"We're looking for a really strong AI engineer."
That's the first line of almost every call we get. So we ask the follow-up: what do you need this person to actually do? The answer is always the same list. Fine-tune the model. Stand up the vector store. Write the backend API. Wire the retrieval pipeline. Own the infra. Build the evals. Handle the on-call. Ideally talk to customers. Senior enough to make architecture calls, cheap enough to fit the budget, available enough to start Monday.
Then we say the thing that changes the conversation: that isn't one job. It's five, and the person who spans all of them at once doesn't exist. You don't have an AI hiring problem. You have a team-composition problem — and you can't fix a team-composition problem with a headcount.
"But we've seen people who can do most of that."
Sometimes you have. When you genuinely find someone who spans three of those roles well, they're expensive, they're gone in a year, and they're a single point of failure wearing a hoodie. The day they take PTO, your AI roadmap takes PTO too.
And the part that surprises people: even that person, hired, is slow. Shipping AI in production isn't a skill you hire for. It's a pipeline. Somebody prototypes, somebody hardens it, somebody makes it observable, somebody catches the hallucination before your customer does. One person doing all of that serially is a bottleneck with a great resume. We've watched this end the same way a dozen times — you post the req, you interview for three months, and either you don't fill it or you hire someone great and they quietly stall six weeks in, because the model was never the hard part. The infra was. The data was. The eval harness was. The thing nobody owned was.
"So what does it actually take?"
The roster for putting one genuinely useful AI feature in front of real users and keeping it alive:
- An AI/ML engineer who understands models, prompts, retrieval, and where they break.
That's not seven hires. It's the minimum viable set of disciplines for one production system. A strong generalist can cover two of these well, so you can compress it — but you cannot compress it to one. Every team we've watched try to skip a role didn't remove the work; they moved it downstream to where it costs 10x to fix.
"Fine — then why not just hire the whole team?"
You can. It takes nine months and it might not gel. That's the reason Black Gibbon is a pod and not a staffing agency handing you résumés.
We rent you a small, cross-functional team that already has the mix: AI and ML engineers, backend, DevOps, data, SRE, and human QA, with senior review baked in and one lead who's accountable for the outcome. They've shipped together. They have a shared toolchain, shared standards, and a shared Slack history of every failure mode they've already hit — so they don't rediscover them on your dime. You're not buying labor by the seat. You're buying a functioning team on day one, with no integration tax between the person who built it and the person who's supposed to run it, because they sit in the same pod.
There's a speed dimension too. Our bench spans Irvine and Vietnam, so work moves across a near-24-hour cycle: something gets specced in California, and there's real progress on it by the time the SoCal team logs back on. That's the difference between a two-week loop and a five-day one.
"A pod costs more than one FTE."
Per month, sometimes. Per shipped outcome, it's cheaper, and it starts today instead of after a two-quarter search plus ramp. You're also comparing against the wrong thing — one FTE can't ship this, so the real alternative is four or five FTEs plus a hiring pipeline plus the risk they don't gel. The pod is the cheaper version of the thing that actually works.
"I'll lose control."
You lose control when knowledge lives in one contractor's head and walks out the door. A pod works in your repo, your stack, your standards, with senior review you can see. You get more visibility, not less — and no key-person risk.
"Then why not a big agency, or an offshore shop?"
From a big agency you get a slide deck, a rotating cast, and a markup on juniors you never interviewed. A pod is small, stable, and named — you know who's on it. From a body shop you get hours billed, not systems shipped: no senior review, no QA, no accountability for whether the thing works in production. We put humans in the loop precisely because "it ran" and "it's right" are not the same claim, and someone has to own the difference.
Where the conversation lands
The companies pulling ahead on AI right now aren't the ones who found the mythical do-everything engineer. They stopped looking, admitted the work is cross-functional, and put an assembled team on it.
If your AI roadmap has been stuck behind an open req for a quarter, that req is the problem. You don't need one more heroic hire. You need the right five disciplines pointed at your problem, accountable as one team, shipping this month. That's the pod. That's what we do.