Germany
The works council comes first
Co-determination over systems able to monitor performance binds harder than the AI Act.
The field guide
The field guide to forward deployed engineering
A job that spread faster than anyone wrote it down, and the practice of getting artificial intelligence to actually change how an organisation works. Written by people who do the deployment, for the engineers considering the role and the companies buying the work.
Written and maintained by Radif Partners
Last reviewed
Every figure names its source and the date it was checked. Our method, and everything we cite.
The premise of this site is that the hard part of artificial intelligence in a company is no longer the model. A system can be demonstrably capable and still produce nothing once it reaches an organisation, and the reasons are almost never technical: data that lives in four systems with three definitions of the same customer, a process that exists in somebody's head rather than in the documentation, a security review that will not approve an outbound call, a team whose bonus depends on the old way of working. Closing that distance is a job, it has a name, and the name spread faster than anyone wrote down what the work involves. That is what these pages are for: the role described honestly for the people considering it, and the practice described honestly for the companies buying it.
The role
What a forward deployed engineer actually is
An engineer paid by the product company, working inside the customer. Where the title came from, why it spread, and the parts of the job that appear in no posting.
Compare
Five roles that get mistaken for each other
AI engineer, solutions engineer, software engineer, solutions architect. Not synonyms. What separates them is five axes, and they sort a job posting faster than its heading does.
For companies
Getting AI past the demo
Most projects do not fail technically. They fail at data access or at adoption, and both are predictable. What a deployment actually involves, and when not to start one.
Worth reading first
Japan
The constraint is people, not law
A promotion-oriented regime, and a shortage of the one profile a deployment cannot do without.
Agents
One answers, the other acts
Same model, same error rate, entirely different consequence. Why reversibility draws the line.
Adoption
Watching people not use a system that works
Correct and unused is the most common outcome, and the reasons are never the ones people give.
By market
The same system is a different project in ten countries, and the difference is almost never the technology. Who must be consulted before people's work changes, whether the data may leave the jurisdiction, which language the output has to be right in, and whether an AI law binds at all.
What the numbers actually say
Generative AI reached 53 % of the population within three years, faster than the personal computer or the internet, and organisational use has followed: the Stanford AI Index records it in at least one business function at 70 % of organisations. Appetite is not the constraint, and it is not distributed the way most people assume. Adoption runs at 64 % of the population in the United Arab Emirates and 61 % in Singapore against 28.3 % in the United States, which ranks twenty-fourth while its companies are among the most active buyers anywhere. That gap is the whole subject in one statistic: where organisations buy far ahead of individual familiarity, the obstacle on a project is never enthusiasm. It is the distance between a decision having been taken and a credential having been issued.
Demand for the skills has moved with it. AI skills are now requested in 2.5 % of all United States job postings, a 297 % rise over a decade. What has not settled is what the jobs are called, which is why a large part of this site is spent separating titles that are used for four different roles.
What this site will not do
It will not publish salary figures. The public numbers for this job title disagree with one another by roughly a factor of four, because aggregators average every posting that shares a heading, and a deployment role at a frontier lab is not a field support role at a mid-market vendor. A single average would rank well and mislead everyone who used it to negotiate. When we maintain our own posting-level data, the figures will appear with the method beside them.
It will not invent a statistic, a client, a testimonial or a job posting, and it will not quote an efficiency percentage for a company nobody has looked at. Where a source has a commercial interest in what it describes, the sources page says so rather than presenting it as neutral.
Sources on this page
- Stanford Institute for Human-Centered AI · The 2026 AI Index Report, Economy chapter
- Stanford Institute for Human-Centered AI · The 2026 AI Index Report
- Wikipedia · Forward Deployed Engineer
Last reviewed . Editorial policy.
Questions people arrive with
Who is this site for?
Two audiences that rarely read the same thing. Engineers deciding whether to take a deployment role, who need the job described including the parts that are unattractive. And companies buying the work, who need to know what a deployment involves before they scope one. The pages are written so that neither group has to wade through the other.
Who writes it, and how is it paid for?
Radif Partners, which also runs deployment engagements. The consulting pays for the site; there are no sponsored posts, no affiliate links and no vendor commission. That arrangement is an obvious conflict, so the test worth applying is whether the editorial pages ever recommend us as the answer. They do not, and several argue against buying what we sell.
Why are there no salary figures anywhere?
Because the public numbers for this job title disagree by roughly a factor of four. Aggregators average every posting that shares a heading, and a deployment role at a frontier lab is not a field support role at a mid-market vendor. A single figure would rank well and mislead anyone who used it to negotiate an offer.
Where should I start if I am considering the role?
Read what a forward deployed engineer is, then the five stages of the job, then the comparison against the role you currently hold. Those three take about twenty minutes together and cover the decision. The interview and routes-in pages are worth reading only once you have decided you want it.
Where should I start if I am buying deployment work?
The ROI calculator first, because it settles in five minutes whether a project is worth having a conversation about. Then the page on how a deployment runs, which describes the five stages and names the two where projects actually die. Neither requires talking to us, and that is deliberate.
How current is any of this?
Every page shows the date it was last reviewed, at the top and in the footer, and the cycle is twelve months or sooner when something material changes. The date reflects a real review rather than the last time the site was rebuilt, which means an old date is telling you the truth rather than hiding.