The Four Gaps — LinkedIn Post
We were demoing Typeface for a customer recently. Halfway through, their CMO stopped us. "The generation quality looks great. But how does this fit into our approval workflow? How does it learn our brand? And what happens to content performance after it's published?"
Those three questions had nothing to do with output quality. And that's exactly the point.
AI content generation is good enough. There's more work to be done, but that's not where the focus needs to be. After 18 months of building enterprise AI systems, watching what holds and what breaks in production, I've come to believe the industry is still solving yesterday's problem. The models have crossed the threshold (other than video). The harder challenge is everything built around them.
What enterprises keep running into isn't a generation problem. It's an infrastructure problem. And like every enterprise transformation, it comes down to people, processes, and technology. Brand guidelines that live in someone's head. Approval workflows still happening in Slack. Performance data sitting in a system that never talks to the content tool. AI that can produce content faster than any organization can actually absorb it.
The technology, it turns out, is the easiest part.
The four gaps I see in nearly every enterprise deployment:
→ Orchestration. Moving content through approvals, localization, versioning, and distribution at scale. Most teams are still doing this manually, which means the AI has made them faster at the easy part while the real bottleneck stays invisible.
→ Knowledge integration. A generic LLM doesn't know your customer, your brand, your regulatory constraints. Without structured context, you get generic content. And generic content, produced at enterprise scale, is expensive noise.
→ Quality evaluation. Subjective feedback doesn't scale. What you actually need are evaluation pipelines with consistent dimensions and repeatability. Most organizations are still running on intuition.
→ Closed-loop measurement. Performance data should flow back and inform what gets created next. For most teams, creation and analytics live in completely separate systems. The loop never closes.
And here's what most teams get wrong when trying to fix this: they iterate on the prompt. They refine the instructions, add more examples, run another cycle. That gets you marginal gains. The real quality leap comes from what feeds into the system: structured context, clean data, well-defined personas. The leverage is upstream, not in the prompt box.
These are systems problems, not model problems. And the companies that win the next phase of enterprise AI won't win on generation quality. They'll win by building the infrastructure layer that handles orchestration, context, quality, and measurement at scale. What I think of as the Content Operating System 🏗️
An instrument doesn't make a symphony. The question worth asking any AI content vendor: are you a violin, or an orchestra?