Buying AI Before Fixing Data — The Disaster That Has Already Started
by DL Keeshin
July 29, 2026
My last post stepped back from individual features to ask why the kDS Data Source Discovery (DSD) App design works, and landed on three properties that reinforce one another: it is genuinely data-driven, it uses AI responsibly and efficiently, and it keeps humans in the middle. Take any one away and the other two weaken.
That post ended on a phrase worth expanding: pre-agentic data infrastructure. This post is about what happens when organizations skip it. Spending money and resources to jump on the AI bandwagon without first having a strategy for data quality is a disaster waiting to happen — except it is not waiting anymore. It has already started.
AI does not fix bad data. It industrializes it. A broken spreadsheet misleads one analyst; a model trained on the same data misleads every decision it touches, at machine speed, with a confident tone and no error bars.
1. It Has Already Started
The failure pattern is no longer a prediction. It shows up in the research the same way from every direction. Gartner forecast that through 2026, organizations would abandon 60% of AI projects unsupported by AI-ready data, after finding that 63% of surveyed data management leaders either lacked or were unsure of the right data management practices for AI. That window is now closing, and the abandonment is visible in budget reviews rather than in vendor case studies.
The generative AI numbers are starker. MIT's Project NANDA found that roughly 95% of generative AI pilots produced no measurable return on the profit-and-loss statement, while RAND reported that more than 80% of AI projects fail — about twice the failure rate of conventional IT projects. The causes recur: unclear definitions of success, weak data foundations, poor integration into real workflows, technology chasing rather than business outcomes, and executive sponsorship that fades. Only the first and last of those are about leadership. The middle three are all about the data layer.
Meanwhile the spending has not slowed. Gartner projected worldwide generative AI spending at $644 billion for 2025, a 76% increase over the prior year. Enormous capital is being deployed on top of foundations nobody has inventoried. That is the disaster in progress: not a dramatic failure, but a slow, expensive accumulation of pilots that never reach production and models whose outputs no one trusts enough to act on.
“AI-ready” is a stricter standard than “analytics-ready”
Part of the trap is that organizations assume the data warehouse they built for quarterly reporting is good enough. It usually is not. AI-ready data demands continuous quality, active metadata, lineage, and governance applied at the cadence the model consumes data — not the cadence the BI team reports on. A monthly reconciliation job is fine for a dashboard someone reads on Friday. It is not fine for an agent making decisions on Tuesday afternoon.
2. Why Bad Data Is Worse Under AI Than It Was Before
Every organization has lived with imperfect data for decades and survived. The reason AI changes the risk profile comes down to four properties of how models behave.
| Property | Consequence |
|---|---|
| No ambiguity signal | A language model cannot tell you its input was malformed, stale, or contradictory. It produces fluent, confident output either way. Bad input does not fail loudly — it fails plausibly. |
| Scale and speed | A human analyst working from a bad extract makes one bad recommendation. An automated pipeline applies the same defect to thousands of decisions before anyone notices a pattern. |
| Loss of provenance | When output cannot be traced to its source, an error cannot be diagnosed — only distrusted. Once a business unit stops trusting a model, adoption ends regardless of accuracy. |
| Unknown coverage | Most organizations cannot enumerate their own data sources. A model trained on what was convenient to collect is silently blind to whatever was not — and silence looks exactly like completeness. |
That last row is the one this blog keeps returning to. You cannot assess the quality of a data source you have not identified. The undocumented shadow spreadsheet that three departments quietly depend on, the overnight file transfer nobody owns since the vendor contract changed, the legacy system whose extract is the real source of truth despite what the architecture diagram says — what we have called data dark matter. Data quality work that starts at the warehouse starts too late.
3. Five Strategies Before You Build
None of these require pausing AI ambitions for a year. They require sequencing the work so the AI you build has something solid to stand on.
Strategy 1 — Inventory sources before you assess quality
Start by finding out what exists. Not what the architecture diagram claims exists — what people actually use to do their jobs. That means interviewing subject matter experts across roles and business units, capturing their answers in a structure that survives the conversation, and consolidating what overlapping respondents say into a coherent picture. Every downstream quality effort depends on this inventory being reasonably complete. It is also the step most often skipped, because it is unglamorous and cannot be bought as a product.
Strategy 2 — Scope quality to a specific use case
“Improve enterprise data quality” is not a project; it is a mood. Enterprise-wide remediation programs have a long history of consuming budget for years and delivering governance artifacts nobody reads. Pick the first AI use case, trace backward to the handful of sources it actually consumes, and make those sources demonstrably trustworthy. Then do the next one. Narrow scope produces finished work; finished work produces the credibility to fund the next round.
Strategy 3 — Establish ownership before tooling
Most data quality defects are organizational, not technical. A field is inconsistent because two departments define it differently and neither has authority over the other. No tool resolves that. Before buying a quality platform, name an accountable owner for each critical data element and write down the definition that owner is accountable for. This is cheap, boring, and disproportionately effective.
Strategy 4 — Instrument continuously, not periodically
Quality is a property of a moment, not a certificate. Freshness, completeness, referential integrity, distribution drift, and volume anomalies should be measured on the cadence at which AI consumes the data. If a model reads a table hourly, a quarterly audit tells you nothing useful about the version it read this morning. Instrument first, then alert, then remediate — in that order.
Strategy 5 — Make lineage a requirement, not a report
Build the ability to answer “where did this number come from?” into the pipeline itself rather
than reconstructing it after an incident. In the kDS DSD App this is what the UUID provenance arrays do:
source_summary_id_set and
interview_id_set carry the trail from synthesized analysis back to the
specific interviews that produced it. The principle generalizes. If a finding cannot be walked back to its
sources, it cannot be defended when challenged — and it will be challenged.
Before funding an AI initiative, ask the sponsor three questions: Which data sources will it consume? Who owns each one? How would we know tomorrow morning if one of them broke? If the answers are not ready, the project is not ready. The answers are the project.
4. Start With Data Quality Agents
Here is the constructive turn. The argument above is not that organizations should postpone AI — it is that they are pointing it at the wrong problem first. The best early application of AI in most enterprises is not a customer-facing assistant or an autonomous workflow. It is a set of narrow agents whose entire job is to make the data foundation trustworthy.
This inverts the usual sequence in a way that pays off twice. The organization gets a real, scoped AI deployment with measurable output, and the output is the foundation the ambitious use cases were going to need anyway. Some agent patterns worth building first:
| Agent | Job |
|---|---|
| discovery | Interview SMEs about where data lives, who owns it, and how it moves — then consolidate overlapping accounts into a single source inventory with provenance. This is precisely what the kDS DSD App does. |
| profiling | Continuously profile tables and files: null rates, cardinality, format drift, distribution shifts, volume anomalies. Summarize findings in language a business owner can act on rather than a metrics dump. |
| reconciliation | Compare the same entity across systems — customer, product, site, account — and surface the disagreements that master data management is supposed to resolve but usually has not. |
| definition | Detect where the same term carries different meanings across business units, draft candidate definitions from how people actually describe the field, and route them to the named owner for confirmation. |
| lineage | Reconstruct and maintain the map from source to consumption, flagging orphaned extracts and undocumented dependencies as they appear rather than during the next audit. |
Build them the way the platform builds everything
The same three properties from the last post apply directly. Keep the agents data-driven: give them clean, typed, structured input and store what they produce in a real schema, not in a chat log. Keep the AI responsible and narrow: an agent that flags a suspected duplicate is useful; one that silently merges records is a liability. Prompts belong in configuration where they can be versioned and audited. And keep humans in the middle: a quality agent should propose, and a named owner should dispose. Every remediation is a promotion, and every promotion deserves a person saying “yes, this is correct.”
The Sequence Is the Strategy
The organizations that will report real AI-driven results are not the ones that started earliest. They are the ones that spent the first phase finding out what data they have, who owns it, and whether it can be trusted — and only then pointed models at it. The ones now quietly writing off pilots did the reverse: bought the capability, then discovered the foundation. The bill for that ordering is coming due across the economy right now, and it is being paid in abandoned projects rather than in headlines.
This is what pre-agentic data infrastructure means in practice. Before agents can act on enterprise data, someone has to answer “where does this data live, and can we trust it?” That question is not a research project to be deferred until after the AI initiative. It is the AI initiative — the first one, the one that makes the rest possible. The good news is that AI is a genuinely excellent tool for answering it, provided you point it there first.
Sources: Gartner, Lack of AI-Ready Data Puts AI Projects at Risk (February 2025); MIT Project NANDA, The GenAI Divide: State of AI in Business (July 2025); RAND Corporation research on AI project failure rates; Gartner generative AI spending forecast (2025).
Interested in what this architecture can do for your organization's undocumented data landscape? Reach out at talk2us@keeshinds.com or visit keeshinds.com. We are actively onboarding select beta partners.
As always, thanks for stopping by!