The world's smartest AI companies are quietly admitting something their pitch decks don't: intelligence isn't plug-and-play yet.
The Summary
- OpenAI launched a "Deployment Company" to embed engineers inside customer organizations. Anthropic and Google are hiring similar roles. This is not a coincidence.
- If AI were truly a utility like electricity or cloud computing, you wouldn't need human integration teams sitting in your office to make it work.
- The gap between "abundant intelligence on demand" and reality reveals what actually blocks enterprise AI adoption: not capability, but implementation complexity.
The Signal
OpenAI's Forward Deployed Engineers will "work with business leaders, operators, and frontline teams to identify where AI can make the biggest impact, redesign workflows, and turn those gains into durable systems." That's consultant language, not utility language. When Amazon Web Services scaled, they didn't need to embed engineers in every customer office. They built APIs and documentation. The fact that OpenAI, Anthropic, and Google are all hiring for embedded engineering roles simultaneously tells you where the real friction lives.
The disconnect is structural. Frontier models can write code, analyze data, and generate content at superhuman speed. But they can't navigate your permissions hierarchy, understand why the sales team hates the CRM, or figure out which compliance requirement is actually enforced versus theater. They can't look at a workflow that exists half in Salesforce, half in email, and a quarter in someone's head and redesign it. Humans still do that. Expensive, experienced humans.
"If intelligence were already a true utility, this would not be necessary. You would not need to send your own engineers to every customer to make the faucet work."
This is the 2025 version of enterprise software's oldest problem: the last mile. SAP, Oracle, Salesforce, they all learned this decades ago. Software at scale requires integration partners, consultants, and implementation teams because every company is a unique snowflake of technical debt and political compromise. AI companies thought they'd skip that phase. The models were supposed to be smart enough to adapt themselves.
They're not. Or more precisely, they're smart enough but not *integrated* enough. The intelligence is there. The interface between that intelligence and actual business operations is still duct tape and human judgment. Forward Deployed Engineers are the duct tape. They're translating between what the model can do and what the organization needs done. That work includes:
- Identifying which processes are actually automatable versus which ones just sound automatable in a slide deck
- Redesigning workflows so AI can slot in without breaking everything downstream
- Building the institutional knowledge about model behavior, failure modes, and edge cases that no documentation captures
- Managing stakeholder expectations when the demo magic hits production reality
The companies doing this aren't failing. They're being honest about where the technology actually is. The AI utility metaphor is marketing. The reality is bespoke implementation at enterprise scale. That's not a criticism. It's a description of the current state of the art.
The Implication
If you're betting on AI agents to "just work" in your organization without serious integration effort, you're going to be disappointed. The companies building the best models in the world don't believe that yet. They're hiring armies of people to make it happen manually, one customer at a time. That's your signal about readiness.
For vendors, this is temporary leverage but permanent learning. Every embedded engineer is gathering data about what breaks, what works, and what customers actually need versus what they think they need. That knowledge turns into better products, clearer APIs, and eventually, actual utility-grade AI. But we're not there. And the companies closest to the technology know it.