AI Labs Are Spending Billions on Deployment, Not Just Models

If your team already has AI licenses and still isn't getting much out of them, the fix the biggest AI companies are actually funding this year isn't a smarter model. It's someone sitting down with your team and teaching them what to ask it to do.

Five moves, one admission

In April, NEC deployed Claude to roughly 30,000 employees as Anthropic's first Japan-based global partner. The part that mattered wasn't the seat count — it was the dedicated Center of Excellence and structured training built into the rollout before a single employee logged in.

In May, OpenAI raised over $4B for a "Deployment Company" that sends engineers directly into client organizations to redesign workflows around AI, not to sell more subscriptions. The same week, Microsoft's own research put a number on why: organizational conditions have twice the impact on AI outcomes that individual skill does, and only 19% of workers were operating in conditions set up for AI to actually help.

That same month, OpenAI's own usage data showed frontier firms using 3.5x more AI per worker than typical firms — not because they prompt more often, but because they've built multi-step workflows instead of one-off questions.

In July, Dario Amodei publicly walked back his "ten years per year" prediction and named three specific blockers. One of them: researchers need time to learn the tools before the tools change anything.

And Shopify's CEO put the company's internal coding agent in public Slack channels only, never DMs — a deliberate choice to make watching a colleague's actual prompts and revisions the training program, instead of scheduling one.

None of these five moves were coordinated. Different companies, different mechanisms, same conclusion stated five separate ways: the constraint was never who has API access. It's whether the person with the task knows how to turn "I have a thing to do" into a scoped request with a checkpoint in it.

What actually closes the gap

You don't need a Center of Excellence or a Deployment Company to do the version of this that works at your scale:

Name the first task in specific terms. "Use AI for writing" doesn't survive contact with a real week. "Draft the Monday risk summary you currently write by hand from the same three sources" does.

Decide the checkpoint before you start, not after the output arrives. Which step in the task is the one where your judgment is non-negotiable — and where does AI just get you to that step faster?

Make the exchange visible to at least one other person. A prompt that worked, sitting in a private chat, teaches exactly one person. The same prompt pasted into a shared channel teaches everyone who reads it.

Write what "done well" looks like in one sentence before you run the task. This is the part the Wharton research behind Amodei's own comments backs up: prompting tricks — personas, tipping, chain-of-thought instructions — returned no measurable gain. Clear goals and named acceptance criteria did.

A training coordinator at a 70-person accounting firm

She'd had a company-wide AI license live for eight months. The usage dashboard showed logins holding steady and nothing else — no growth in how people used it, no shift in what they used it for. Rather than run another all-hands demo, she booked one 20-minute conversation with each of six team leads: name the two most repetitive tasks that eat more than half an hour a week, pick one, and build a first request for exactly that task together, using their actual inputs. By the end of the quarter, five of the six teams had at least one workflow still running unprompted. The license hadn't changed. The gap between having it and using it had.

A sales director at a 45-person agency

He tried the Shopify approach at a much smaller scale: a Slack channel called #ai-drafts, with one rule — any AI-assisted proposal draft gets posted with the prompt that produced it before it goes anywhere near a client. He expected resistance. What he got instead was newer account managers picking up the senior team's framing habits within weeks, without a single training session, because the actual exchanges were sitting there searchable. The one caution he'd pass on: he made the channel about what taught someone something, not a showcase of polished work — a review-culture version of the same idea would have shut it down in a month.

The one thing

Every company in this piece could have spent its money making the model better instead. They spent it making the user better. That's the tell worth acting on before the next consultant shows up to tell you the same thing.