Brandtech's Rebecca Sykes: AI Does Not Need Perfect Data

By Priya Raman

Conventional wisdom holds that clean data must come before artificial intelligence. Rebecca Sykes, a partner at The Brandtech Group who leads its emerging technology practice, is not so sure. She says generative AI can work with the kind of fuzzy, siloed and sometimes badly labeled data that most marketers actually hold.

That single claim reframes one of the most common barriers cited in enterprise AI adoption. For years, practitioners have been told to fix their data foundations first: unify identifiers, standardize taxonomies, complete labeling. Only then, the argument ran, could models deliver trustworthy output. Sykes turns that sequencing on its head.

Her view comes from inside an organization that builds AI tools for large advertisers. The Brandtech Group operates an emerging technology practice, and Sykes leads it. That position puts her close to real deployments rather than laboratory conditions — and to the messy data estates that global brands bring to those deployments.

What does this change for marketing teams?

If Sykes is right, the practical consequence is direct. Teams waiting on a full data-quality program before piloting generative AI can start earlier. Siloed data, imperfect labels and inconsistent structure stop being blocking issues and become constraints the technology itself can absorb.

This matters because generative models behave differently from earlier analytical systems. Classical machine learning pipelines degrade sharply when input data is dirty or fragmented. Generative systems, by contrast, can often produce useful output from partial, ambiguous or poorly organized source material — the everyday condition of most marketing data.

The distinction changes budget conversations too. Investment that was earmarked for multi-year data remediation can be redirected toward use cases, governance and measurement, provided the outputs are checked against the business question at hand.

Why frogs and socks?

The original discussion, framed around lessons from frogs and socks, uses everyday analogies to make the implementation argument concrete. The underlying point survives the shorthand: AI implementation succeeds through experimentation with real, imperfect inputs rather than through extended preparation for idealized ones.

For practitioners, the guidance is to test against the data you have, not the data you wish you had. Where output quality falls short, that gap itself becomes a precise map of which data fixes actually carry business value.

The so-what

Sykes's position does not argue that data quality is irrelevant. It argues that quality work should follow, and be justified by, live AI use cases. Marketers who treat imperfect data as a starting condition rather than a disqualifier will learn faster which parts of their estate genuinely need repair.

The open question for buyers is whether their vendors can demonstrate results on imperfect data, not on sanitized demos. That is now the question worth asking in every pitch.