The chart was fine. The number was not.
A sales dashboard can be sharp, current, and completely wrong. Not because someone edited the slide. Call this an illustration, not a case file: a transformation reads cents as dollars, leaves the source row untouched, and sends a confident figure downstream. The warehouse looks healthy. The forecast does not.
A flawed chart embarrasses a meeting. An agent that trusts the same figure can book the discount, reroute the lead, or tell a rep the deal is healthy. Picture a duplicated account, a currency unit nobody checked, and a concession auto-approved before anyone opens the CRM. AI for sales gets expensive without a dramatic breach. One trusted field and permission to act is enough.
Regulated clinical and genomics labs got strict about identity and traceability because the harm is not theoretical. A misidentified sample can enter a patient’s record or steer a diagnosis. A flawed churn forecast usually ends in a tense meeting, not a legal file. The damage is not equivalent. The mechanics still rhyme: bad input, a transform that leaves the original looking innocent, history that vanishes, a field reused for a job it was never meant to do. Softer consequences explain why BI discipline slips. They do not justify keeping the slack once AI agents stop displaying the number and start spending against it.
Low maturity is a habit, not a score for a board slide. Spreadsheet metrics, ownerless dashboards, definitions nobody agreed, and no one who will vouch for the record. That pattern is the evidence you can actually see. Tim King at Solutions Review flagged a gated Gartner note, Four Steps to Improve Data and Analytics Capabilities When Business Intelligence Maturity Is Low, for subscribers on 11 December 2018. Its sample and wording are not publicly quotable, so leave any headline percentage off the deck.
Four ways a clean-looking number fails
“Data quality” is too blunt a phrase for a figure that smells wrong. The chart is a symptom. A useful review asks which layer broke. What follows is a working model, not a regulation. It earns its keep by stopping teams from fixing the dashboard when the defect lives three steps upstream.
Input integrity asks whether the record was genuine when it arrived. A mistyped opportunity amount. A duplicated customer. A file that was never the system of record. If you cannot check the source while it is still in reach, you are already late.
Processing integrity asks whether transformations preserved meaning. The cents-as-dollars slip is the cartoon version: the original row is correct, the calculated field is not. Lineage is what lets you walk the reported value back to the row that did not lie.
Storage integrity asks whether the retained copy stayed correct, and whether later edits left a history. Silent overwrites hide change. Last quarter’s truth disappears without a fingerprint.
Fit for purpose asks whether a later use still matches the job the data was collected to do. A support-ticket field can be accurate and still be the wrong evidence for a churn model or a sales score. Right bytes, wrong job. A renamed field hides misuse. That is a different failure from a silent overwrite.
Three habits worth stealing
Labs did not invent chain of custody because they enjoy paperwork. They tie a result to a specific sample, a handling history, and a person. Automation did not erase that question. It moved the check earlier: refuse to start until the setup can be inspected. The lesson for a pipeline is less heroic cleanup, and more refusal at the door. You do not need to buy a robot.
I would not import a clinical quality system wholesale. A churn forecast is not a diagnostic. I would import three habits, and treat them as non-negotiable once models and agents sit in the path.
Validate at capture. Check the record when it enters, while someone can still open the source system and see the truth. A schema contract at ingestion (agreed columns, types, and allowed values) is a cheap gate. So is rejecting a file that fails a row count or a currency unit. Point-of-capture checks will not erase every reconciliation meeting. They stop one class of bad records from ever reaching a reporting mart or a training set. They are not the whole control. Lineage and history still have to hold.
Make lineage mandatory. If a reported value cannot be followed through each transformation back to a source, it is an opinion with a chart attached. Version the logic. Prefer transformations you can rerun and explain. When finance asks why pipeline moved, “the model said so” is not an answer. “This job, this commit, this source extract” is.
Keep provenance the way a lab keeps custody. Attach where the data came from, who or what handled it, and what changed. A field that meant “open ticket count” in 2023 and “escalation severity” in 2025 should not wear the same name in an agent’s memory.
Map those habits onto the four layers and the gaps show up fast. Capture checks defend input. Lineage defends processing. History defends storage. A written purpose for each reused field is the start of fit for purpose. This is not a product. It requires an owner.
Agents do not forgive a sloppy input
AI automation gets more expensive the moment the system stops advising and starts doing. Feeding stale or incomplete inputs into a model is a reliable way to get a fluent wrong answer. How often, and at what dollar cost, is not a figure worth pretending to have pinned down. The practical claim is narrower. Once a bad value has been copied into marts, features, and agent context, you may not know every place it landed. Fixing the source does not automatically fix the copies. Catching it at entry is the bet with the better odds.
A category name is not a substitute for that record. In a market guide dated 18 February 2025, Market Guide for AI Trust, Risk and Security Management, Gartner analysts Avivah Litan, Max Goss, Sumit Agarwal, Jeremy D’Hoinne, Andrew Bales, and Bart Willemsen set out the firm’s view of AI TRiSM (trust, risk, and security management for AI). Palo Alto Networks’ page on that guide frames the topic as governance plus runtime inspection and enforcement. Public summaries describe those themes. They do not establish a chain-of-custody requirement for training data.
Gartner is not handing you a standard you comply with by buying a logo. Its own disclaimer is the right size for this: its publications consist of the opinions of Gartner’s Research & Advisory organization and should not be construed as statements of fact. They are not a vendor endorsement either. If you want a model you can defend, you still have to know where the rows came from and how they were handled.
Runtime inspection matters for a different reason. Even clean inputs can be misused once an agent is loose in your systems. Input discipline without inspection is half a control. Inspection without input discipline is a fast way to audit garbage. The input problem does not shrink as the system gets more ambitious. It compounds.
Questions worth answering on Monday
- Why borrow lab practice if a bad forecast will not harm a patient?
The harm is not the same: a misidentified sample can reach a patient, and a bad forecast usually reaches a meeting. The mechanics still rhyme, and the penalty rises once an agent turns the metric into an action, so the old slack is no longer a good bargain.
- Which three habits are actually worth copying?
Validate at the point of capture, keep lineage so a reported value can be walked back to its source, and retain provenance closer to chain of custody. They are practices, not a product, and nobody has handed you a studied return with a price tag.
- When is a bad record too late to fix?
It is too late to be sure once you no longer know which marts, features, and agent memories copied it, not merely when the source row can still be edited. No public case here puts a reliable dollar figure on the cleanup. The cheaper move is still to refuse the record at the door.
- What should you be able to name before the next pilot?
The metric that already moves money, its source, the job that transforms it, who may change it, and what the field is not allowed to mean. If that list takes more than an afternoon, the gap is custody, and model choice will not close it.
Own the number before you buy the model
The labs did not get rigorous because their instruments looked advanced. They got rigorous because an unchecked result had a victim. Yours is quieter: a board that funds the wrong motion, a rep coached off a phantom pipeline, an agent that automates a mistake at 2 a.m.
Before the next AI-for-business pilot, pick one metric that already moves money. Name its source. Name the job that transforms it. Name who is allowed to change it, and what the field is not allowed to mean. If that takes more than an afternoon, the gap is a custody problem, and a new model will not close it. Fix that, and the dashboard, the forecast, and the agent all get harder to fool.