Your AI doesn't know your customers and that's a problem for growth

When a traveller fills in a survey after a flight, the response typically lands in a customer experience platform that counts how often certain words appear and which ones cluster together. Any employee seeking to understand what the customer meant exports the results to a spreadsheet and pastes them into ChatGPT, taking customer data outside the organisation. This is what we’ve accepted as the state of the art in data science for the AI era.

It isn’t just anecdotal either: a recent PagerDuty survey of office workers in Australia and three other markets found 40% of respondents upload customer data to public AI tools. Meanwhile, PwC found that 74% of real financial returns on AI investments in Australia are only captured by a mere 20% of companies. The emerging correlation is those who invest in a strong AI-native architecture see better returns. And yes, you would be correct to assume that those who offload customer data to external tools are excluded from that cohort.

Centralised customer intelligence forms an important part of this architecture; it is the foundational layer of growth in the AI era, but many mishandle it. From my observations, there are three factors poking holes in enterprise AI strategies and they all tie back to how they collect and analyse customer data, organise around it, and use it to decide what to build.

Understanding the what but not the why

Let’s be honest, enterprises are drowning in unstructured customer data. Support tickets, call recordings, NPS and CSAT verbatims, app store reviews and sales calls: all of it arrives in volumes that keyword counts and co-occurrence matrices were not designed to handle.

Those techniques are good at identifying keyword frequencies, but they miss the important context behind it. What did the customer actually say and why did they say it? While large language models can process this material at a scale no analyst team could match, it is only useful if the model understands the context. So, a summary generated from an exported CSV may seem like a quick win, but it ultimately gives teams an incomplete picture for important product decisions — and operating on vibes is never good business practice.

Different teams, different views on reality

Even good analysis tends to stay with the team that commissioned it and that’s because product, sales, marketing and support in most large organisations use different systems. They are also likely working from different continents, with go-to-market in North America and product or support in Sydney or India, for example.

Disparate tools and geographies tend to create organisational siloes, with CX platforms holding survey data but little revenue information, while sales and marketing tools contain call and pipeline data. The result is a business that has four partial views of the same customer and no direction on what to do to retain them or grow that segment.

What’s more is that AI will inherit these blind spots because it can only reason with the data it is given. Now imagine you are an enterprise deploying fleets of AI agents at scale to autonomously execute tasks for different cross-functional teams, but they all carry a fragment of the broader context. Would you trust their outputs? I would fathom a guess and say no.

When anyone can build quickly, deciding what to build is crucial

February’s sell-off in software stocks, dubbed the SaaSpocalypse, rested on the idea that companies could vibe code the tools they license. That fear was grounded in the reality that AI-assisted development lets a capable team build almost anything far faster than before.

When every competitor has that speed, the differentiator moves to deciding what to build and for whom, and without centralised customer insights behind that decision, AI is mostly going to help you ship well-engineered software that nobody asked for.

Many businesses have moved quickly to put AI to work, but a modern enterprise chasing after the promise of ROI without first laying a strong customer intelligence foundation is akin to building an empire on sand. It is doomed to lose market share to competitors who have invested in the plumbing to steer AI in the right direction.

The next round of growth will go to those that stop treating the post-flight survey as a keyword count and start treating what customers tell them as the brief for everything they build.

*Doug Rathbone is VP Engineering at Dovetail.

Image credit: iStock.com/rawintanpin