Most AI customer service projects fail for a reason that has nothing to do with the technology. The company deploys before it has decided what good looks like: which outcomes it wants, what experience standard it holds, who is accountable. The AI is handed an empty strategy and optimises into it.
When an AI deployment disappoints, the review usually looks at the model, the vendor, the integration. Those are rarely the cause. The cause was set months earlier, in what the company did not decide before it started.
This article is about the strategy gap behind AI adoption: why deploying first and steering later fails, and what has to be decided before launch.
AI optimises whatever you point it at
An AI system is very good at pursuing a goal. If you give it a clear goal, it pursues that. If you give it no goal, it does not wait. It optimises for whatever measurable thing is in front of it, which is usually the easy operational metric: contacts closed, handle time, cost per contact.
Mark Levy, who writes Decoding Customer Experience, makes this point sharply: AI scales whatever you aim it at. Aim it at cost per contact and it will get very good at driving that number down, including in ways that quietly damage the customer relationship. The technology is not the problem. The aim is.
Deploying before deciding is the actual failure
Blake Morgan, in her customer experience newsletter, frames the gap as a strategy problem. Her piece on the CX strategy gap behind AI adoption describes companies buying and deploying AI while the questions of business outcome, experience standard, and accountability stay unanswered. The deployment goes ahead because the technology is available and the budget is there, not because the company knows what it wants from it.
So the AI launches into a vacuum. It picks up the nearest metric and optimises hard. By the time anyone notices the customer experience drifting, the system has months of momentum pointed at the wrong target. Running a change-management programme after launch does not fix this, because the steering was missing from the start.
What has to exist before launch
Mark Levy, in a companion piece on what CX leaders need to fix before AI can deliver, makes the sequencing argument directly. Three things have to be settled before deployment, not after.
First, the outcome. Not "deploy AI", but the specific result you want: faster resolution, fewer repeat contacts, protected retention in a segment. The outcome is what the AI will be aimed at, so it has to be named. Second, the experience standard. What a good AI interaction looks and feels like for your customer, written down clearly enough that you can tell when the AI is below it. Third, accountability. One named person who owns the AI's customer outcomes and has the authority to change its target when it drifts.
The fix is sequence, not effort
This is not an argument for more governance meetings. It is an argument about order. The same decisions, made before launch, steer the AI from day one. Made after launch, they are a correction fighting months of momentum.
An AI customer service project is largely decided in the weeks before launch, in whether the company settled what it wanted the AI to do. The technology cannot supply an aim it was never given.