MELLORCATalk to Mellorca
Patterns

Data, Reporting & Management Visibility

Why AI Cannot Fix Data Nobody Trusts

When staff already doubt the numbers, an AI tool inherits the doubt and delivers it faster. Fixing sources, definitions and ownership has to come before the pilot.

Mellorca Patterns2026-10-11

Observable condition

The leadership team decides it is time to use AI. A pilot is approved: a tool to forecast demand, summarise customer history or draft collections follow-ups. The demonstration is impressive. Then the tool meets the real data, and the results are puzzling. The forecast ignores a product line. The customer summary shows two customers who are really one. The collections draft chases an account that was settled last week.

The reaction inside the business is often to blame the technology. It is called unreliable, immature or not suited to a company of this kind. The more accurate observation, made quietly by the people who maintain the records, is that they would not have trusted those numbers either.

An illustrative example: a Johannesburg wholesale distributor trials an AI assistant to answer sales queries about stock and pricing. The assistant is fluent and fast. It also quotes a price that was changed two months ago and reports stock for a warehouse that was closed. The sales team stops using it within a fortnight, not because it is slow, but because they cannot tell which answers are right.

Realisation

The realisation is that AI is a consumer of data, not a cleaner of it. A model can draft, summarise, classify and predict, but it can only do so from what it is given. If the underlying records are duplicated, out of date or defined differently in different places, the model will produce confident-sounding output built on shaky ground. Speed makes the problem larger, not smaller.

IBM's overview of data quality describes the standard dimensions on which data is judged, such as accuracy, completeness, consistency and timeliness. These are not new concerns, and AI does not exempt a business from them. If anything, it removes the human who used to notice that a figure looked wrong before acting on it. The Insight why AI makes data quality more important, not less explains this in detail.

Realisation prompt: take one report your management team already relies on. If you removed the person who knows which columns to ignore, how many people would still trust it?

Identification

This is a pattern of distrusted data feeding an expensive tool. The visible symptom is an AI project whose outputs people double-check, discount or abandon. The common assumption is that the model needs better prompting or a different vendor. The structural cause is that the business never settled which source is authoritative, what its key terms mean or who maintains the records, so there is nothing reliable for the tool to stand on.

The Discovery why AI tools fail without authoritative data covers the technical side. The Pattern when nobody knows which system holds the correct data describes the human side of the same problem: if staff cannot say which source is correct, a model cannot either.

Diagnosis

  • No agreed source of truth. Customer details, prices or stock exist in several systems, each partially right. IBM's discussion of master data management addresses the idea of a single, trusted reference for core entities.
  • Definitions drift. "Active customer", "order value" and "margin" mean different things in finance, sales and operations. The model inherits whichever definition it happens to see.
  • Corrections happen downstream. Staff fix numbers in spreadsheets or reports rather than at the source, so the system of record stays wrong while the visible report is right. The model, reading the source, learns the wrong version.
  • Nobody owns quality. Data is entered by many people and used by many more, but no one is accountable for how accurate or complete it is.
  • Staleness is invisible. Records carry no clear sign of when they were last confirmed, so a price from last year looks the same as one from this morning.
  • Quality is found by the consumer. Problems are discovered when a report or an AI output looks odd, rather than being caught at entry. The Discovery why data quality problems are found after reports are published shows how late that is.

Commercial impact

The commercial value at stake is the return on AI investment, decision quality, customer trust and management time.

The first cost is the failed or stalled project. Licences, integration effort and staff time are spent on a tool that people then decline to rely on. In a market where budgets are watched closely and the Rand does not stretch as far as it once did on imported software, an abandoned pilot is a visible loss.

The second cost is the quiet one: a tool that is used and wrong. If a customer-facing assistant quotes an outdated price or a collections tool approaches a customer who has already paid, the business pays in credibility and sometimes in money. The Insight why bad data becomes more dangerous when AI can take actions explains why the risk rises once a system can do more than advise.

The third cost is the verification burden. If staff must check every AI output against the source, the saving the project promised disappears. The fourth is the loss of confidence: after one poor experience, teams become sceptical of later, better-prepared initiatives.

Common misidentification

The usual response is to switch vendors or models. A different model will make different mistakes on the same inputs, but it cannot make the inputs reliable. Another response is to add a clean-up phase at the start of the project, run by consultants, and then assume the problem is solved. Without owners and routines, the data drifts back to its earlier state within months.

A third response is to give the AI tool access to everything and let it work out which source to believe. That makes the confusion larger, because the tool now sees every conflicting version at once. The Insight why giving AI access to more systems is not automatically better makes the point more fully.

Possibility

In a better state, the business can state for each important kind of data where it lives, who owns it and what it means. Core records have a single reference. Corrections are made at the source. Records show when they were last confirmed. An AI tool pointed at that foundation behaves predictably, and when it errs, the error can be traced to a specific record and fixed once.

That foundation is valuable whether or not AI is ever used. Reports become quicker to produce and easier to defend, and staff stop keeping private copies of the numbers.

Intervention

  1. Choose one use case and its data. Before any tool is bought, list exactly which records the use case will read: customers, prices, stock, invoices.
  2. Test trust directly. Take a sample of fifty records from each source and ask the people who use them whether they are correct. Count the disagreements.
  3. Name the authoritative source. For each kind of record, decide which system is the master and write it down. Make other systems read from it rather than keep their own copies.
  4. Fix definitions. Agree one meaning for each term that appears in reports and prompts, and publish it where staff and tools can see it.
  5. Assign owners. Give each data set a named person responsible for accuracy, completeness and a regular check.
  6. Correct at the source. Change the practice that lets staff patch reports instead of the record, and make the right path the easy path.
  7. Start the AI pilot small. Run it on the cleaned data set, compare outputs with a human check, and widen only when results hold.

Diagnostic questions

  • For the data your AI idea needs, can you name the system of record and the person accountable for it?
  • Do finance and sales give the same answer to the same question about a customer or a product?
  • How often do staff correct a number in a report instead of in the system?
  • Would a new employee trust the data without being warned which parts to ignore?
  • If an AI tool gave a wrong answer tomorrow, could you trace it to the record that caused it?

Bottom line

AI amplifies what it is given. Where the data is trusted, that is an advantage. Where nobody trusts it, the tool simply spreads the doubt faster and more confidently. The practical order is therefore data first: agree the source, the meaning and the owner, and only then ask a model to work on top.

Sources and further reading

Related Mellorca reading

Article summaryWhen staff already doubt the numbers, an AI tool inherits the doubt and delivers it faster. Fixing sources, definitions and ownership has to come before the pilot.
Explore our solutionsStart a conversation