Obtainium.ai

Why 95% of AI Projects Fail: The Data Foundation Problem

MIT Technology Review recently published findings from a conversation with leaders at Databricks and Infosys that should give every small business owner pause:

The Uncomfortable Truth About AI Failure Rates

MIT Technology Review recently published findings from a conversation with leaders at Databricks and Infosys that should give every small business owner pause: 95% of enterprise AI projects fail to deliver meaningful business value. The reason is not what most people assume. It is not that the AI models are weak, or that the prompts are wrong, or that the technology is not ready.

A contributing reason is the data itself.

AI is only as good as the data it can see, trust, and act on. Most organizations cannot give their AI clean, unified, governed data — so the AI cannot do useful work.

For small and mid-size businesses, this finding is both a warning and an opportunity. The warning: rushing to deploy AI agents on top of messy, scattered data is a recipe for wasted money. The opportunity: SMBs have far less data sprawl than enterprises, which means the foundation can be fixed in weeks rather than years.

What "Bad Data Foundation" Actually Looks Like

When we talk about a weak data foundation, we are not describing some abstract enterprise problem. Here is what it looks like in a typical small business:

When you bolt an AI agent onto this kind of environment, the AI has to guess. It guesses wrong. It hallucinates. It gives different answers to the same question depending on which system it happened to query first. Then leadership concludes "AI doesn't work" and shuts the project down.

The AI worked fine. The data did not.

The Three Things That Have to Be True

The MIT piece highlights three foundational requirements that distinguish AI projects that succeed from the 95% that fail. They translate cleanly to small business reality.

1. Unified Data in Open Formats

Your customer data, your operational data, and your historical records need to live in one place that all your tools can read from. "Open formats" means standard file types and database structures that are not locked to a single vendor. If your data is trapped inside a SaaS tool that charges you to export it, you do not actually own your data.

For an SMB, this often means consolidating into a single source-of-truth database (Postgres is a common choice) with your various tools writing to and reading from that one place — instead of each tool maintaining its own private copy.

2. Access Controls and Governance

Databricks calls their version of this Unity Catalog. The principle is simple: every piece of data has a clear owner, a clear set of permissions, and a clear audit trail. When the AI agent reads a customer record, you can prove which agent read it, when, and why.

For small businesses, this matters for three reasons:

You do not need a $100K enterprise governance platform. You need a clear policy and a system that enforces it.

3. Tied to Measurable Business Outcomes

The most damning finding in the MIT piece is that most AI deployments are stuck in pilot purgatory. Companies build a prototype, demo it to leadership, get applause, and then never connect it to a real business metric. Six months later nobody can prove the AI saved a dollar or earned a dollar.

Before you deploy any AI agent, you should be able to answer:

If you cannot answer those four questions, you do not have an AI project. You have a science experiment.

What Lakebase and Operational AI Data Actually Mean

Databricks announced Lakebase, a new operational database designed specifically for AI agents. The technical details are not what matters for SMBs — what matters is the underlying shift it represents.

Traditional databases were built for humans clicking buttons in software. AI agents work differently:

The industry is racing to build databases that fit this pattern, and the cost-per-query is dropping fast. For SMBs, this means operating an AI agent on real business data is becoming dramatically cheaper — provided your data is in a format the agent can actually use.

Two years ago, running an always-on AI agent against your CRM was a five-figure-per-month proposition. Today the same workload can run for tens of dollars. The bottleneck is no longer cost — it is data readiness.

The Practical SMB Playbook

If you are a small business owner reading this and wondering where to start, here is the order of operations that actually works:

  1. Inventory your data sources first. List every place customer or operational data lives. Most SMBs are surprised to find 8-15 separate systems.
  2. Pick a single source of truth. Usually a database you control, not a SaaS tool. This is where the consolidated, clean data lives.
  3. Build one boring pipeline. Get data flowing from your most important source (often your CRM or booking system) into the source of truth. Make it work before adding more.
  4. Add governance from day one. Who can read what. Who can write what. Logged. Even a simple permissions table is better than nothing.
  5. Define the business outcome before you deploy AI. Pick one metric. One. Write down the baseline.
  6. Deploy a narrow AI agent against the clean data. Not a general assistant — a specific agent solving a specific problem.
  7. Measure against the baseline. Kill it or scale it based on the number, not the demo.

Notice that AI is step 6 of 7. That is not an accident. The 95% failure rate happens when companies start at step 6.

The Takeaway

The single most valuable thing a small business can do to prepare for AI is not to buy AI. It is to clean up the data first. Unified storage, clear ownership, measurable outcomes. The companies that get this right will deploy AI agents that pay for themselves in months. The companies that skip it will join the 95%.

This is unglamorous work. It does not demo well. It will not impress anyone at a networking event. But it is the difference between AI that quietly compounds value year over year, and AI that becomes another expensive cautionary tale.

If you are evaluating AI tools or vendors right now, the question to ask is not "what can your AI do?" The question is "what does my data need to look like for your AI to work?" A vendor who cannot answer that clearly is selling you a demo, not a solution.

Ready to Put AI to Work?

Whether you know exactly what you need or want help figuring it out, we have a path for you.

Know what you need?

Book a Free Call

15 minutes. We'll map your workflows to the automations that'll move the needle fastest. No pitch deck, no pressure.

Book a Free Call
Not sure where to start?

AI Readiness Audit

A full analysis of your operations — specific automation recommendations, ROI projections, and a custom implementation roadmap.

Learn About the Audit

Obtainium.ai builds custom AI automation for service-based small businesses. 30+ years in IT and IT security, CISSP and CAISS certified — we build systems that run in production, not demos that look good in a sales meeting. Based in Reno, NV, serving businesses nationwide.