Nobody sets out to deploy an AI agent hoping it fails. Three years into the enterprise AI race, it’s happening at scale.

Enterprises have collectively spent $30 to $40 billion on generative AI, according to MIT Media Lab’s Project NANDA. Much of what matters is still locked in mainframes, legacy databases, spreadsheets someone built years ago, PDFs, and applications never designed to share with each other. Ask three business units to define “active customer” and you’ll get three different answers. Ownership’s rarely clear, and the one person who understands why a field is formatted that way is a few years from retirement.

None of that is new. What is new is that AI does not forgive it the way a person would.

The same 2025 MIT report found that only about 5 percent of generative AI pilots produce real, measurable value. In almost every case, it comes down to brittle data workflows, missing context, and systems never built around how the business runs.

Analysts have always worked around bad data. Someone notices a number looks off and fixes it before it causes a problem. AI agents have no such instinct. They approve claims, route cases, and make decisions that affect people’s lives, and when the data underneath is incomplete, they make the wrong call with confidence. AI will not fix a weak data foundation. It will scale the consequences of one.

Getting the data in is still the hardest part

Traditional ETL was built for a simpler world. It holds up when sources are documented and stable and falls apart when they are not — when the format changes without notice, the field names came from someone else’s system, and nobody on your side controls either.

“The hardest work happens before anything gets integrated, getting data out of a legacy or external system and into a form the business can use,” said Tim Bond, Chief Product Officer at Adeptia. “Proprietary formats, undocumented fields, a spreadsheet with macros nobody remembers writing, PDFs with handwritten notes in the margin. Cleaning up one employer’s enrollment file still takes 15 to 20 hours in a lot of shops.”

Retirement-plan onboarding shows how fast this becomes a revenue problem. A provider change means pulling data from a prior recordkeeper, a payroll system, an HRIS platform and sometimes a broker, with every discrepancy resolved before go-live. On a $50 million plan, each month of delay defers roughly $40,000 in fees.

Regulated industries have more on the line

There, the exposure extends to compliance risk, legal risk, and larger financial consequences. Industry analysts estimate that 80 to 90 percent of enterprise data is unstructured. Group health insurance is a prime example.

Enrollment files arrive with merged cells, half a dozen tabs, custom macros, and PDFs that don’t render the same way twice. When a dependent’s eligibility gets mis-mapped or an effective date goes missing, that’s a denied claim, a compliance flag, or a check that never should have gone out.

Carrier transitions expose the same weak points. They typically run 60 to 90 days, most of it mapping and testing, because one carrier calls a field “EE_DOB” and another calls it “BirthDate,” and historical claims data rarely follows the same structure twice. Every mismatch gets reconciled by hand. That debt grows more expensive as AI pulls from more systems, more often.

“Real-time” doesn’t mean much without trust

“Everyone wants real-time data yesterday, but it isn’t worth much if you can’t say where it came from or whether you’d trust it to decide on its own,” added Bond. “An agent needs to know whether it can reach the data, whether the data is current, which system wins when two disagree, and whether the decision is its to make at all. That is what a governed data dictionary and a real semantic layer are for. Without shared definitions, an agent runs with whatever was baked into the query it was handed.”

Quality checks must start the moment data comes in the door. You need to know that a process ran, that the output is the shape you expect, and that the numbers hold up against the business. Lineage matters too. An agent needs enough visibility into where data came from and how fresh it is to flag an uncertain answer or hand it to a person.

This is not a fringe concern. According to a recent survey from my company, Prosper Insights & Analytics , nearly 40 percent of executives and business owners named a lack of human oversight as one of their top concerns about AI, close to the 40.4 percent of the general public who said the same. Concern about oversight is strongest among the people signing the checks.

Bringing ETL into the AI era

Enterprises do not need another point solution. They need an integration foundation that handles real-world variety without giving up governance and control.

Intelligent ETL means AI-driven pipeline design and built-in industry knowledge paired with deterministic execution. Let AI do what it is good at, speeding up schema discovery, field mapping, and validation. Keep the workflows predictable, so what happened can be traced, repeated, and defended.

Even among executives and business owners, just over 33 percent think handing decisions to a fully autonomous system is a good idea, according to the same Prosper Insights & Analytics survey. Nearly a third said no outright and another third are unsure. Trust will not catch up on a better prompt, but when the data underneath is something you can stand behind.

None of this is a one-time fix. It requires ongoing ownership, observability, lineage, governance, and maintained definitions as the business changes. Every new use case tests the foundation. The systems that hold up improve with each correction rather than needing the same fix twice.

Where to start: Agree on definitions before automating anything, check quality at intake rather than after an incident, and keep lineage good enough that an agent knows when to stop and ask. The enterprises that get the fundamentals right onboard faster, recognize revenue sooner, and keep customers longer.

Disclosure: The consumer sentiment study referenced above was conducted by my company , Prosper Insights & Analytics . This is the same dataset used by the National Retail Federation, and available from Amazon Web Services, Databricks, and the London Stock Exchange Group for economic benchmarking.