Stop Treating AI as a Tool — It's a Digital Employee That Needs Onboarding

Most companies adopt AI the way you'd buy a software license — then wonder why nothing changes. The reframe that stuck with me from a recent listen: a large language model isn't a tool you operate, it's a digital employee you hire — and like any new hire, it produces nothing until someone onboards it.

I picked this up from an episode of the Mandarin-language podcast 《十字路口》 (Crossroads), featuring two ex-BCG consultants — A-Gan and Liu Kai — who now run an AI consultancy called Rolling AI. They spent an hour on a concept getting a lot of attention inside OpenAI and Anthropic: the Forward-Deployed Engineer (FDE). These are my notes and where my own thinking landed after.

A quick note on sourcing: the case studies and figures below are recounted by the guests on the podcast and I haven't independently verified them. Treat them as illustrative claims, not audited results.

The Reframe: AI Is a Digital Employee, Not Software

Here's the mental model shift. When you buy a SaaS product, it's a tool — a human learns to operate it. A frontier LLM is closer to digital labor. You don't say you "bought" an employee; you say you "hired" one. And a new hire needs onboarding before they're useful.

That's the job of the Forward-Deployed Engineer. The term originated at Palantir, where engineers embed directly with a customer to make the software actually work against messy real-world data — and it's the same idea OpenAI and Anthropic now borrow for their enterprise teams (Anthropic labels its version "Applied AI Engineer"). An FDE doesn't sit in the back office writing code in isolation; they go to the front line of a business, help the AI understand the company's context, and map the actual workflows so the model can start doing real work.

The way I'd put it: an FDE is less "engineer" and more HR business partner for your AI — the person who onboards a brilliant new colleague who happens to be a model.

My takeaway: This is exactly why so many AI rollouts feel useless. Companies drop a model into the org and walk away — like hiring a genius graduate, handing them no data, explaining none of the rules, and being shocked when they deliver nothing.

Consulting's Disruption: From Selling Decks to Shipping Agents

The part I enjoyed most was the jab at traditional consulting. The old McKinsey/BCG model: months of research, then a polished 200-slide deck telling you what your strategy should be.

Firms like Rolling AI do the same upfront research — then skip the deck entirely and ship a working AI agent, with a rule that it goes live within 15 days. They call this model "Service as a Software": instead of top-down strategy, you extract the street smarts of your best frontline employees and replicate that judgment into an agent the whole company can use.

The deliverable changed from a recommendation to a running system. That's the whole disruption in one sentence.

Three Case Studies That Reframed What AI Is For

The guests shared three deployments that broke my mental image of AI as a thing that "just writes copy and makes images":

Cutting a service interaction from ~16 yuan to ~0.1 yuan. A dairy company wanted to sell premium supplements, which required professional nutritionists — but there aren't enough nutritionists in China to staff it. They fine-tuned an "AI nutritionist" that served 6 million users. The clever part: its first move wasn't a recommendation, it was emotional value — asking the customer, "You're not even overweight, why do you want to lose weight?" That landed far better than a hard pitch.
Freeing the property manager buried in busywork. On a rental platform, human building managers spent all day on complaints — leaks, noisy neighbors — with no time to sell high-margin services. Letting AI absorb the routine tickets freed the humans to do the one thing humans do best: warmth and genuine care. Revenue went up.
The AI assistant store manager that breaks the rules. For a retail chain, HQ's inventory forecasts are usually wrong — headquarters doesn't know a new supermarket opened next to one store yesterday, or that tomorrow brings a downpour. Pairing each store with an "AI assistant manager" that folds in the store manager's local knowledge reportedly saved the company millions a year in waste.

Why AI Projects Fail: Three Traps

I wrote down their "three traps of AI transformation" because they read like hard-won advice for anyone pushing AI into an organization:

The CEO expects a superhero. Treating AI as omnipotent — assuming it solves everything the moment it's switched on.
Letting IT own the project. The big one. AI adoption has to be led by the business team, not IT. IT understands the technology but not how to pick products or talk to customers. AI without a business context is just an expensive toy.
Not changing the incentives. This one rings true. If you're a top salesperson and the company asks you to teach the AI your playbook, of course you'll resist — you're training your replacement. Leaders have to design a new profit-sharing model so the people who contribute their expertise earn more by having the AI replicate their ability.

Frequently Asked Questions

What is a Forward-Deployed Engineer (FDE)?

A Forward-Deployed Engineer is someone who embeds directly inside a customer's business to make an AI system work against that company's real data and workflows. The role originated at Palantir and is now used by AI labs like OpenAI and Anthropic. Think of it as the person who onboards an AI the way HR onboards a new employee — supplying context, mapping processes, and getting the model productive.

What does "Service as a Software" mean?

It's a consulting model where the deliverable is a working AI agent rather than a strategy report. The firm does the usual upfront research but, instead of handing over a slide deck, ships a deployed system — often within a fixed window like 15 days — that encodes the judgment of a company's best employees.

Why do most enterprise AI projects fail?

Three recurring reasons: leadership expects AI to be an instant superhero; the project is owned by IT instead of the business team that understands the real workflows; and incentives aren't restructured, so the experts whose knowledge the AI needs have every reason to withhold it.

Who Survives the AI Era

The question I couldn't stop chewing on: in an AI era, who actually gets displaced?

The guests were blunt. The most exposed are pure information relays — the middle layer that just passes messages along — and people who compete on execution alone: typing fast, good memory, following the script. We have to move from being knowledge workers to knowledge decision-makers.

The valuable employee of the future needs strong commercial judgment and a genuine curiosity for solving problems. And — the part AI can't replicate — the ability to deliver emotional value: making a customer feel happy, respected, understood.

AI is a new electricity. It's rapidly taking over repetitive cognitive labor, and you can't stop the grid from spreading. So the move is to stop resisting and start learning to wire the building — and eventually, to build the factory that runs on it.

If business process change with AI interests you, the full 《十字路口》 episode is worth the hour. I'd love to hear where you land — what's the most useful AI deployment you've actually seen in the wild?