5 things you should do now to get AI ready before the competition

The organizations winning with AI are not the ones with the best models. They are the ones who moved first.

Every major AI capability, such as retrieval-augmented generation, predictive analytics, and autonomous agents, depends on one thing before it can deliver value: data that is complete, current, and accessible. 

The model is the easy part because foundational models are available to everyone. What is not available to everyone is the breadth of data, the pipeline reliability, and the organizational discipline to centralize the right data before the use case demands it.

This post is not about technology selection. It is about strategic urgency. The 5 actions below are practical, sequenced, and measurable. They are also time-sensitive: every quarter spent in analysis rather than execution is a quarter your competition may spend building the data foundation that makes their AI initiatives work while yours stall.

[CTA_MODULE]

1. Define the real data problem

Most organizations have a rough sense that they have a lot of data sources and that getting the data together will take time. Few have actually done the arithmetic, but we’ve laid out the steps to do it so you can attach both a timeline and a risk estimate to the result.

Step 1: Start with a count of your applications. 

A mid-sized enterprise today uses approximately 106-130 SaaS applications on average, and that number rises sharply with company size — enterprises with more than 5,000 employees average 131 or more, and those above 10,000 employees average 447. Add your databases, file sources, proprietary internal systems, and operational platforms to that count. 

Step 2: Apply the math. 

Using a traditional ETL approach — building, testing, deploying, and stabilizing a pipeline for each source — a realistic estimate is 4–6 weeks per source for an experienced engineer, with significant ongoing maintenance cost thereafter. The enterprise data infrastructure benchmark report 2026 from Fivetran found that organizations using legacy approaches spend an average of $1,900 per pipeline per year on maintenance alone, with teams of 35–60 engineers spending 53% of their time on upkeep rather than new work.

Step 3: Write down your source count, estimated weeks per source using current method, and confidence level in that estimate. 

Then do the risk adjustment. If your estimate is 6 weeks per source with 50% confidence, the risk-adjusted expected timeline is double your headline number; the actual timeline could easily be twice what you wrote down. Now run the same calculation with Fivetran, assuming 1 week per source and 99% confidence. Multiply the 2 numbers together; your expected timeline drops to a fraction of the traditional approach, and the risk-adjusted figure stays close to the headline because the confidence level is so high.

Step 4: Quantify the gap and bring to stakeholders

Open a spreadsheet. Count your sources. Run both calculations. Then, show the result to your CEO because this is the difference between having an AI strategy and having the infrastructure to execute it.

2. Examine your competition

The numbers you calculated above only matter in context. In isolation, 8 years sounds like a long time. In competition, it depends entirely on what your competitor's 8 years looks like compared to your 90. 

Imagine the scenario with specificity. Your competitors have centralized data from their top 200 sources. Their AI models have training data that is 18 months fresher and 3x broader than yours. Their agents have access to current customer, operational, and market data that yours do not. They are deploying personalization, dynamic pricing, predictive maintenance, and intelligent routing at scale. Their data team is spending 80% of its time on insight and innovation rather than pipeline maintenance. 

Now ask 4 questions: 

• What does that mean for the relative profitability of your businesses? 
• What does it mean for revenue growth rates? 
• What does it mean for your respective reputations with customers and talent? 
• And what does it mean for your market capitalization if investors begin pricing in their AI-native capability as a forward earnings multiplier?

The gap across these four dimensions — the delta between the business you are running today and the business your data-ready competitor is becoming — is not hypothetical. AI-driven competitive advantages compound, and the organizations that start now build a lead that becomes structurally difficult to close.

3. Build for agents now

Agentic AI is not a future technology. It is being deployed now in customer service, sales, operations, clinical decision support, and financial services. The data requirement for agentic systems is fundamentally different from the data requirement for batch analytics: agents need current state, not historical summaries. They need data that reflects what happened in the source system minutes or seconds ago, not what was loaded last night.

With that in mind, every pipeline you build today should be built with change data capture (CDC) as the default. CDC reads committed changes directly from the transaction log of source databases, propagating each insert, update, and delete to the destination as it happens. For SaaS sources, Fivetran's 1-minute sync intervals on Enterprise plans provide the closest available approximation. 

The result is a destination that agents can query and trust to reflect current reality rather than a snapshot that may be hours out of date. Equally important: the destination must be consistent across every platform where agents will run. A CDC pipeline that feeds Snowflake but not Databricks, or that works in production but not in the ML environment, creates exactly the schema inconsistency that breaks agentic workflows at deployment time.

4. Ensure any process, any time

The history of enterprise data infrastructure is a history of technology transitions — on-premises data warehouses to Hadoop to cloud warehouses to lakehouses to LLMs and beyond. In each transition, organizations that had stored data in the proprietary format of the platform they were committed to faced expensive migrations. Their data was not wrong; it was locked. The organizations that did best were those that stored data in open, portable formats and swapped the compute layer rather than reprocessing the data.

[CTA_MODULE]

The current equivalent is Apache Iceberg™, which can be queried simultaneously by Snowflake, Databricks, Amazon Athena, BigQuery, Apache Spark, Trino, and any future query engine that supports the standard. Fivetran's Managed Data Lake Service writes directly to Iceberg (or Delta Lake) on your object storage, automatically managing schema evolution, table compaction, and metadata cataloging — updating AWS Glue, Databricks Unity Catalog, and Google BigLake metastore so downstream engines always see current tables without manual DDL or scheduled refresh commands.

The strategic principle is ownership. Data in open Iceberg format on your own object storage belongs to you in a way that data locked inside a proprietary warehouse does not. When the next compute paradigm arrives — and it will — your data is already in the right format to serve it. You are not committed to any single vendor's future roadmap; you are committed to an open standard with broad ecosystem support and an active development community.

5. Think of new data sources

Earlier, you counted your known data sources. This step asks you to think about the ones you have not counted yet — the categories of data that are not in your current inventory but that will materially change what AI can do for your business. Your competition may already be working on these. The organizations that will have the most powerful AI in 3 years are not just the ones that have connected their CRM and ERP; they are the ones that recognized the competitive value of data categories that were not obvious to them 18 months ago.

The advantage compounds — but only if you start

Run the numbers on your data problem. Model the scenario where your competitor ran them six months ago and acted on it. Build every new pipeline with CDC and open formats as the default. Then expand your thinking beyond the sources you already know — to the data categories that will change what questions your AI can answer.

The organizations that will lead with AI in three years are building the data foundation for it right now. The window for first-mover advantage in data centralization is not infinite. The question is not whether to start. The question is whether you are starting before or after the competition.

[CTA_MODULE]

Apache Iceberg is a trademark of the Apache Software Foundation.