Tech Explained: AI engines in firms frequently run on adulterated oil in Simple Terms

“Data is the new oil” is an often-used phrase. By extension then, AI and analytics are the engine spinning data into insight, action and (ideally) value. An engine running on adulterated oil sputters, stalls and underperforms.
AI rarely fails because algorithms are weak — more often it underdelivers because organizations are blind to the state of their data. In this article I will argue for the need to build and deploy a Data Quality Assurance function (DQA) which will engineer data accuracy into organizational processes; doing this in my experience can improve ROI from investments in AI and analytics by a factor of 7-10X.

The Invisible Problem
Few organizations measure or worry about, the quality of their data. Yet beneath the surface, poor-quality data can quietly undermine ambitions: afflicted companies will continue to get only a fraction of the return their AI investments could generate. In most organizations, data quality is everyone’s job — which means it is effectively no one’s job. CXOs, meanwhile, are consumed by the promise of what AI can do with good data, not by the drudgery of ensuring it’s trustworthy. The resultant is a perfect organizational recipe for quiet neglect.
How to recognize if this is happening at your firm: Symptoms
The problem is most apparent if in business reviews, participants frequently digress into debates on the accuracy of data being presented or functions want the data circulated for review before the meeting. If you see these symptoms, you can safely conclude that the data quality is likely poor!

Why Fixing It Is So Hard, Especially in Manufacturing
In manufacturing, both customer-related data that is scattered across CRM, billing, support and service systems and operations data both tend to be incomplete and unreliable. And hence, while investments in AI and analytics still add value, they stop short of delivering the step-change that is possible – not because the technologies lack potential, but because the underlying data foundation isn’t strong enough to fully unlock it.

Fixing this data quality isn’t trivial. It is invisible, slow, painstaking work that cuts across almost every function, plant, process and system. And it is tough since, as mentioned earlier, failure modes tend to be more administrative, not technological:
• Split accountability: Functions generate data but think tech owns it. Tech meanwhile builds and maintains systems but can’t fix upstream process behavior.
• Within functions, short term efficiency often trumps accuracy; after all no one wants to slow a workflow just to fill a field properly.
• Useful feedback loops are not designed-into processes: The people capturing the data rarely see immediate benefit to themselves from it being correct.
CXO distance from the problem and solution compounds the challenge: While enormous energy and resources are poured into building infrastructure — cloud migrations, data lakes, shiny visualization tools — the data is poor and that hardly receives any attention.
Fixing it needs organizational discipline and governance, not glamour or bold investments that can be announced to fanfare.

The solution needs a function many firms forgot to create Manufacturing long ago largely solved an analogous problem. Most companies have a quality assurance (QA) function that independently assures both the output and more importantly the reliability of the processes that generate the output. Some form of a QA function works with operations to fix systemic issues and drives up process capability. The importance is understandably higher in industries like Pharma where quality cannot be “measured into” the product. Yet when it comes to data there is no such function in most companies that assures reliability of the “data product”. The assumption is that data quality will somehow emerge from good intentions and decent tooling.

Imagine, instead, a Data Quality Assurance (DQA) function with a clear mandate:
– Independently assess data integrity through sampling and audits.
– Work hand-in-hand with functions to trace issues back to their root causes.
– Define and publish data quality metrics as business KPIs, not IT footnotes.
– Build simple feedback loops so data creators actually see how their work fuels decisions.

Additionally, going even more upstream of the problem, the role of the DQA function will be to ensure that automated data checks are built into the design of any process and even the installation of new machines itself, so that data quality isn’t inspected in, but engineered in before any process or system “goes live”.

Organizationally to have clout and accountability, the function could report into the underlying function (operations or business head) and act as a necessary member for deployment of any product, process or system.
It is far from a radical idea, but a very practical and underutilized one. Regulated industries — financial services, pharmaceuticals, medical devices — already do this in some form or fashion because regulators demand it. But for everyone else, the discipline is optional. And so, predictably, it’s absent.

The Way Forward
Our collective obsession with AI has, ironically, distracted us from its most basic enabler: clean, reliable data. Data quality won’t fix itself. Nor will it be rescued by another platform migration or AI pilot. What’s needed is a shift in mindset: from treating data as exhaust from operations to treating it as a product or by-product of operations — one that needs standards, stewardship, systematic QA and operational “engineering”.

This cause can benefit from a data quality assurance function – not data governance policies, but an operational function that sits at the intersection of process design, behavioral incentives, and governance: the glue that makes everything else — AI, analytics, digital transformation — work.

The author is Managing Director and Senior Partner, Boston Consulting Group. Views are personal.