Preparing data for AI is not about “cleaning up every system,” but about collecting and verifying exactly the data needed for one chosen use case: for a knowledge base, that means current documents with access permissions; for forecasting, it means event history at the required level of detail; for an AI agent, it means reference data and system access; for computer vision, it means labeled images. For each dataset, you define an owner, a source, and check completeness, timeliness, and consistency, then record it in a data sheet. This approach makes it possible to launch a pilot in weeks, instead of waiting for a years-long “data lake” project to finish.
Prepared by AIrassvet and verified against public sources October 1, 2026. It is intended for business owners and company executives, IT directors, and department heads planning their first or next AI project. Standard references are based on information about their texts as of the verification date; this is not legal advice on personal data processing. The AIrassvet section at the end describes the publisher’s services.
Contents
- Why AI Projects Run Into Data Problems
- What Data Different AI Use Cases Need
- Data Quality: What to Check
- A Data Sheet for an AI Project
- Preparing Documents for a Knowledge Base
- Preparing Tabular Data for Forecasting
- Data for AI Agents and Automation
- Personal Data and Trade Secrets
- Step-by-Step Preparation Plan
- How to Tell Whether the Data Is Ready
- Frequently Asked Questions
- Source Documents and Verification Method
Why AI Projects Run Into Data Problems
Models are now accessible: you can call them via API or deploy them on your own servers. But results are determined not only by the model, but by what you feed it. Typical situations that stop projects include:
- Documents are disorganized. There are multiple versions of policies and procedures, no clear current one, and some knowledge exists only in email threads.
- Reference data doesn’t match. The same customer or product is recorded differently in CRM, 1C, and spreadsheets.
- The history is incomplete. A forecast needs daily sales by store for several years, but only monthly totals were saved.
- There is no owner. No one is responsible for making sure the data is correct and updated.
- There is no access. The data exists, but it can only be obtained through a manual export.
It’s important to understand that none of these problems need to be solved “for the whole company” before you start. It is enough to solve them for the data needed for the first use case. If the use case has not yet been chosen, start with a business process audit before implementing AI.
What Data Different AI Use Cases Need
| Use Case | What Data Is Needed | Key Requirements | Common Problem |
|---|---|---|---|
| Knowledge base and employee assistant (RAG) | Policies, procedures, contracts, FAQs, support correspondence | Current version, document owner, access rights, readable text | Outdated and duplicate versions |
| Demand, churn, and pricing forecast (ML) | History of sales, inventory, prices, promotions, customers | Required granularity, history length, unified reference data, event markers | Missing data and a change in accounting method partway through the history |
| AI agent and process automation | Customer and product reference data, statuses, rules, access to CRM and 1C | Aligned reference data, API or a stable access method, documented rules | Process rules are not documented |
| Computer vision | Photos, video, document scans | Labels, variety of shooting conditions, examples of defects and rare cases | Too few examples of rare defects |
| Analysis of inquiries and reviews | Text of inquiries, calls, and reviews with metadata | Connection to the customer and the outcome, anonymization | No link between the text and the final outcome of the inquiry |
The table shows the main point: you cannot “prepare data for AI” in general; you can only prepare data for a specific use case.
Data Quality: What to Check
To assess data quality in analytics and machine learning in Russia, there is a national standard series, GOST R 71484-2024, based on the international ISO/IEC 5259 series. GOST R 71484.1-2024 provides an overview and terminology and describes a six-stage data lifecycle: requirements definition, planning, acquisition, preparation, provision, and decommissioning. GOST R 71484.2-2024 defines data quality characteristics and indicators. Both standards were approved by Rosstandart orders dated October 28, 2024, and came into effect on January 1, 2025; the series also includes parts on requirements for data quality management and the structure of that process.
You don’t have to apply the standards in full in a small project, but their list of characteristics is a good checklist. For the first project, the following are usually the most important:
| Characteristic | Verification Question | Example Problem |
|---|---|---|
| Completeness | Are the fields needed for the use case filled in? | Half of all orders are missing the acquisition channel |
| Consistency | Do the data match across different systems? | Different codes for the same product in 1C and on the website |
| Timeliness | How fresh is the data and how often is it updated? | The price in the knowledge base is from last year |
| Reliability | Can the source be trusted? | Figures from a manual spreadsheet with no author |
| Traceability | Is it clear where the record came from and who changed it? | The assistant’s answer cannot be verified against the source |
| Representativeness | Do the data reflect real-world conditions? | Training photos were taken only in daylight |
| Balance | Are there enough examples of rare classes? | Five defect examples per ten thousand normal ones |
| Accessibility | Can the data be obtained without manual work? | Export only through an employee once a month |
The characteristic names are based on the list in GOST R 71484.2-2024; the questions and examples are an editorial adaptation for business use.
Data Passport for an AI Project
Data Passport is a short document for each dataset in the project. It replaces long debates about whether the data are ready with clear answers. Minimum contents:
- Purpose: for what scenario and what model task the dataset is used.
- Source and collection method: system, spreadsheet, API, export frequency.
- Owner: who is responsible for accuracy and updates.
- Contents and volume: fields, period, number of records or documents.
- Quality: results of checks against the checklist characteristics, known issues.
- Restrictions: personal data, trade secrets, who is allowed access.
- Version and date: when the dataset was assembled and when it will be updated.
The passport is useful after launch too: when the AI starts making mistakes, the first thing to check is whether the data have changed.
Preparing documents for a knowledge base
For RAG-based assistants, data are documents. Preparation includes:
- Inventorying. Where the documents are stored and which ones are needed for the scenario.
- Selecting current versions. Outdated versions are excluded or marked; each document has an owner and a date.
- Access rights. The assistant should show employees only what they are allowed to see. Permissions are carried over from the source systems, not configured again manually.
- Text extraction. Scans are OCR-processed, and tables and diagrams are checked manually — that is where meaning is most often lost.
- Control questions. 50–100 real employee questions with correct answers and document links are the benchmark for quality checks.
More on how such systems work in the articles AI-Powered Enterprise Knowledge Base and RAG Accuracy and Hallucinations.
Preparing tabular data for forecasting
Forecasting demand, churn, or inventory optimization requires historical tables. Preparation checklist:
- Granularity: data at the level at which the decision will be made, for example by day, location, and product.
- History length: for seasonal processes, the history should cover several seasons; if it does not, this limitation must be recorded in the passport.
- Reference data: consistent product, customer, and location codes across the entire period; renamings and mergers are consolidated.
- Events: promotions, stockouts, price changes, and location closures are marked; otherwise, the model may treat a shortage as a drop in demand.
- Accounting gaps: a change in the accounting system or methodology in the middle of the history is described and taken into account.
Read more about using such data in the article Demand Forecasting with AI.
Data for AI Agents and Automation
An AI agent works with live systems: CRM, 1C, email, messengers. What matters to it is not historical datasets, but:
- Aligned reference data. The agent must be able to identify the customer, product, and contract unambiguously.
- System access. An API or a stable way to work through the interface, with an account that has the minimum necessary permissions.
- Written rules. How requests are handled, who approves exceptions, and what statuses exist. What “everyone knows,” the agent does not know.
- Action log. Where the agent’s actions are recorded so they can be reviewed and rolled back.
Documenting process rules is often harder than exporting data, but without it the agent will behave unpredictably. The approach is described in the article Technical Specification for an AI Agent.
Personal Data and Trade Secrets
Client and employee data are personal data, and their processing is governed by Federal Law No. 152-FZ “On Personal Data”. At the preparation stage, you need to decide:
- which fields are truly needed for the scenario, and which can be excluded or anonymized;
- where the model will run: in an external service or within the company’s own environment;
- who will have access to the prepared datasets and to the system’s responses.
For trade secrets, the logic is the same: the more sensitive the data, the stronger the case for a model on your own servers. Read more in the article Can Customer Data Be Sent to AI?.
Step-by-Step Preparation Plan
Step 1. Choose One Scenario
Describe the task, the solution it supports, and the pilot success criterion. Without a scenario, it is impossible to determine what data is needed.
Step 2. Build a Data Map for the Scenario
List the required datasets, their sources, and owners. Note what exists, what is missing, and what has to be collected from scratch.
Step 3. Check Data Quality
Go through the characteristics checklist for each dataset. Measure simple indicators: the share of empty fields, number of duplicates, reference data mismatches, and the last update date.
Step 4. Fix What Blocks the Pilot
Not everything at once: only the issues without which the scenario will not work. Record the rest in the data profile as known limitations.
Step 5. Prepare a Benchmark for Testing
Control questions with answers, a holdout history period for forecast validation, labeled examples for computer vision. A benchmark is needed to assess AI quality fairly.
Step 6. Set Up Updates
Define how the data will be updated after launch and who is responsible. A one-time export is fine for a pilot, but not for a live system.
How to Tell Whether the Data Is Ready
| Criterion | Readiness Indicator |
|---|---|
| Scenario | The task and pilot success criterion are documented |
| Composition | All scenario datasets are listed, and each has an owner |
| Quality | The checklist review has been completed; blocking issues have been fixed and the rest documented |
| Access | Data is obtained without manual work or with a clear export procedure |
| Limitations | It has been decided which personal and confidential data are used and where the model runs |
| Benchmark | There is a dataset for evaluating result quality |
| Updating | A responsible person and update method after launch have been assigned |
If even one criterion is not met, that is not a reason to cancel the project, but a list of tasks for the coming weeks.
Frequently Asked Questions
Do you need to build a data warehouse before implementing AI?
Not necessarily. For the first project, it is enough to prepare the data for one scenario: the required datasets, their owners, quality checks, and update method. A data warehouse makes sense when there are many scenarios and they use shared data.
How much data do you need for an AI project?
It depends on the scenario. A knowledge base needs up-to-date documents on the topic, forecasting needs historical data at the required level of detail for several seasons, and computer vision needs examples of each class, including rare ones. There is no universal threshold, so the amount is validated on a benchmark set during the pilot.
What standards define data quality for AI?
In Russia, the GOST R 71484-2024 series “Artificial Intelligence. Data Quality for Analytics and Machine Learning,” prepared on the basis of ISO/IEC 5259, is in force. Parts 1 and 2 were introduced on January 1, 2025, and describe terminology, the data life cycle, characteristics, and quality metrics.
Can customer data be used for AI?
Yes, provided the requirements for processing personal data under 152-FZ are met. During preparation, unnecessary fields are removed, data is anonymized where possible, and the location where the model will run is chosen: in an external service or within the company’s own environment.
Who in the company should be responsible for data preparation?
Each dataset should have a business owner who understands the meaning of the data and is responsible for its accuracy, and a technical owner responsible for exports and updates. Without an owner, data quality quickly declines after launch.
What should you do if there is too little data or the quality is poor?
Start with a scenario that can be supported by the data you already have, for example a knowledge base built on existing documents. At the same time, set up collection of missing data for more complex scenarios and document the limitations in the data profile.
Primary Sources and Verification Method
Verified on October 1, 2026:
- GOST R 71484.1-2024 (ISO/IEC 5259-1:2024) — overview, terminology, data life cycle; Rosstandart Order No. 1537-st dated 10/28/2024, effective January 1, 2025.
- GOST R 71484.2-2024 (ISO/IEC 5259-2:2024) — data quality characteristics and metrics; Order No. 1551-st dated 10/28/2024, effective January 1, 2025.
- Federal Law No. 152-FZ “On Personal Data” — the framework for processing personal data.
The standards were reviewed based on publications in the open regulatory database. The scenario table, checklist, data profile, steps, and readiness criteria are an editorial analysis by AIrassvet.
How AIrassvet Helps Prepare Data for an AI Project
If an AI project is being delayed because of uncertainty about the data, it is useful to start with one scenario and validate only its data. On the AIrassvet website, which also uses the name “AI Studio by Vladimir Lomtev,” the following directions relevant to this task are described:
- Process audit and data preparation for the selected scenario: data map, quality check, data profile, and benchmark for testing.
- Preparing documents and building a knowledge base with RAG, taking versions and access rights into account.
- Preparing historical data for ML forecasting of demand, churn, and inventory.
- Deploying models on company servers if data cannot be transferred to external services.
The first step is to choose one process, describe its current metrics, data sources, constraints, and the criterion by which you will consider the pilot successful.
Bottom line: prepare data for the scenario, not in general
Preparing data for AI starts with choosing a scenario: only that determines which documents, spreadsheets, reference data, or images are needed and what quality is sufficient. The GOST R 71484-2024 series provides a useful list of quality characteristics, and the data profile turns the review into a concrete task list.
Choose one process, map the data, remove what blocks the pilot, prepare a benchmark, and assign owners. That way, the company gets a working AI project and, at the same time, brings order to the data that actually creates value.