--- title: "Is Your Data Ready for AI? The Checklist Before You Launch Your First Agent" description: "Is your data AI-ready? A practical checklist covering sources, definitions, quality, owners and access to review before you put your first AI agent to work." url: https://addabra.com/en/blog/ai-ready-data lang: en date: 2026-10-16 category: "Services" keyword: "AI-ready data" publisher: Addabra alternate: https://addabra.com/blog/datos-listos-para-ia --- # Is Your Data Ready for AI? The Checklist Before You Launch Your First Agent > Is your data AI-ready? A practical checklist covering sources, definitions, quality, owners and access to review before you put your first AI agent to work. It's 8 p.m. and the operations manager is still at the office. Tomorrow the pilot starts for the agent that will confirm deliveries with customers, and she has just discovered that many addresses in the system have no ZIP code, that some customers appear twice, and that order status is updated in a spreadsheet only one person maintains. The question she should have asked a month ago arrives late: **is our data AI-ready?** You don't have to end up in that evening. Data readiness can be checked beforehand, with a short and honest checklist, on the data your first agent will touch. Not on the entire company. ## Why data holds back so many agents An AI agent does three things with data: it reads it to understand the situation, decides something and sometimes writes back into your systems. If what it reads is incomplete or contradicts another source, the decision goes wrong even if the model is excellent. For years, people filled those gaps without noticing. They knew which column to ignore, whom to call to confirm a detail, which customer has two records. That invisible correction layer disappears when an agent does the work. That's why data problems that "were always there" surface in the very week of launch. The good news: almost all of them can be caught with simple questions. ## The AI-ready data checklist: seven questions before your first agent Take the AI use case you want to launch and answer each question only for the data that use case needs. If you've already done a [process diagnosis](https://addabra.com/en/blog/process-diagnosis-for-ai), start there: you already know which process and steps the agent will touch. ### 1. Do you know exactly what data the agent needs? Write the list: which fields it reads, from which system, and which fields it writes. If the answer is "the whole CRM", the use case isn't clear yet. A well-designed agent sees only what it needs, the same idea we explain in [context engineering](https://addabra.com/en/blog/context-engineering-for-companies). ### 2. Do you know where each piece of data lives? For each field, one official source. If order status exists in the ERP and in a tracking spreadsheet, decide which one wins. Spreadsheets aren't the enemy, but the ones that are the real source for a process need to be identified. We get into this in [operations without spreadsheets](https://addabra.com/en/blog/operations-without-spreadsheets). ### 3. Is there an agreed definition for each concept? Active customer, delivered order, margin, close date. If sales and finance define them differently, the agent inherits the confusion. One definition per concept, written down and approved by the teams. ### 4. Does the data have an owner? Each group of data (customers, products, pricing, orders) needs a person who decides what's correct and fixes what isn't. If there's no name, the error stays. ### 5. Is quality monitored automatically? Duplicates, empty fields, inconsistent formats. Cleaning once isn't enough: you need rules that run daily and alert the owner when something fails. If quality today depends on someone "checking now and then", it isn't monitored. ### 6. Is it clear who can see what? By role and also by agent. A collections agent doesn't need health data or salaries. Identify personal or sensitive data and define who has access, what gets masked and how long it's kept. Your legal or compliance team decides the policies; what matters is that they're enforced in your systems, not just written in a document. ### 7. Can you trace where a piece of data came from? If the agent says a customer owes an invoice, can you see which system that came from and when it was updated? Without that trail (lineage), every error turns into an argument. ## How to read the result This isn't a pass/fail exam. It's a map of where to start. | Result | What it means | What to do | |---|---|---| | Yes to everything | Your data can support the use case | Launch with evals and monitoring | | 1 or 2 questions fail | Specific gaps | Fix them in parallel with building the agent | | Questions 2, 3 or 4 fail | The foundation is missing | Sort out sources, definitions and owners before launch | | Question 6 fails | Access risk | Resolve it before any test with real data | > **Start small** You don't need perfect data across the company. You need reliable data for the first use case and a way to catch new errors. The rest gets sorted out one use case at a time. ## Illustrative example: the delivery confirmation agent *Illustrative example:* back to our operations manager. With the checklist two weeks earlier, the result would have looked like this: - **Data needed:** customer, phone, address, order, promised date and status. - **Source:** status lived in a spreadsheet; the team decided the ERP wins and the spreadsheet is fed from it. - **Definition:** "order ready for delivery" meant different things to the warehouse and to sales; they agreed on one. - **Owner:** the warehouse lead became responsible for order status. - **Quality:** a daily rule flags addresses with no ZIP code and alerts the owner. - **Access:** the agent reads only contact and order data, nothing from billing. The agent launched later than planned, but without the late-night panic and with a team that trusted what the agent said. ## Common mistakes - **Reviewing the whole company instead of the use case.** It becomes a never-ending project and the agent never ships. - **Cleaning once and forgetting about it.** Without rules that run on their own, data gets messy again. - **Assuming the model will "figure out" messy data.** The model reasons on what it receives; it doesn't guess which source is right. - **Leaving permissions for last.** An agent with too much access is a risk from day one. - **Not measuring.** If you don't know how many errors the data had before, you can't show it improved. ## What to do on Monday Bring together the process owner and the person who knows your systems best. Walk through the seven questions for a single AI use case, in one hour. Log every "I don't know" as a task with an owner. At Addabra, every data project starts this way: a readiness assessment for your AI use case with a prioritized gap map, and then we build the rules, catalog and permissions in your own systems. To see what it includes, take a look at our [Data Governance for AI service](https://addabra.com/data-governance-for-ai/en). ## Frequently asked questions ### How do I know if my data is ready for AI? Review the data your first use case will touch, not the whole company. If you know where it lives and it has an agreed definition, an owner, quality rules that run on their own and clear permissions, you are ready to start. If any of those fail, that is your first job. ### Do I have to clean all my data before launching an agent? No. The data that agent touches needs to be in order, and you need a way to catch and fix new errors. Waiting until everything is perfect is the surest way never to launch. ### What if my data lives in spreadsheets? You can work with them, but you should know which ones are the real source for a process, who maintains them and how they are validated. Often the first step is connecting them to a shared place with rules, without forcing the team to drop them overnight. ## About Addabra Addabra is an AI-software-first software factory that designs and builds custom agentic systems for companies: one agent per process, a shared business memory, orchestration, human approval by role, auditing and cost control. Free diagnosis: https://addabra.com/en/blog/ai-ready-data#contacto