They Just Revealed Their Internal Process for Becoming an AI FDE

AAI LABS
Computing/SoftwareManagementJob Search

Transcript

00:00:00if you're someone who knows how to use ai there's a huge opportunity in front of you right now and
00:00:04interestingly enough even though it's one of the roles companies need the most they can barely find
00:00:09anyone who can do it that opportunity is a role called forward deployed engineer or an fde and
00:00:14it's not that the work is hard to learn it's just that people don't know this role exists in the
00:00:19first place and over the past month this role has gotten important enough that the people running
00:00:24these teams at openai anthropic and cursor have all come out and given talks on it and you don't
00:00:29normally get to hear any of this because they're walking through real projects inside real
00:00:33businesses so we've gone through all of them and pulled out the parts you can actually use now if
00:00:38you're new to the channel then welcome we're a software company and this is ai labs where we show
00:00:43you how to optimize a business with ai using proven methods from our own team and in this video we're
00:00:48going to go over what an fde is how the big companies actually do this work and how you can start doing
00:00:53it yourself but before we go into anything let's start by defining what an fde actually
00:00:58is a forward deployed engineer or an fde for short is basically someone whose whole job is getting a
00:01:03business to actually use ai and that doesn't mean going in and advising them or handing over some
00:01:09strategy document it means getting ai running inside the systems they already use so that the work
00:01:14their people were doing by hand just gets done by ai instead no matter what kind of process it is
00:01:20there are three things working in it either it is a human who's doing the task or it is software that
00:01:25follows hard set rules to do the task or it is an ai and an fde is someone who actually knows which of
00:01:30those three should handle which task in the process for example a company put an agent on its refund
00:01:36requests and on paper it worked it read each request checked it against the refund policy and turned
00:01:41down the ones that didn't qualify but a few weeks in the company started losing customers they'd had for
00:01:46years and nobody could work out why because every refund the agent turned down was one the policy said to
00:01:52turn down so the fde went and sat with the person who used to do that job and turned out they had a step
00:01:57of their own which wasn't documented they checked how the order had been paid for and if the purchase
00:02:02was on a company card they would just approve it without reading any further because those come
00:02:06from businesses that buy every month and arguing over one refund costs you the account the policy
00:02:11said nothing about that they'd worked it out themselves years ago so that check went into the
00:02:16workflow as a fixed rule ahead of anything the model decided now when you're working with people there is a
00:02:22problem most of them can't tell you what they need they'll describe something they've already imagined
00:02:26instead of the problem they're actually having and that's the biggest issue of getting a business to
00:02:31adopt ai here's why the people that actually do the work always just need to complete their job they
00:02:36just want that the output comes out as easily as it can they don't look at the process itself so it's
00:02:41really hard to build something that they would adopt and that's exactly where most businesses are right
00:02:46now ever since ai got popular they've been adding it into everything they do without actually knowing
00:02:51if they even need it in the first place so that's why everyone's suddenly hiring someone who can
00:02:55actually sort this out for them job postings for this role are up 729 in a year aws put a billion
00:03:02dollars into building a whole department of forward deployed engineers and openai's own fde team went
00:03:08from two people in january to 39 and all of that happened really fast because three years ago basically
00:03:13nobody needed an fte and now there are over a hundred startups hiring for that role inside y combinator
00:03:19alone which is where a lot of the biggest startups today got started colin jarvis runs the forward
00:03:24deployed team at openai and he says there just aren't a lot of fdes out there so the demand turned
00:03:29up before the people did before we dive deeper into this role it would be great if you subscribe to
00:03:34the channel and hit the hype button this small gesture of support goes a long way for us so the demand is
00:03:40there but that still doesn't explain why an fde has to be a separate job because the tools aren't the
00:03:45hard part anyone can sign up for them and most businesses already have last year mit ran a study
00:03:51on exactly this they looked at 300 ai projects and surveyed hundreds of people inside the companies
00:03:56that ran them and they found that 95 of those projects produced no measurable return at all and mit's
00:04:03own conclusion was that the companies themselves caused the failures not the models the ones that got
00:04:08nothing were using generic tools that looked good in a demo and fell apart the moment they were used on
00:04:13real work and nothing they built around those tools ever learned how their business actually ran now
00:04:18vasuman moza does this kind of work for a living he was an engineer at meta and now his company gets ai
00:04:24working inside big companies essentially he is an fde so he's watched a lot of these projects go wrong
00:04:30up close and he says ai is just getting slapped on top of broken processes because nobody actually looks at
00:04:36the process first one executive he spoke to burned through a 10 million dollar budget in three months
00:04:41and it was meant to last a year his company had handed the tools to everybody and just left them
00:04:46to it so everyone spun up whatever they felt like the money went and none of it made the business any
00:04:51better at anything it was already doing and the reason that work has to be somebody's whole job is
00:04:56that nobody already inside the business is going to do it the engineers have their own work to get
00:05:00through and the people running the process have been doing it for so long that the workarounds in it just
00:05:05look completely normal to them and palantir lost a whole year to one of those workarounds a move
00:05:10onto a new file format was stuck because one engineer kept saying the new format was worse and nobody
00:05:15could work out why until someone actually watched her work turns out she'd been checking the data
00:05:20by double clicking the files open and the new format had nothing you could double click so the team built her
00:05:26something that night that let her open the new files the same way and she approved the move two days later
00:05:31so that's why it's worth having somebody whose whole job is to work all of that out before anything
00:05:36gets built because the alternative is a waste of all your cost and efforts a lot of big companies like
00:05:42anthropic and open ai have recently given talks about how they handle this whole process internally
00:05:47we compiled the best of those talks and when we analyzed them we found that the same patterns kept showing
00:05:52up in every single workflow the first step is to pick the process that's actually costing the business
00:05:57something and with ai that basically means volume so you're looking for the person who handles the
00:06:02same kind of message 50 times a day because automating one message saves nobody anything but automating
00:06:08a message that needs to be sent a thousand times a week is where the business actually becomes more
00:06:13productive and earns more and that job is easy to find now because a business already has its whole past in
00:06:19its support history you hand that to a model and it tells you where the volume is open ai's team did
00:06:24exactly that at one of the biggest banks in the world they went after the one job thousands of
00:06:29advisors did every day and around 98 of them ended up using what they built the second step is to build
00:06:35on top of whatever the business already runs and with ai that matters way more than it used to
00:06:40because an agent is only worth what it can reach so if a team already keeps all their work in notion you
00:06:45don't build them a separate system and move all of that across just so an agent can read it more easily
00:06:50you connect the agent to their notion with an mcp and everyone carries on working the way they always
00:06:56did that's how most of them work because they don't want to give up a system they've already put a lot
00:07:00of time and money into one of moza's clients had spent five million dollars and five years getting onto their
00:07:05finance system so moving them off of the system was never going to happen the whole job was getting
00:07:10everything else in that business talking to it the third step is to not change the way people work more
00:07:16than you have to if somebody's been running an 11 step process for years and you hand them a one-step
00:07:21version they stop using it they used to check the work as they went and now the middle is gone and
00:07:26they're being asked to trust an answer that just appeared so moza says you need to leave the process
00:07:31looking like the one people already know you let the agent do the work inside each step but you keep
00:07:35the steps there so people can still have something they can see which would let them know the answer
00:07:40is right and the last step is to budget way more time for trust than for building at that same bank
00:07:45open ai went into the technical side was finished in six to eight weeks and then it took another four
00:07:50months of pilots and testing before the advisors would actually rely on it because if you're doing
00:07:55something every day the same way you've done it for the past year then any change to that process has
00:08:00to earn its way in whether you're handing it to 5 000 people or to five but before we move forward
00:08:05let's have a word by our sponsor dex ai the talent agent for software engineers you're not short on
00:08:10opportunities you're short on the right ones the best roles never get posted they go to whoever's
00:08:15in the network dex fixes that it works entirely on your behalf not a job board not a recruiter it's
00:08:20the insider who gets you in the room so i tried it you brief dex once on what you do and want honestly
00:08:25the conversation sold me it asks the right questions and gets what you're after then it does the work no
00:08:30applications no recruiter screens dex comes back with a short list of vetted roles from startups like
00:08:36linear and atio averaging over 150 000 every role is handpicked for fit never scraped by a bot you get
00:08:43a direct line to the hiring team with salary up front referral to offer averages 29 days versus 55 on a
00:08:49job board it's free for engineers so stop applying and start getting introduced meet dex at the link in the
00:08:55pinned comment so that's how the big companies get ai into a business now if you want to actually run
00:09:00this yourself the roadmap is five steps and the first one happens before you build anything first you need
00:09:06to watch how a job is really done and write down every single step of it in the order it actually
00:09:11happens then you need to ask why each of those steps is done the way it is and if nobody can give
00:09:16you a concrete reason that's usually a workaround somebody put in years ago that nobody has questioned
00:09:21since second you have to decide which parts of that process should actually become ai and which parts
00:09:26shouldn't because if you hand the whole thing over to a model that's exactly how you end up in that 95
00:09:32of ai projects that never produce anything so you need to run every step you wrote down through three
00:09:38filters the first is whether that step follows a fixed rule and has to come out right every single
00:09:43time and if it does it just stays as ordinary software and doesn't need to be automated with ai the
00:09:48second is whether somebody has to read something messy and make a judgment call on it because that's
00:09:53the part the model is actually for and the third is what it costs when it goes wrong because if getting
00:09:58it wrong is expensive that step stays with a person even when a model could do it and once you've done
00:10:03that you'll almost never find that the whole process should be run by ai some steps stay as software
00:10:08some go to the model and some stay with a person in moza's example of an eight-step process four of the
00:10:14steps ended up running on their own another three ran with a person checking the output before it went
00:10:19anywhere and the last one stayed fully human and the reason for that step being fully human is because
00:10:24it's a business call which ai cannot do yet so either the cost of getting it wrong is too high to hand
00:10:29over or the step doesn't come up often enough for a model to be worth the trouble third you need to
00:10:34build for the ways the system can fail and not just for the case where everything goes right and moza says
00:10:39that when there's only one way something can go right there are also a thousand ways it can go wrong
00:10:44so if you only build for the way it goes right what you built is worth nothing and with ai that mostly
00:10:50means handling the places where the model isn't sure because it'll hand you an answer either way
00:10:55even when it has no idea fourth you need to make sure the system actually works before you let it near
00:11:00anything real now a model gives a slightly different answer every time even if you ask it the same
00:11:05question so if you only test the system a few times you're going to miss the rare ones where it falls
00:11:09over what you do instead is take real examples out of the process you wrote down where you already know
00:11:14the right answer then you run the system across all of them and count how many it got right and then you
00:11:20go and read the ones it missed and fix those and last you have to put a number on what it was actually
00:11:25worth and you should be able to say whether it brought money in or took a cost out or made a risk
00:11:30smaller because nothing else counts now cursor had somebody complained to them that an agent was
00:11:35costing him two thousand dollars a day and when they asked him what that agent was actually doing
00:11:39turns out it was just picking which engineer to send out to fix broken equipment so then they asked him
00:11:44what it was costing him to send the wrong person out and it was more than two thousand a day and he
00:11:49agreed straight away because all he'd been looking at was what the agent was costing him and never what
00:11:54it was saving him and those five steps are how we built the internal chatbot in our own software
00:11:59company now if you've seen our previous videos you'll know it's basically just a chat interface
00:12:03that lets the non-technical people on our team use claude code and everything it can do we built it
00:12:09because the people in hr and accounts aren't engineers and everything around claude code assumes
00:12:14you are one before claude code could do anything useful for their work it'd have to be connected to
00:12:19slack and email and the rest of what they use every day and that's a setup job none of them were ever
00:12:24gonna do so the chat interface does all of that connecting for them and they just type what they
00:12:29need but we didn't start building our engineers went and sat with the people in hr and accounts
00:12:34first because engineering had no real idea how those departments actually spent their days so they spent
00:12:39proper time in those departments and wrote down every single repetitive thing that was being done by hand
00:12:45and then worked out which of those a model could reasonably take over and then that document turned
00:12:50out to be most of the build those written up processes became the instructions the agents follow
00:12:54and the knowledge the chatbot answers from then we handed the first version back to those departments
00:12:59and had them use it on their own real work and they came back and told us it was actually saving them time
00:13:05now the last question is what you actually need to be good at to run any of this one of the companies
00:13:09that hires for exactly this job wants people who are wide across business process and technology with
00:13:15one area they're genuinely deep in and they say the business half gets taught on the job while the
00:13:20technical half doesn't but for you that's probably backwards because you just have to focus on the
00:13:25business half because the building is handled by agents now and that's fine because a palantir exec says
00:13:30that the person who fails at this job is the careful engineer who wants code that still holds up in 10
00:13:36years the job is getting something rough in front of a real user quickly which is basically what an
00:13:41agent like claude code hands you on the first pass anyway so now that you know what an fte does and
00:13:47how the work actually gets done let's look at what you go and do with it the place to start is the
00:13:51smallest business you can physically walk into so wherever you work or somewhere a friend runs then you ask
00:13:57whoever runs it if you can spend an hour sitting next to whoever does the most repetitive job in
00:14:02the place while they're actually doing it and then you just run the five steps on what you saw what
00:14:07comes out the other end is called an audit and it's a real thing businesses actually pay for moza who
00:14:12runs those ai projects inside big companies starts every single one of them with an audit before anything
00:14:18gets built and he says that the first phase is the biggest bottleneck in the whole job and that's what
00:14:23you put in front of people whether you're applying for an fte job or pitching a business directly
00:14:28because inside it you've got one real company's process written down the steps you'd hand over to
00:14:32ai and what each of those is worth to them in money now we have curated roadmaps and guides for those
00:14:38who want to dive deeper into the role of fte which can be found in ai labs pro which is our community so
00:14:44if you've found value in what we do and want to support the channel this is the best way to do it the
00:14:49links in the description that brings us to the end of this video if you'd like to support the channel
00:14:54and help us keep making videos like this you can do so by using the super thanks button below as
00:14:59always thank you for watching and i'll see you in the next one

Key Takeaway

Successful artificial intelligence deployment requires Forward Deployed Engineers to map existing human workflows, separate fixed rules from judgment tasks, and integrate models directly into current software stacks.

Highlights

  • Job postings for Forward Deployed Engineer roles surged by 729 percent within a single year.

  • A study by MIT across 300 artificial intelligence projects found that 95 percent produced zero measurable return due to poor business integration.

  • OpenAI expanded its Forward Deployed Engineer team from two people in January to 39 shortly after.

  • AWS invested one billion dollars into establishing a dedicated Forward Deployed Engineer department.

  • A five-step roadmap guides successful artificial intelligence adoption by auditing workflows, filtering tasks, planning for failure, rigorous testing, and measuring financial impact.

Timeline

Defining the Forward Deployed Engineer Role and Market Demand

  • Forward Deployed Engineers integrate artificial intelligence directly into existing business systems rather than delivering theoretical strategies.
  • Market demand for Forward Deployed Engineers increased by 729 percent over twelve months.
  • Over one hundred startups within Y Combinator actively hire for the Forward Deployed Engineer role.

Businesses struggle to adopt artificial intelligence because employees complete tasks without documenting informal workarounds. Companies like OpenAI, Anthropic, and Cursor report an acute shortage of technical professionals capable of bridging this gap. Consequently, major technology providers allocate massive capital to build specialized deployment teams.

Why Standard Artificial Intelligence Projects Fail

  • An MIT study of 300 projects revealed that 95 percent yielded no measurable return on investment.
  • Failed initiatives typically slap generic artificial intelligence tools onto broken or undocumented internal processes.
  • Unidentified manual workarounds frequently break automated workflows until technical teams observe actual user behavior.

Organizations frequently burn through substantial budgets by handing generic software to employees without analyzing underlying operations. Palantir lost an entire year on a file format migration because an engineer relied on double-clicking legacy files, a workaround resolved only after directly watching her workflow.

Internal Deployment Frameworks Used by Top Artificial Intelligence Labs

  • Deployment teams target high-volume communication channels where automation generates immediate productivity gains.
  • Integrating tools via mechanisms like Model Context Protocol allows workers to retain familiar software interfaces.
  • Allocating four months for testing and trust-building ensures user adoption in critical daily operations.

Analysis of talks from leading laboratories highlights consistent operational patterns. Successful projects identify repetitive tasks using historical support data, build directly on top of existing software like Notion, and maintain familiar multi-step processes so users can verify intermediate outputs.

A Five-Step Practical Roadmap for Implementation

  • Mapping every step of a job reveals undocumented human workarounds before writing code.
  • Filtering tasks divides responsibilities between standard software, artificial intelligence models, and human judgment.
  • Rigorous testing against historical ground-truth data prevents unexpected failures caused by model variance.

Effective implementation starts with a detailed process audit. Tasks are filtered through three lenses: fixed rules stay as software, messy data evaluation goes to models, and high-risk decisions remain human. Systems must be tested against real historical examples rather than quick demos, and final deployments require clear financial metrics proving cost reduction or revenue generation.

Applying the Forward Deployed Engineer Playbook

  • Building internal tools, such as chat interfaces connecting non-technical staff to advanced coding models, streamlines internal operations.
  • Conducting an operational audit for local businesses creates an immediate commercial service offering.
  • Prioritizing rapid deployment over long-term code perfection drives successful artificial intelligence integration.

Practitioners can start by shadowing repetitive jobs in small businesses to write comprehensive operational audits. These audits identify exact automation targets and potential cost savings, serving as a primary deliverable for both job interviews and direct client pitches.

Community Posts

View all posts