If Your AI Isn't Working, Fix The Foundation
Merlin Bise is the CTO of Inbenta. He has built conversational AI since before generative AI.
gettyA lot of the AI deployed into the customer experience over the past three years isn’t performing. That’s just the reality. And the default response I hear, over and over, is to start all over again, evaluating and paying for a new platform.
In most cases, that’s an expensive way to rebuild the same problem.
I’ve been building AI for customer experience since before generative AI showed up. I watched intent matching give way to generative layers, and now we’re into agentic systems that can reason and act on their own. The platforms keep changing. But it isn’t the model that causes the deployment to fail—it’s the knowledge underneath it. I’ve seen that pattern in nearly every underperforming deployment I’ve looked at, going back years.
And look, none of this is a knock on the decision you made three years ago. You made it with the tools that existed then. Those were very different tools.
Here’s what’s new. Fixing the knowledge layer used to be a multiyear, hand-built project. That’s why re-platforming seemed faster, even when it wasn’t. AI now does the heavy lifting up front. It reads, structures and tags your content before anything goes live, so what customers hear is approved, not improvised. Work that used to take a room full of contractors a full year can now get done in weeks, not quarters. If you wrote off your knowledge problem 12 or 18 months ago as too slow or too expensive, it deserves a second look and, honestly, a lot more optimism than it would have earned back then.
Before you run another evaluation, check three things. You can diagnose all of them without a new vendor.
Content written for someone browsing a help site does a different job than content a model has to parse. A model can have the right document and still give the wrong answer. Why? Because two records that aren’t the same can look the same to a system reading them. If a customer asks about one product configuration and gets information about another, that isn’t because the company chose the wrong model or the customer botched the prompt. It’s a structure and authorship problem, and no new platform is going to fix it for you.
Count them. Most organizations have more duplication than they think. As mentioned earlier, a lot of platforms started with intent matching, then added a generative layer later and kept the original intent base underneath. So, now you’ve got two copies of your content, two governance surfaces and two possible answers to the same question. Contact center stacks do the same thing by channel—one store for digital and another for voice. The answer changes depending on which door the customer walked through. If a policy lives in more than one place, you have a consistency problem, not an AI problem.
This is the one that kills deployments. If the subject matter expert who spots a wrong answer has to file a ticket and wait on an engineer, corrections will never keep pace with the business. The person with the knowledge can’t act. The person who can act has no way to judge whether the answer was right. Give that six months, and the system has drifted away from how the company is supposed to operate, without a single error in the logs. You don’t get an alert—you get confused customers.
In the end, you should have one knowledge store instead of several, content structured so a machine can find the right answer and the people who own the policies in control of the answers. And there’s no need to replace your platform.
The economics back this approach up. A full platform replacement is a six-figure commitment before you even count internal time, and moving off an entrenched stack can run several times that. Fixing the knowledge layer costs a fraction of it. And you keep the asset, no matter what you run on top. If you want to switch platforms later, fine. You’ll be moving a structured, governed knowledge base, instead of hauling the same mess to a new vendor and paying them to discover it.
There’s a second-order benefit, too. A consolidated, governed knowledge layer gives you an audit trail. When a regulator asks about a specific conversation, you need to show the question, the answer, the source and who approved it. It doesn’t matter what platform was used. If you can’t see why your AI said something, you can’t defend it.
Here’s my takeaway. Before you run another evaluation, figure out whether you have a platform problem or a foundation problem. Most of the time, it’s the foundation. And it’s more fixable today than it was when you signed that original contract.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
