Everyone is hiring for a skill set that barely exists. The capability you need is already on your payroll. This first article of a three-part series[1] examines the role of forward deployed engineers (how they work and the expense and information ownership associated with these individuals) at companies embracing AI vs. developing AI capabilities inhouse with the people who do the work.
Everyone is hiring for a skill set that barely exists. The capability you need is already on your payroll. This first article of a three-part series[1] examines the role of forward deployed engineers (how they work and the expense and information ownership associated with these individuals) at companies embracing AI vs. developing AI capabilities inhouse with the people who do the work.
An organization with real maturity in lean, in technology, and in AI is a purple unicorn in its own right. So when a partner invited us to help kick off a Lean AI project on improving how one company ran its kaizen events, I jumped.
The starting point is straightforward: run a hackathon, trystorm with the technology, see what you can spin up before committing to a design. Beforehand we review the internal tooling. This company had done better than most, with an enterprise secure deployment of Claude that people used regularly, but a few tests turned up two things: The tool had limited access to their enterprise software, and something in the system prompt steered anyone who raised an integration question toward the core software group.
Wanting the team unfettered while they iterated, I asked their technology leader what could be opened up for the exercise.
He was helpful, and walked me through the pace at which security could grant access and the deliberate work underway to build durable pathways to the technology rather than fast ones. He committed to go to work on it.
Then we got to the bigger problem. Technology leaders are being handed two mandates at once: scale AI across the organization, and build real capability with it. The industry’s answer has become the forward deployed engineer. Those can help, particularly when an organization hires them onto its own payroll rather than renting them. But what you need is capability sitting close to the work, with a feel for the technology, able to iterate on it, and to partner with the core software group instead of queuing behind it.
Not a forward deployed engineer, I said. A forward developed one.
He took the phrase immediately, and asked whether he could use it. Then he described his own purple unicorn: the hire specified in requisitions he had been tasked with filling, several of them, across sales, marketing, and each of the technical functions where the company’s advantage lived. He was struggling to fill the roles. The intersection of skills is narrow, and it is the most in-demand profile in the market right now. Watching him consider a possible alternative to help, you could see something ease.
Build capability at the layer closest to the work, and you need less of the unicorn because you no longer need all of those skills inside one person.
He will keep staffing the role, and he should. But the problem is not only finding a new skill set. Build capability at the layer closest to the work, and you need less of the unicorn because you no longer need all of those skills inside one person.
The market is moving fast and the incentives are clear. In the eight weeks between mid-May and early July, OpenAI launched a majority-owned deployment company with more than $4 billion dollars behind it.[2] Anthropic launched a $1.5 billion venture with Goldman Sachs, Blackstone, and Hellman & Friedman to embed engineers inside mid-sized companies.[3] AWS committed $1 billion to the same model,[4] and Microsoft announced a $2.5 billion operating unit staffed with roughly 6,000 embedded experts.[5] Salesforce has committed to a team of a thousand; postings for the role grew more than 800 percent in the first nine months of 2025.[6] One of Microsoft’s taglines for the new unit is “No Pilots. Scale from Day One.”[7]
The firms behind this are no dummies, and there is enormous money motivating the move. They have also identified something real. And the forward deployed engineer is also an easy thing to sell.
The forward deployed engineer has a job title, a salary band, a start date, and a vendor who will stand behind it. It appears in the board deck as a line item with a number attached. And it asks almost nothing of you. The management structure stays as it is. No manager diverts attention or changes how they work. Someone arrives, studies the process, and hands back a new one, a redesigned workflow with a headcount reduction attached, and the manager’s job is to follow it. There are no hard conversations with your people. The quiet assumption is that if the disposal happens in the background, everyone else can move forward.
The alternative — forward developed engineer — puts your leadership and your teams to work. It requires honesty about where the redesign leads and a plan for the freed-up labor before you free it. It arrives as a request for engineering time from people who currently have other jobs, with no title, no owner, and no number to put beside it. In the short term it can feel like missing the boat. In the long term, the compound interest on trust, connection, and loyalty exceeds whatever temporary edge the arms race was going to hand you.
Under pressure to show returns, one of those is defensible in the room and the other is not. The asymmetry is not payback. One can be bought and the other has to be led.
The vendors get a great deal of this right.
Earlier this year the CEO of a well-funded deployment firm, Vasuman Moza of Varick Agents, told a room of industry peers that execution is no longer the constraint. The models are good enough and so are the harnesses. What remains is how deep you can go into a customer without hiring an army to do it. He named the bottleneck outright: designing how the work gets done around AI.[8]
He went further. Documentation describes the golden path and maybe an edge case or two. The reality is that when the invoice does not match the purchase order, Sarah hands it to Chris and four days of cycle time disappear. Every company handles that moment differently, and the difference is what a model cannot guess. He is describing the need for effective gemba, and he is right that it is the critical lever for reengineering a workflow.
His head of engineering, JD Puett, then showed the room what that costs. He described his own office: the platform team on one side, in his words having a great time, hanging out and building with the coding agents. Across the room, the forward deployed engineers were stressed, sleep deprived, clients emailing them around the clock. Their process was to upload roughly 150 pages of documentation into Claude, write a prompt, wait a couple of minutes, and get back something verbose and incorrect. Sitting with process leads, watching the work, and iterating in a messy layer until patterns stabilize into something you can call a standard is deliberate, detailed, and social.[9] The cycle can be run fast. It cannot be skipped, and the upload and hope is a skip. Skipped work does not disappear. It comes back as churn, and the burnout in that room is the evidence. Anyone who has run a kaizen event recognizes the shape.
There is a second thing hiding in the pitch. A great deal of what gets sold as AI return is return on attention. These are processes nobody has examined in years. Put a capable person in front of one with a mandate to understand it, and you will find time and money whether or not a model is involved. That is an argument about attribution, and about who ought to be doing the examining.
What happens to that knowledge next is the part worth watching. Forward deployed engineers interview every process lead in a department. What they capture is encoded as a dependency graph and used to train a model, all of it running on the vendor’s platform. Your process knowledge, the part your own documentation never captured, goes in one end. Out the other end comes a trained model and a toolset the vendor owns, improves, and carries to the next client in your industry. Microsoft’s launch language promises the opposite: your intelligence compounds, and your IP stays protected. That the largest vendor leads with the promise tells you what buyers have started to fear.
None of this is concealed. Palantir’s posting for its Forward Deployed AI Engineer offers, as a benefit of the job, “ample opportunity to contribute learnings from the field back to the Palantir AIP product suite.”[10] The loop is not a secret. It is a selling point, aimed at the engineer.
But there is one capability you never rent, and it is the ability to see and improve your own work.
Then there is the arithmetic. An engineer whose base salary runs to $400,000, and whose loaded cost runs well past it, is not underwritten on a single return.[11] Immediate savings inside your operation. Reusable pattern knowledge that makes the vendor faster and better at your competitor. Architectural dependency on a stack, a platform, and a compute relationship you did not previously have. Only the first — operations improvements — flows to you. The second is manufactured out of your process knowledge, and the third is the bill for maintaining what they built.
That is not a swindle. Pooled learning across many clients produces tools no single company could build alone, and renting capability is frequently the right call. Nobody smelts their own steel.
But there is one capability you never rent, and it is the ability to see and improve your own work.
Lean AI Basics gives your practitioners a repeatable framework for putting AI to work in lean, taught live by Art Smalley. Two sessions, online.
Org Strategy
Tyson Heaton spent 15 years running operations in manufacturing at JBS, Schreiber Foods, and Greencore, learning lean on the floor before moving into enterprise technology leadership at O.C. Tanner. The frustration he found there became a throughline: organizations that had mastered continuous improvement on the shop floor were treating their technology layer as a black box, a dependency to manage rather than a capability to master. He now leads LEI's LeanTech/AI initiative, working with organizations done accepting that tradeoff -- the work is moving faster than most can keep up with, and he'd rather be part of sharpening that thinking than watching from the side.