Everyone got the same efficiency gains. Where's your advantage?

The last two years of enterprise AI can be summarised in one sentence: everybody bought efficiency. Copilots for the teams, deflection bots for the contact centre, generation tools for the content pipeline. It broadly worked. Costs came down, cycle times shortened, and the board pack got its AI slide.

The uncomfortable part is that your competitors bought the same tools, from the same vendors, at roughly the same price, in the same quarter. An efficiency gain that everyone gets at once is not an advantage. It's a new baseline. The savings get competed away, into price and into margin expectations, and strategically nobody has moved.

The deployment data shows the ceiling on this approach. A Sinch survey found that 75% of enterprises have rolled back or shut down customer-facing AI agents after launch. That's what happens when AI is bought as a cost line rather than designed as an experience. The agent deflects, and the customer defects.

What the movers are doing instead

None of the interesting activity right now is on the cost side.

Kroger just made a conversational assistant the default digital front door across all of its banners. It builds carts from budgets, dietary needs, and photos of handwritten shopping lists. Kohl's has grown a narrow seasonal gift finder into a persistent shopping assistant that compares products, surfaces deals and tracks orders. Sierra has wired Plaid into its platform so customer-facing agents can complete transactions mid-conversation rather than just answering questions about them.

These are revenue-side moves. Each one takes a specific, high-value customer interaction, whether that's filling a basket, finding a gift or finishing a transaction, and rebuilds it on an AI-native model alongside the existing experience, with a commercial number attached. Not a transformation programme. A migration, one interaction at a time.

Where advantage actually lives now

If the tools are symmetric, advantage has to come from what can't be procured. In AI-native CX that means two things.

The first is judgment about sequence: knowing which interaction to migrate first, for which customers, and what it's commercially worth. That knowledge doesn't come from a vendor roadmap. It comes from behavioural data, from what customers actually do rather than what they tell a survey they would do. The gap between stated and revealed preference is where most AI CX programmes quietly die. Customers say they want a human, then choose the assistant that builds the basket in forty seconds. Design for the behaviour and validate on the behaviour.

The second is taste, and the market has just put a price on it. The creators of DesignArena, where millions of people vote on which AI-generated designs are better, raised $7.9M from Index Ventures to sell aggregated human design judgment to the frontier labs. The companies building the models are paying for human taste, because generation is abundant and knowing what's good is not. If judgment is the scarce input for the labs, it's certainly the scarce input for your CX roadmap.

The better question

The planning question that dominated the last two years, "where can AI cut cost?", has been asked by everyone and answered the same way for everyone. As a source of advantage, it's exhausted.

The question that isn't: which customer interaction, rebuilt AI-native, would your customers choose more often, and what is that choice worth?

Answer it one interaction at a time. Pick the interaction where volume and value are provably concentrated. Rebuild it alongside what already exists. Attach a commercial number before you start, and measure revealed preference rather than satisfaction scores. Then move to the next one.

Everyone got the same efficiency gains. Advantage is what you design next.