AI. It is necessary to prepare the domains that determine the outcome of the use case.

Examples:

Customer service

Customers and contracts.
Products and services.
History of interactions.
Knowledge base.
Care policies.
Complaints.
Levels of service.
Consent forms and personal data.

Finance

Accounting catalog.
Cost centers.
Invoices.
Purchase orders.
Suppliers.
Rules of authorization.
Budgets.
Documentary evidence.

Manufacturing

Assets and components.
Sensors.
Fault events.
Maintenance orders.
Shifts.
Lots.
Quality parameters.
Spare parts.

The inventory must include structured and unstructured data: tables, emails, images, audio, manuals, contracts, tickets, transcripts, and scanned documents.

Step 3. Measuring quality using objective criteria

Quality must be evaluated by dimension and by critical field. A general average may conceal serious errors.

The minimum dimensions are:

Completeness: percentage of mandatory fields available.
Accuracy: Correspondence with reality or authorized source.
Consistency: coincidence between systems and periods.
Uniqueness: absence of duplicate records.
Validity: compliance with the format and business rules.
Opportunity: availability within the required time.
Referential integrity: correct relationship between entities.
Representativeness: coverage of the scenarios that the model will face.

As the operational control proposed by Scanda, the identifying fields and the rules that enable a critical decision must meet a requirement higher than the average of the set:

Unique and valid critical identifiers.
Mandatory fields with explicit threshold.
Conciliation between master sources.
Documented tolerance for missing data.
Evidence of the origin of the records.
Alerts when a quality rule is violated.
Blocking the flow when the error exceeds the accepted risk.

Scanda addresses this stage with objective measurement, preventive controls, monitoring, and governance, rather than treating quality as an extraordinary cleanup before each project.

Step 4. Assign owners and governance rules

A data project fails when everyone can consume the information, but no one is responsible for its meaning or quality.

Each domain must have:

Data Owner: responsible for the use and definition of the data from the business perspective.
Data Steward: responsible for applying rules and resolving incidents.
Technological Custodian: responsible for storage, availability and backup.
Security Officer: validates access, classification and display.
Responsible for the AI case: is responsible for performance, adoption and results.

They must also be documented:

Authorized purpose of use.
Legal basis or consent.
Rules of retention.
Classification of information.
Access by role.
Use allowed by external models.
Requirements for anonymity.
The elimination process.
Third-party management.
Evidence for audit.

The NIST AI Risk Management Framework structures the management of AI risk around four functions: govern, map, measure, and manage. This reinforces that data control should not be performed after deployment, but throughout the entire lifecycle.

Step 5. Integrate the sources without creating another layer of chaos

An AI project can be connected directly to multiple systems, but each point-to-point connection increases complexity, maintenance costs, and the likelihood of inconsistencies.

Architecture must define:

Authorised registration systems.
ETL or ELT patterns.
Batch integrations, near real-time or in real time.
Rules of reconciliation.
Data schemes and contracts.
Handling errors and retries.
Pipeline monitoring.
Change history.
Consumption layer for AI applications.
Cost controls.

Depending on the maturity and the use case, the solution can rely on:

Data Warehouse.
Data Lake.
Lakehouse.
Master Data Management.
Business APIs.
Streaming of events.
Data products.
Corporate knowledge platforms.

Scanda integrates data from ERP, CRM, operating systems, SaaS platforms, and legacy environments through governed architectures, monitored pipelines, and layers prepared for analytics and AI.

Step 6. Prepare the unstructured data for generative AI

Loading thousands of files in a vector format does not guarantee correct answers.

A document must be turned into a usable asset. To do this, the following is required:

Extract text, tables, and images.
Preserve the relationship with the original file.
Identify title, date, version, area, and owner.
Classify sensitivity.
Eliminate duplicates and obsolete versions.
Segmenting content without losing context.
Generate metadata for filtering.
Associate document permissions with the user.
Define validity.
Maintain verifiable references.
Evaluate the recovery before evaluating the response.

In a RAG architecture, quality does not depend solely on the model. It also depends on whether the system retrieved the correct document, the correct fragment, and the current version.

McKinsey points out that scaling AI requires connecting structured and unstructured information in a governed, traceable, and reusable basis. A PDF, for example, can generate text, tables, images, metadata, tags, and quality scores that must remain linked to the original document.

Step 7. Build an evaluation set before production

A solution should not be approved because “the answers look good.”.

A representative assessment set is required with:

Frequently asked questions.
Exceptional cases.
Ambiguous situations.
Incomplete data.
Contradictory documents.
Unauthorized requests.
Cases that require escalation.
Examples of correct answer.
Examples of abstention.
Criteria for human review.

Metrics can include:

Precision.
Recall.
Rate of substantiated responses.
Correct recovery rate.
False positives.
False negatives.
Answers without evidence.
Violations of permits.
Response time.
Cost per transaction.
Percentage of human escalation.
Impact on the business KPI.

For high-risk cases, an incorrect answer should not be compensated for with a good overall average.

Step 8. Process the data as a permanent product

Data alone are not ready to be used. Sources, processes, rules, customers and market conditions change.

The operation must monitor:

Pipeline collapses.
Schematic changes.
Loss of completeness.
Increase in duplicates.
Outdated information.
Variations in distribution.
Loss of representativeness.
Changes in permissions.
New documentary versions.
Performance of the AI system.
Cost per consultation or process.
Incidents and corrections.

Data preparation is not a phase that ends before go-live. It is an operational capability.

Scanda Data Preparation Matrix for AI

Scanda proposes to evaluate each use case using six weighted dimensions. The goal is not to obtain a decorative rating, but to make an investment decision.

Dimension	Suggested weighting	Evidence required
Alignment with business value	15%	KPI, baseline, expected benefit, owner and decision affected
Quality and representativeness	20%	Profile, rules, exceptions, coverage and test results
Access and integration	15%	Authorized sources, pipelines, frequency, APIs, and availability
Government, privacy and security	20%	Owners, permits, purpose, classification and retention
Traceability and metadata	15%	Lineage, version, source, catalog and reproducible evidence
Operation and monitoring	15%	SLA, alerts, observability, incident response, and costs

Executive interpretation

Less than 60 points: Do not invest yet in construction or scaling up. Fix the critical gaps.
From 60 to 79 points: execute a controlled pilot with a formal remediation plan.
80 points or more: move on to production if no critical control points are not met.
Mandatory blocking: Although the rating is high, it should not be scaled up when there is no authorization for use, ownership, traceability, evaluation, or access control.

The matrix should be applied on a case-by-case basis. A company may be prepared for an internal consultant, but not for automating credit decisions or critical maintenance.

How to translate data quality into EBITDA, ROI and risk

The data preparation budget should be linked to avoided losses and captured benefits.

Annual cost of unprepared data

A financial approximation can be calculated as follows:

Cost of unedited data =

Hours of manual correction × cost per hour.
Operational procedures.
Billing, inventory, or service errors.
Delays in projects.
Wasteful infrastructure and cloud.
Incorrect decisions.
Expected loss due to incidents.
Fines, claims or compensations.
Margin not captured by late automation.

Expected loss due to risk

Expected annual loss = probability of the incident × estimated financial impact

The impact should include:

Operational interruption.
Recovery.
Research.
Notification.
Legal fees.
Compensation.
Fines.
Loss of customers.
Reputational damage.
Loss of productivity.

ROI of data preparation

ROI = [(operating savings + incremental margin + avoided losses) − total investment] ÷ total investment × 100

The total investment must include:

Architecture.
Integrations.
Licenses.
Remedy.
Government.
Security.
Evaluation.
Operation.
Adoption.
Maintenance.

The goal is not to clean everything up. It is first to invest in the data that protects or generates greater economic value.

Errors that should be avoided before implementing AI

Buying technology before defining the business case

Choosing a model without knowing the decision that needs to be improved produces eye-catching pilots, but difficult to justify financially.

Assuming that more data always leads to better results

Adding duplicate, outdated, or irrelevant information can degrade accuracy and increase processing costs.

Delegate all responsibility to the IT department

TI can manage the platform, but the definitions of client, onboarding, risk, product, or compliance belong to the business.

Clear the data once only

Without controls at the source, the errors reappear after each load.

Ignoring permissions from source systems

An agent should not reveal a document just because it could be indexed. The permissions must be applied during retrieval and generation.

Only evaluate average responses

The exceptional cases are those that pose the greatest financial, operational or regulatory risk.

Scaling without traceability

When an organization cannot explain what data supported a decision, it cannot defend it either.

Measuring adoption without measuring results

The number of users, queries or prompts does not equal EBITDA, productivity or risk reduction.

Predictive conclusion

During the next adoption cycles, the advantage will not lie in accessing the newest model, because that capability will become increasingly available. The difference will lie in which companies can connect models and agents with reliable, governed, and usable data within real processes.

The expansion of AI agents will increase the risk: an incorrect assistant generates a bad response; an incorrect agent may execute a bad action. The market is already showing a gap between usage and value: adoption is widespread, but scaling and the impact on EBIT remain concentrated in a minority.

Companies that delay data preparation will accumulate technical debt, manual controls, cloud consumption, and pilots that are difficult to scale. Those that turn their data into governed products will be able to automate entire processes, deploy AI with less risk, and measure their contribution to the business.

The executive question should no longer be “what model are we going to buy?”, but “what decisions are we willing to delegate and what evidence demonstrates that our data can support them?”.

FAQ

1. Is it necessary to modernize the entire data architecture before implementing AI? No. The priority should be to prepare the necessary domains for one or two business cases with measurable impact. However, the solution must be designed with reusable integration, security, lineage, and monitoring patterns. Creating isolated connections for each pilot reduces initial investment but increases the cost of scaling.

2. How can a CFO justify the investment in data preparation? You should compare the total cost of the preparation to three components: operational savings, incremental margin, and avoided losses. The financial case should include current rework, errors, delays, regulatory exposure, cloud cost, and expected benefits from automation. The budget is justified not by “having better data,” but by protecting EBITDA and enabling measurable results.

3. Can SharePoint documents, ERP, CRM or legacy systems be used in an AI solution? Yes, as long as there is authorization for use, controlled access, metadata, version identification, traceability, and update mechanisms. It should also be evaluated whether the AI retrieves the correct source and respects the user's permissions. Connecting a repository does not mean that its content is ready for use.

Before investing in AI, confirm that your data can support it

Scanda helps companies move from scattered information and isolated pilots to a reliable, governed, and ready database for intelligent automation.

We evaluate:

Feasibility of the use case.
Quality and representativeness of the data.
Integrations and architecture.
Government, privacy and security.
Traceability and documentation.
Financial and operational risk.
Ability to scale up.
Expected return.

The result is a prioritized roadmap that allows for deciding what to fix, what to integrate, and which AI cases can move forward without turning innovation into another source of technical debt.

Request an AI Value Discovery session with Scanda and determine if your data is ready before committing budget, infrastructure, and operations.