AI
AI Where It Adds Business Value, Not Where It Is Fashionable
The discipline of building an AI product for enterprise data is mostly the discipline of leaving AI out.
Dave Welensky, Director · Published August 2026 · 4 min read
When we built Ixia Data Quality AI, the tempting version was AI everywhere. The useful version is narrower: AI classifies, matches and summarises, with confidence scores a human reviews. Rules, reconciliation and severity stay deterministic, which is why auditors can follow every finding.
The tempting version and the useful one
When we built Ixia Data Quality AI, the tempting version was AI everywhere: a chat interface over your master data, generated rules, generated fixes, generated confidence. The useful version turned out to be narrower, and much more valuable.
AI earns its place in our platform in three specific jobs. Classifying columns, because a model reading thousands of fields and proposing semantic types with confidence scores saves a steward days, and the confidence score tells the steward exactly where to look. Matching duplicates, because fuzzy and semantic similarity finds the record pairs that exact matching never will. And summarising runs, because turning 20,644 findings into a prioritised fix plan is pattern work a model does quickly and a human can verify cheaply.
What stays deterministic, and why
Everywhere else, we kept determinism. Rules execute exactly as the business wrote them. Reconciliation is arithmetic, not inference. Severity is assigned by policy. An auditor can follow every finding from rule to record, which is the entire point of a data quality control.
The test we apply is simple: if the AI is wrong here, does a human catch it before it costs money? Column classification, duplicate candidates and summaries all pass, because a person reviews the output with the confidence scores in view. Silent automated correction of financial data does not pass, so the platform does not do it.
AI applied where it adds real business value is not a slogan. It is a scoping decision, made line by line, and it is why enterprise teams can put the output in front of their auditors.
Key takeaways
- AI carries load in classification, matching and summarisation. Humans review, guided by confidence scores.
- Rules, reconciliation and severity stay deterministic and auditable.
- The scoping test: if the AI is wrong here, is it caught before it costs money?
Recognise your programme in this?
Written by the people who deliver the work. The conversation works the same way.