AI Doesn’t Fail Because the Technology Is Wrong

Every enterprise AI project starts the same way. A proof of concept, built on clean historical data, performs beautifully in a sandbox. Confidence builds, budget follows. Then it goes live and quietly underperforms, month after month, until someone finally asks: was this ever going to work?

The instinct is to blame the model: wrong algorithm, wrong vendor, wrong use case. That instinct is almost always misplaced:

• Gartner: 60% of AI projects lacking AI-ready data will be abandoned through 2026
• MIT: 95% of organisations saw no measurable profit and loss return on their generative AI investment, across 300+ enterprise deployments reviewed

An analyst firm and a university lab, with no reason to agree. Same outcome.

So if it isn’t the model, what is it?

The Uncomfortable Answer

Almost always, it’s the data, and how it’s organised.

Most organisations accumulate data the way they accumulate complexity: incrementally, reactively, in silos. CRM holds one view, the contact centre another, billing and analytics a fragment each, none telling the whole story. That collection has a name: a data estate, everything an organisation knows about its customers, scattered across every system it has ever bought.

The question isn’t whether you have a data estate, every organisation does. It’s whether that estate was organised well enough to support a model making live decisions on it. For most, the answer is no: testing data that never reflected real journey complexity, signals that existed but were never captured usably, no shared definition of “success” to tell afterward whether the AI helped or hurt. None of that is a technology failure. It’s an organisation failure dressed up as one, the gap between a demo that impresses and a live system that moves the business.

What Organised Actually Means

Built from scratch to support AI, a data estate would look almost nothing like the average enterprise landscape. It would be structured around the customer’s journey, not internal departments, so every data point carries context: what stage this customer is at, what they’ve done, what they’re trying to reach, how far they have to go. It would treat progress, not profile, as the primary signal, since an agent that can’t tell a pause from a stall will either hover uselessly or go silent exactly when it’s needed. It would be obsessive about outcomes over interaction counts, because a model with no feedback loop cannot learn, it just becomes an expensive rule engine wearing an AI costume. And it would combine behaviour with sentiment, since either alone misleads: a customer moving smoothly but grumbling is different from one stalled but polite about it.

This isn’t an abstract wish list. It’s what Customer Journey Analytics and Orchestration, CJA/O, was built to do: define the goal, structure the journey into stages a model can reason about, capture the signals that separate progress from stagnation, and tie every AI decision back to whether it moved someone closer to the goal or further away. Done seriously, AI becomes an asset. Skipped, it produces exactly the numbers above.

Here’s what should concern anyone already bought in. Gartner’s 2025 Market Guide for CJA/O found 65% of senior marketing leaders have adopted these tools, but use only 43% of the capability inside them. This isn’t about lacking the technology, most already have it. It’s about buying the fix and still not using it.

“The challenge with the application of LLMs in engaging customers is that they are fluent but blind. Without customer context, they can be very confidently wrong, ruining the customer experience.”

Trent Rossini, CEO, inQuba

Where inQuba Fits

This is the argument behind Outcomes CX, the methodology our platform is built around. Most CX tools are still survey machines: ask how the customer felt, plot it on a dashboard, call it experience management. Outcomes CX starts from behaviour and outcome first, tracking customers along their real journey so we can see who’s progressing and who’s stalled. Sentiment is layered on top, not standing in for it.

We have watched this play out with our own clients, not as a hypothetical.

At Sanlam Corporate, the challenge was shifting from an employer-led, B2B model to a direct relationship with individual policyholders, the kind of shift that fails without journey-organised data. Onboarding was restructured around the member’s actual journey.

The result was that inQuba helped Sanlam Corporate increase their customer engagement rate over 500% and member portal registration by 340%.

The product didn’t change. What changed was whether the data and the intervention were organised around the customer’s actual position in their journey.

At Discovery and Cumulate, the challenge was completion rates on financial education content customers kept abandoning. Journey-aware nudges replaced a one-size-fits-all approach.

Course completion rate: 2% → 25%, a twelve-fold difference between doing nothing and intervening at the right moment.

This is also, concretely, what “AI agents engaging customers along the journey” means in practice. Our AI Navigator, released this year, is a conversational agent that nudges and assists customers across WhatsApp, SMS, and email, using their full journey context to hold a coherent conversation rather than fire off a generic message. It is fluent and, unlike an LLM without that context, not blind.

The upside shows up elsewhere too. McKinsey’s Next in Personalization research found companies growing faster derive 40% more revenue from personalisation than slower-growing peers. That gap doesn’t come from a smarter model. It comes from knowing precisely where every customer stands, exactly what moved the needle at Sanlam and Discovery.

The Real Choice

The AI opportunity in customer experience is real: agents that guide customers to their goals instead of just responding to complaints, systems that sense frustration before it becomes churn. It’s achievable, and organisations are already proving it.

But only in that order. Organise the data around the customer’s journey and outcomes first, then let AI loose on it. Reverse it, model first, data as an afterthought, and you’ll get an expensive version of the abandonment rate Gartner is already predicting.

The organisations that win with AI over the next few years won’t be the ones with the most advanced models. They’ll be the ones that did the unglamorous work of organising their data around the customer first.

Trent Rossini will be discussing this further at CEM Africa, 18 to 20 August, Century City, Cape Town. Find inQuba at stand C21 to continue the conversation.