Alaaeddine Hammoudi

Writing / article

Data maltreatment: the quiet tax on the P&L

Data quality is the symptom, data maltreatment is the process that causes it. Its mechanisms, its cost, why incentives keep it alive, and what it does to organisational knowledge and AI.

Author
Alaaeddine Hammoudi
Published
Reading
7 min

Quality engineers worked out decades ago that inspecting finished products doesn’t remove defects. You have to fix the process that creates them. Data management has mostly kept the inspection habit. We measure data quality, which is a state, and we rarely name the behaviour that produced it.

I’ve started calling that behaviour data maltreatment: any recurring action or omission in the life of data that makes it less fit for the purpose it serves. The word is deliberately strong, but it says nothing about intent. Most maltreatment is ordinary people making sensible shortcuts inside processes that reward them. Only a small part is misconduct. Naming it moves the conversation away from cleaning data after the fact and towards how data is created, changed and used every day.

Where I first ran into it

I didn’t come to this through governance frameworks. I came to it through migrations and AI projects, which is where maltreated data stops being abstract.

Move a compliance codebase from one platform to another and you rediscover, rule by rule, which fields people actually trust and which ones everyone has quietly learnt to work around. Build a text pipeline on customer comments and you find free text where codes should be, and codes that no longer mean what their label says. Put a retrieval assistant on top of internal documents and, within days, it surfaces outdated versions and contradictions that nobody had noticed for years.

Each time, the fix that mattered was rarely technical. It meant finding the step in the process where the data was being bent, and understanding why someone had a good reason to bend it.

Four ways data gets maltreated

FormMechanismTypical signal
NeglectNo owner, no definition, history overwritten, data collected and never usedUndefined columns, missing lineage; in one 2020 survey, 68% of the data available to businesses went unused (Seagate and IDC)
ManipulationValues changed because the true value is inconvenient: a date moved to avoid a delay, a category changed to avoid a control, a case closed to meet a target and reopened under a new numberEdit spikes before deadlines, reopen rates, status changes with no matching event
Source alterationNotes written on documents during processing: margin codes, highlights, sticky notes scanned with the pageArchived documents that differ from what was received, extraction errors on annotated pages
MisuseFields repurposed (“01/01/1900” for “unknown”), corrections made in spreadsheets or emails, data reused outside its original contextSentinel values, free text in coded fields, manual reconciliations

What strikes me about all four is that each one is locally rational. It saves the person a few minutes, or avoids a problem today. Nobody is being careless on purpose. The cost simply shows up somewhere else, later, on someone else’s desk.

What it costs

Good numbers are rare. Most of the ones in circulation are estimates or surveys, several are a few years old, and I’d treat each one with some caution. I’ve kept the year of each so you can judge for yourself. Taken together, though, they point the same way.

ImpactEvidence
Operating cost$12.9M per year on average per organisation (Gartner, 2021, based on costs self-reported by surveyed companies); 15 to 25% of revenue for most companies (Redman, 2017, an estimate)
Error rates47% of newly created records had at least one critical, work-impacting error; only 3% of the data quality scores studied could be rated acceptable (Nagle, Redman, Sammon, 2017)
RevenueUnity Software: management estimated that bad data ingested by one of its tools would cost about $110M of 2022 revenue (Q1 2022 earnings call)
RegulationCitibank: a $400M civil money penalty from the OCC in 2020, citing deficiencies in risk management and “data governance”; then about $135.6M more from the Federal Reserve and the OCC in 2024, with the Fed citing “insufficient progress remediating its problems with data quality management”
ProductivityNearly 20% of the workweek spent looking for internal information or tracking down colleagues who can help (McKinsey Global Institute, 2012)
AIIn 2024 Gartner predicted that at least 30% of GenAI projects would be abandoned after proof of concept by end of 2025. Gartner later reported that over 50% had been abandoned by then, with poor data quality first among the reasons. It also forecasts that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data (2025 forecast)

The 1-10-100 rule, a quality-management heuristic usually credited to Labovitz and Chang, gives the shape of the curve. The ratios are illustrative and won’t match any real cost base exactly. Preventing an error costs 1, correcting it costs 10, and letting it reach a customer or a regulator costs 100. Maltreatment amounts to a standing decision to pay at the expensive end.

Why it keeps happening

Once you look at it as an economist would, the persistence stops being surprising. Four things are true at once.

The person who creates or alters the data keeps the time saved, while the cost lands downstream, on reporting, risk, customer service and models. Economists call that an externality.

The saving is immediate and the cost arrives months later, usually in another budget.

Nobody owns the whole life of a dataset. IT owns systems, the business owns processes, projects end. With no residual claimant, nobody has a reason to optimise it.

And there is no line in the P&L called “bad data”. The cost hides in overtime, rework and write-offs.

Put those four together in any sector, with any technology, and you get the same result.

How errors climb into knowledge

The DIKW pyramid (data, information, knowledge, wisdom) is useful here as a dependency chain. Each level is built from the one below, so errors don’t stay at the bottom. They travel up and compound.

A quick calculation makes the point. If a decision depends on five fields, each 95% correct and independent of the others, only 77% of decisions rest on fully correct inputs (0.95⁵). With ten fields it drops to 60%. Field-level quality that looks fine on a dashboard turns into decision-level quality that isn’t.

Organisational knowledge sits at the top of that chain. Some of it is written down in documents and rules. Much of it lives in the heads of experts who know which fields to trust and which to double-check. In practice those experts are an error-correction layer, undocumented and impossible to scale, and it walks out of the door when they retire or move on.

GenAI skips that layer entirely. A model reads raw text and produces something that looks like knowledge, with no idea which value was manipulated or which page was annotated. It inherits the error rate, then serves it, fluently, to anyone who asks.

The strategic angle

In finance, alpha is the return above the market. With AI, foundation models are the beta: you and your competitors can buy the same ones. What can’t be bought is your own data and organisational knowledge: claims, contracts, customer interactions, the decisions your experts made over twenty years.

Maltreatment erodes exactly that. Two companies running the same model on data of different quality end up with very different assistants, and the one with maltreated data pays the same price for the worse one.

What I would do about it

  1. Pick the critical data elements: the handful of fields that drive revenue, cost or regulatory reporting. Measure their quality and express defects in money and risk, because that is the language that gets budget.
  2. Act at the point of creation, with controlled values and validation where the cost of a fix is lowest.
  3. Look for manipulation openly: reason codes and audit trails for overrides, and a regular look at edit patterns around deadlines. The aim is to find the targets and rules that make people bend data, rather than to catch them doing it.
  4. Keep sources intact: preserve the original document and store annotations as structured data next to it.
  5. Give each critical dataset one owner, with quality targets in their objectives.
  6. Hold AI-produced data to the same rules, with provenance, version and quality checks for every value a model writes. I come back to this in why I’m still a Data & AI leader.

Seen this way, data maltreatment is an incentive problem with a measurable cost. Incentive problems tend to change once the cost is visible and lands on someone who can act on it.

Sources