The Real Bottleneck of AI Implementation Was Never the Model

Deep thoughts on AI and aspirations —— ByteDance Deep Thinking Circle

Over the past two years, everyone’s been fixated on models: whose model is stronger, who released a new version, whose benchmark scores are higher. But an increasingly obvious fact is this: whether AI actually lands in enterprises almost never bottlenecks on the model itself.

Models are already strong enough. The real problem is—once it enters an enterprise, who will actually make it work? This layer currently has no one doing it systematically, but the value may be even greater than the model itself.

The Last Mile Is an Undervalued Opportunity

Selling AI to enterprises was never the hardest part. The hardest part comes after the sale: the model is installed, but it doesn’t connect with the actual business operations, won’t run, has no one maintaining it, and ultimately becomes a showpiece.

Where’s the problem? It’s the lack of someone who “understands tech, understands business, and understands delivery” who can connect these three things. Engineers understand tech but not client business, business people understand business but not tech, delivery folks touch neither side. So AI enters the enterprise and gets stuck at this final step.

The value of this “last mile” is being rediscovered. Some leading AI companies have already started creating this specific role—having people who understand technology walk into client sites and actually integrate AI into business operations to make it run.

Why This Role Is a Moat, Not Outsourcing

What makes this role valuable isn’t “getting the job done,” but distilling what’s learned in the process.

If you only get the client’s job done, that’s essentially human services—you charge once per job, and once people leave, the value disappears. But if during this process you distill “what clients in this industry need, how to deliver, how to replicate” into methods and products, then the value becomes reusable—next time you work with similar clients, you don’t start from zero.

This is the difference between a moat and outsourcing: one turns knowledge into assets, the other uses knowledge and discards it. The same work, done differently, yields entirely different orders of magnitude in value.

But There’s a Critical Pitfall to Avoid on This Path

The biggest risk with this role is gradually becoming high-end outsourcing.

If a team only does delivery without distillation, going to client sites each time to re-interview, re-map processes, starting from scratch, then no matter how sophisticated they are, they’re just expensive outsourcing. The real test is: after completing each job, how much reusable material does this team have?

So for people doing this work, the validation criterion is straightforward: after finishing each client, are the methods, templates, and cases you’ve distilled growing, or still at zero?

Judgment for Entrepreneurs

My judgment is: this last mile of AI implementation is one of the most worthwhile opportunities right now.

But it’s not an easy business—it requires teams to simultaneously understand tech, business, and delivery, plus the ability to productize experience. Done right, it’s a moat. Done wrong, it’s outsourcing.

The key question is one: are you helping clients get the work done, or are you making “helping clients get the work done well” itself into a reusable business? These two answers determine the value ceiling of this business.

Key points: AI implementation bottleneck is at the last mile (hybrid role understanding tech + business + delivery); value lies in distilling into assets, not just completing jobs; to avoid becoming outsourcing, look at how much reusable material each deal generates.