AI & Data Readiness Trends 2027: Why 60% of AI Projects Stall on Data

Key Takeaways

• The famous 60% number is narrower than it sounds. Gartner predicts organizations will abandon 60% of AI projects that lack AI-ready data. That is a statement about data, and it is good news if you fix the data.
• Ambition outran infrastructure. AI deployment went from two in five organizations to four in five in about two years. Data practices did not move nearly that fast.
• Context is becoming infrastructure. In 2027 the companies that write down what their data means will get more out of the same models than the companies that don't.
• Readiness belongs to a workflow, not a company. "Is our data AI-ready?" has no answer. "Is our data ready for this workflow?" does.
• Agents raise the bar. A chatbot that misreads a field gives a bad answer. An agent that misreads a field takes a bad action.

There's a Gartner prediction that shows up in nearly every AI-and-data conversation I'm part of, and its deadline is about two months out.

In February 2025, Gartner predicted that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. The same release cited a survey of 1,203 data management leaders in which 63% said they either do not have, or are not sure they have, the right data management practices for AI.

That number gets quoted as "60% of AI projects fail." Read it again. It says 60% of the projects without AI-ready data get abandoned. The failure has a named cause, and the cause is fixable.

Meanwhile, the ambition side kept accelerating. At its 2026 Data & Analytics Summit, Gartner said AI deployment grew from two out of five organizations in 2024 to four out of five today. Twice as many companies are running AI. Not twice as many companies cleaned up their data.

That gap is the story heading into 2027. Ambition sprinted. Infrastructure walked. Most stalled projects I hear about sit right in the middle of that gap, and the trends below are all versions of companies trying to close it.

Ambition Is Outrunning the Data Underneath It

The pattern is predictable at this point. A leadership team sees a demo. The demo works on a clean sample. A pilot gets funded. Then the pilot meets the real data: three CRMs from three acquisitions, a shared drive nobody has pruned since 2019, and a "revenue" field that finance and sales calculate differently.

The pilot doesn't blow up. It stalls. Accuracy sits at "pretty good," nobody trusts it enough to turn off the old process, and six months later it quietly loses its budget.

The survey data backs this up. In Informatica's CDO Insights 2025 survey, two-thirds of data leaders said they had not moved even half of their GenAI pilots into production. Back in 2024, Gartner predicted at least 30% of GenAI projects would be abandoned after proof of concept by the end of 2025, and poor data quality was the first reason on the list.

I don't think this means companies were wrong to move fast. Experimenting was the right call. The mistake was assuming the data work could happen later.

Price the data work into the pilot

If a pilot budget covers the model, the prompts and the interface but not the data cleanup, it is underfunded. Put a line in the budget for getting the source data into shape, and give it an owner. If nobody will own it, that tells you something about whether the pilot will survive.

Context Is Becoming the Infrastructure Layer

This is the trend I'm most convinced of. The model is no longer the scarce part. Every company has access to roughly the same frontier models. What they don't share is context: what their data means, how their business actually works, which numbers are trusted and which ones are folklore.

Gartner landed in the same place this year. The opening keynote at the 2026 summit called context the "new critical infrastructure" (that's from a vendor's recap of the session, so take the framing as theirs), and Gartner's own release tells data leaders to build "a single, unified context layer" to avoid misunderstandings and hallucinations.

Strip away the jargon and it means something plain. An AI system can read your tables. It cannot know that "status = 4" means the deal is dead, or that the Q3 numbers in one folder were restated. Somebody has to write that down.

Companies that do this work get compounding returns. Every new model plugs into the same context and gets smarter for free. Companies that skip it pay the setup cost again on every project.

Write down what your data means before an agent has to guess

Start small. Pick the 20 fields and documents your first workflow depends on. For each one, write what it means, where it comes from, who owns it, and when it's wrong. That document becomes the most valuable AI asset you own, and it is a spreadsheet.

Readiness Belongs to a Workflow, Not a Company

"Is our data AI-ready?" sounds like a reasonable question. It doesn't have an answer.

Gartner's own definition makes this clear: AI-ready data has to be representative of the specific use case, including its errors and outliers, and readiness is an ongoing practice rather than a one-time project. Your invoice data might be ready for an accounts payable agent and useless for a forecasting model. Same data. Different answers.

This changes how I think companies should approach the whole thing in 2027. A company-wide data overhaul is a two-year project that rarely finishes. A readiness check on one workflow takes weeks and produces a yes, a no, or a short list of fixes.

Score data against one workflow at a time

Here is the check I'd run before any AI project touches production data. Five questions, asked about one workflow.

If a workflow fails two or more of these, fix the data before building anything. If it passes all five, build fast.

Agents Raise the Bar Higher Than Chatbots Did

The first wave of enterprise AI was mostly chat. Ask a question, get an answer, a human decides what to do with it. Bad data produced bad answers, and a person caught most of them.

Agents change that. An agent reads data and then acts on it. It updates the record, sends the email, routes the ticket, files the report. If it misreads a field, the mistake doesn't sit in a chat window waiting for a human. It goes out into your systems.

So the same data that was "good enough" for a chatbot pilot is often not good enough for an agent. Quality matters more. Permissions matter a lot more. An agent with read access to the whole shared drive can surface a salary file in a summary nobody expected it to touch.

I'm bullish on agents. I also think a lot of 2027's stalled projects will be agent projects built on chatbot-grade data.

Map permissions before you map prompts

Before writing a single prompt, list every system the agent will touch and what it can do there: read, write, delete, send. Then cut that list to the minimum. It's the least exciting hour of the project and it prevents the most expensive kind of failure.

Governance Moves Upstream of the Build

For most companies, governance has been the thing that shows up at the end. The pilot works, then legal and security review it, then it waits.

That order is getting flipped. Gartner's 2026 guidance tells data leaders to make governance a value accelerator and to prevent "exposing the wrong data to the wrong people" as part of the foundation, not as a final gate. I agree with that, with one complication: most mid-market companies don't have a governance team to move upstream. They have a CFO, an IT lead and a policy document somebody wrote in 2023.

That's fine. Governance at that size can be small. Who owns each data source. What is sensitive. What the AI can and can't see. Written down before the build, and reviewed when a new workflow comes online.

Decide sensitivity before the data moves

Tag each source your workflow touches as public, internal, confidential or restricted. It takes an afternoon. Every later decision, from which model to use to whether an agent can act without approval, gets easier once that tag exists.

Pilots Will Be Judged on Data, Not Demos

Here's a prediction of my own for 2027, offered with honest uncertainty. The question boards and operating partners ask about AI will shift from "what are we piloting?" to "what is in production, and what does it run on?"

A demo proves the model works. It says nothing about whether your data will hold up at volume, with real users, on a Tuesday when the upstream system is late. The companies that ship in 2027 will be the ones that ask the data question first, before the demo gets anyone excited.

We see this in diligence work too. When PressW helped Ecliptic Capital evaluate a healthcare AI company, data quality and data governance were two of the five pillars we scored, alongside modeling, testing and scalability. An impressive model on weak data is a risk flag, and investors are starting to read it that way.

Make the data check the funding gate

Make the readiness check the gate for real funding. A pilot that passes gets a production budget. A pilot that fails gets a data budget first. Either way, nobody spends six months polishing a demo that was never going to ship.

The Gap Closes One Workflow at a Time

I don't know whether Gartner's 60% will turn out to be exactly right. Predictions are predictions, and "abandoned" is hard to measure. What I'm confident about is the direction. AI ambition keeps getting cheaper to act on. Data readiness doesn't. The distance between the two is where projects go to stall.

I also don't think the answer is a giant data transformation program before anyone touches AI. That's how you lose two years. The companies doing this well pick one workflow, check whether the data can carry it, fix what's broken, ship it, and move to the next one. Each pass leaves behind cleaner data and more written-down context for the next project to use.

If you want an outside read on where your data stands, that's what our AI & Data Readiness Assessment is for: a scored baseline across data, systems, security and teams, and a roadmap ranked by what your data can support today. If you'd rather start by comparing notes on one workflow, Book a Discovery Call.

FAQ