What the AI companies figured out

The companies that build the AI models — the ones with the best software on earth and every reason to want you to just log in and use it — have quietly concluded that handing you a login does not work. So they send a person instead.

The role has an ugly name: forward deployed engineer. Palantir invented it for its early government accounts, on a simple premise — an engineer sits inside the customer's business, learns how that business actually runs, and builds against its real data rather than shipping a product and hoping. It stayed a Palantir curiosity for the better part of two decades.

Then the AI wave hit and everyone copied it. OpenAI, Anthropic, Google, Databricks and Cohere all run the function today; Anthropic keeps it under its Applied AI team, and OpenAI stood up a roughly $4 billion entity built specifically around these engineers. Job postings for the role grew more than 800% in nine months.

Read what that actually admits. A company whose entire business model is software-at-scale does not staff thousands of on-site engineers for fun. It does it because it measured what happens without them.

Why buying the tool is not the same as getting the result

The number everyone quotes comes from an MIT report published in August 2025: of the enterprise AI pilots it looked at, 95% showed no measurable financial return. It drew on 52 executive interviews, surveys of 153 leaders and 300 public deployments. The figure has been argued with since, and it is worth arguing with — but nobody has seriously claimed the real number is small.

The interesting half is not the 95%. It is what the other 5% did differently, because it was not budget and it was not better models. They built for one specific process instead of "the business". They judged the thing by a business outcome instead of a demo. And they insisted it connect to the systems the work already happens in.

All three of those are integration decisions, and none of them survive contact with a free trial. A trial account gives you the model. What it cannot give you is the half-day where somebody works out which of your processes is worth automating first, and wires it to the place your orders actually live.

The failure is rarely the AI. Pilots die on brittle connections, missing context and work that never reached the tool people already have open. That is plumbing, and plumbing is done by a person on a specific afternoon.

What the expert actually does

Strip the job title away and the work is unglamorous, which is exactly why it gets skipped.

Learns the business the way a new employee would — from your site, your past conversations and whatever documents you already wrote. Not from a form you fill in.
Decides what the AI should not touch. The boundary is worth more than the capability. An AI that answers everything is the one that eventually answers something wrong.
Connects it to where the work already is — the WhatsApp number customers already message, the store, the calendar, the CRM. Nothing that needs your team to open a new tab.
Tests it on your real questions, not sample ones. The gap between a demo and your Tuesday is the entire project.
Comes back and fixes what broke. First contact with real customers changes the answers. Someone has to be there for the second week.

Notice that only one of those five is technical. The rest is judgement about a specific business — which is why it cannot be shipped in a product, and why the companies with the best products send people.

Why a small business needs this more

The obvious read is that embedded engineers are an enterprise luxury: big company, big budget, big deployment. The obvious read is backwards.

A bank running a failed AI pilot has an internal team that noticed, a budget line that absorbed it, and next quarter to try again. A business with six people has an owner who spent three evenings on it, got a mediocre result, and now believes AI does not work for their industry. The same failure costs far more, because it is usually the only attempt.

And the thing standing between a small business and a working setup is never ambition. It is that the person who understands the business — the owner — is the same person serving customers all day. There is no spare month. Enterprises solve that by assigning someone. Everyone else has been told to watch tutorials.

The self-service trap. Self-service is not cheaper if it does not finish. A tool you set up halfway costs its subscription plus every enquiry it half-answered.

What an integration session contains

It is a working session with a screen shared, not a presentation. The useful version has a fixed slot in a calendar and produces something that exists when it ends.

The business gets learned — from the website, from real past conversations, from the documents you already have. You answer questions; you do not fill in a wizard.
The channels get connected — WhatsApp, Instagram, Facebook. Where your customers already write to you, not a new inbox for them to discover.
The systems get connected — store, calendar, CRM. This is the step most people skip and the step the 5% did not.
The limits get set with you in the room. What it answers alone, what it hands to a human, what it never touches.
It gets tested on your real questions — the ones from last week, including the awkward ones.

What makes it work is the fixed slot. Open-ended help that everyone means to get round to is the thing that never happens; an hour in the calendar with a person waiting is the thing that does.

Five questions to ask before you book one

These work against anyone offering this, including us. If a question gets a vague answer, that is the answer.

What is connected by the end of the session? Name the channels and the systems. "We'll explore your needs" is a meeting about a meeting.
Who does the work — you or me? If the output is a list of things for you to go and do, it was a consultation with a different name on it.
What do I keep if I stop paying? The connections, the content it learned, the account. Know this before, not after.
What happens in week two? The first real customers always change the answers. Ask who fixes that, and whether it costs extra.
What will it not do? Anyone who cannot answer this quickly has not deployed enough of them to know.

One more, for yourself. Pick the single job you would hand over first, before the session starts. Owners who arrive with one specific chore get something working; owners who arrive wanting "AI in the business" get a conversation.

Was this helpful?