Forward-deployed engineers are the new startup wedge
In May 2026, OpenAI did something that looks nothing like a software launch. It put $4 billion into a separate deployment company, valued at about $14 billion, whose job is to send engineers into client offices and build AI systems next to the customer’s own staff (SiliconANGLE, May 2026). Its first acquisition, a London services firm called Tomoro, arrived with roughly 150 forward-deployed engineers and a client list that already included Virgin Atlantic and Tesco. Six months earlier OpenAI had taken a stake in Thrive Holdings and agreed to place its research, product, and engineering teams directly inside Thrive’s portfolio companies, starting with accounting and IT services (techstrong.ai, December 2025).
Read those two moves together and a pattern shows up that I think most founders are misreading. The most valuable AI company on the planet is not winning its enterprise business by shipping a slicker self-serve product. It is winning by putting humans inside the customer. I have come to believe that for most AI-native companies the first real product is not software at all. It is an engineer sitting in the customer’s workflow, and the software is what you extract from that seat over time.
That idea runs against almost everything the last decade of startup advice taught us. So let me make the argument properly, because the mistake it corrects is an expensive one.
Why the self-serve default quietly broke
For about fifteen years the default go-to-market for a software company was product-led growth. Build something a user can try alone, remove the salesperson, let the product sell itself, and keep gross margins in the high seventies or low eighties. That worked because the buyer already knew what the software was for. A CRM, a design tool, a payments API: the category existed, the workflow was legible, and the user could map the product onto a job they already understood.
AI agents break that assumption in a precise way. There is often no incumbent product to copy and no settled workflow to map onto. The analysis of the forward-deployed model at startuphub puts it plainly: for a lot of agentic products there is no incumbent, so discovery has to happen by hand, inside the customer’s environment (startuphub.ai). The customer cannot tell you what to build, because they have never seen the thing you are trying to build. They can describe their pain in the language of their current tools, which are the wrong tools, which is exactly why the description sends you in the wrong direction.
The market data is starting to reflect this. Gartner, cited in Deloitte’s 2026 technology predictions, expects that by 2030 roughly 35 percent of point-product SaaS tools will be replaced by AI agents or absorbed into larger agent systems, and that at least 40 percent of enterprise software spend will move toward usage, agent, or outcome pricing (Deloitte, 2026). Maxio found that 83 percent of AI-native software companies already offer usage-based pricing (Deloitte, 2026). When the price stops being a flat seat fee and starts tracking work actually done, the old self-serve motion begins to leak, because that motion assumes the buyer configures the value themselves. Someone has to make the work happen the first time. More and more, that someone is an engineer you sent.
The engineer you embed is the wedge
The model is not new. Palantir adapted military forward deployment to software back in 2003, on the founding bet that product-market fit is found inside the customer’s environment rather than on a developer’s dashboard (fourweekmba, 2026). The structure it built is worth copying carefully. Palantir split the work into two roles: field people who sat with users in intelligence agencies, refineries, and logistics hubs and surfaced needs nobody had written down, and headquarters engineers who turned those field insights into changes in the core platform (startuphub.ai; fourweekmba, 2026). One analysis estimates that roughly half of Palantir’s employees were forward-deployed at the model’s peak (fourweekmba, 2026). That ratio sounds insane to anyone raised on pure software economics. It is also the cost of knowing something your competitors do not.
Here is how I would describe what the embedded seat actually does, in plain terms. The engineer builds a rough, ugly, custom thing for one customer first, the equivalent of a gravel road that gets that one customer from A to B. The central team then watches which gravel roads keep getting rebuilt in the same shape across customers and paves those into real product, the superhighways everyone can drive on (startuphub.ai). You are not choosing between services and product. You are using services as the sensing layer that tells you which product to build, and you only build the part the field proves out. This is the same instinct I argued for in treating founder-led sales as a research method, extended from the sales conversation into the delivery itself.
The reason this beats a survey or a design sprint is that people are unreliable narrators of their own work. They tell you the official process. The embedded engineer sees the real one: the spreadsheet that overrides the system of record, the one analyst everyone quietly routes exceptions to, the step that officially takes an hour and actually takes three days of waiting. That gap between the stated workflow and the real one is where the durable product lives, and you cannot see it from a dashboard. It is the same reason I keep telling founders that distribution and context beat a clever feature: the information you need is in the room, not in the roadmap.
What the seat actually buys you
When I weigh the two motions against each other, the embedded one looks worse on every metric a spreadsheet tracks and better on every one that compounds. Self-serve gives you a clean funnel, high margins, and fast logos. Forward deployment gives you something slower and, I think, more valuable: a map of a workflow nobody else has, data generated by real use rather than a demo, and a level of trust that a free trial cannot manufacture. Here is the honest trade.
| Dimension | Self-serve product | Forward-deployed |
|---|---|---|
| First customer | Whoever signs up | One or two design partners you pick |
| What you learn | Funnel and retention metrics | The real workflow, including the undocumented parts |
| Early gross margin | 70 to 80 percent, the SaaS norm (Growth Unhinged, 2026) | Nearer 50 percent, the AI-product median target (Growth Unhinged, 2026) |
| Source of the moat | Brand and shipping speed | Proprietary workflow data and switching cost |
| Main way it fails | You build something nobody needed | You stay a consultancy forever |
The margin line is the one founders fixate on, and I understand why. Fifty points of gross margin feels like a tax after a decade of eighty-point SaaS stories. But a lower margin on revenue you can actually get, from a workflow you now understand better than the incumbent, is worth far more than a high margin on a product the market is not ready to buy alone. The margin is a problem you earn the right to solve later, once the gravel roads are paved. The alternative, a beautiful self-serve product with a trial funnel that nobody completes because they do not yet know what the tool is for, has a margin of roughly zero on roughly zero revenue.
The money looks like services, and that is fine for a while
Let me be concrete about the finances, because this is where the model gets argued down fastest. The 2026 State of B2B Monetization report found the median target gross margin for AI products sits around 50 percent, against the 70 to 80 percent or higher that traditional SaaS still enjoys (Growth Unhinged, 2026). The same report found that AI-native companies pull a meaningful share of their revenue from services and headcount budgets, roughly 35 percent from services, where older software companies draw almost entirely from technology budgets (Growth Unhinged, 2026). In other words, the companies closest to the frontier are already booking revenue that looks partly like services, and the market is funding it.
The move that makes this work is land and expand. You win a high-value first contract to embed, you deliver a result the customer can measure, and then you expand into the next workflow in the same account at a far lower cost of sale, because you are already inside and already trusted. The first deal pays for the learning. The second and third deals are where the margin starts to recover, because by then you are selling the paved product, not rebuilding the gravel road. I walked through the mechanics of this kind of staged capture in more detail when I wrote about revenue models for AI products, and the forward-deployed motion is the cleanest real example of it I know.
The pricing is also quietly shifting under this. Hybrid pricing, blending a usage component with a base fee, is now the single most common model at 37 percent of companies, up from 25 percent a year earlier (Growth Unhinged, 2026). That matters for embedding, because when you price on work done rather than seats occupied, the services you deliver during deployment and the product you sell afterward sit on the same meter. The customer stops drawing a hard line between consulting and software, which is the line that used to make services revenue look like a weakness. If you want the deeper version of why unit economics, not benchmark scores, decide these outcomes, I made that case separately on AI product gross margins.
When the biggest lab copies Palantir, pay attention
I keep coming back to the OpenAI move because of what it signals, not because most founders can copy it at that scale. A company with the best self-serve distribution in the history of software, hundreds of millions of people using its product with no salesperson involved, chose to spend $4 billion standing up a separate company whose entire job is to put engineers in client buildings (SiliconANGLE, May 2026). It bought Tomoro to acquire roughly 150 people who already do this work (SiliconANGLE, May 2026). Through its investment and consulting partners it says it now sponsors more than 2,000 businesses around the world (SiliconANGLE, May 2026). And through the Thrive Holdings stake it is going further still, embedding its own research and engineering teams inside the operations of companies it part-owns, beginning with accounting and IT services (techstrong.ai, December 2025).
If pure self-serve were enough to capture enterprise value from AI, the company with the best self-serve engine on Earth would not be hiring an army of implementation engineers. The read I take from this is simple: the model does not get better than the workflow it is applied to, and the workflow is owned by the customer, not the lab. Whoever gets inside the workflow and shapes the data and the trust around it captures the value. A model is a commodity you rent. A workflow you understand from the inside is not. This is the uncomfortable version of a point I have made before about why wrapping a model is not a business: the wrapper that wins is the one with a human-earned understanding of the work underneath it, and you get that by sending someone in.
The honest case against embedding engineers
I do not want to pretend this is free, so let me argue the other side as well as I can. The strongest objection is that this is just consulting wearing a software costume. The forward-deployed literature has a name for the failure: the consulting trap, where a company gets addicted to bespoke contract revenue and never generalizes anything into product, so it scales linearly with headcount and dies at the margin ceiling (startuphub.ai). That risk is real and common. Plenty of companies have embedded, delivered, collected the check, and then embedded again from scratch, learning nothing transferable. That is not a wedge. That is an agency.
The second objection is about people, not margins. This model scales with how many strong, trusted engineers you can put in rooms, and those people are expensive, hard to hire, and slow to train. Palantir carried roughly half its workforce in this role and accepted the cost (fourweekmba, 2026); a seed-stage company cannot. There is also the investor reaction. A lot of venture capital is still allergic to anything that reads as services revenue, and you may be penalized on valuation for a revenue mix that a traditional SaaS comparison does not reward, even when that mix is exactly what is buying you the moat. These are genuine costs. I think they are worth paying when the category is new enough that no self-serve path exists yet, and a mistake when the workflow is already well understood and a clean product would sell itself.
How to run the first ten deployments
So here is what I would actually do if I were starting an AI-native company into an unformed workflow today. First, pick one or two design partners on purpose, not whoever shows up, and choose them for how representative their workflow is, not how big their logo is. Second, write into every deployment contract that you keep the right to generalize what you build. The custom work is yours to pave; a contract that locks your learning inside one customer turns the wedge back into an agency. Third, instrument everything the embedded engineer touches, so the gravel roads that keep reappearing across accounts are obvious in the data and not a matter of opinion.
Fourth, set a hard internal rule for when a repeated custom build graduates into the core product, and hold to it, because the pull toward one more lucrative bespoke contract never stops. Fifth, staff the first deployments with your strongest engineers, including founders, because the point of the seat is to learn the workflow, and you cannot delegate the learning that your whole strategy depends on. The failure mode is treating deployment as a cost center to be handed to juniors. It is the research lab. If you want the broader frame for sequencing a one-person or tiny team through this, I laid it out in the solo founder scaling playbook and in the AI-native founder playbook, and the forward-deployed motion drops cleanly into both. For a vertical specifically, the same logic shows up in the vertical AI SaaS playbook, where owning one industry’s workflow is the whole game.
Questions founders ask me
Is this just consulting with a better story
No, and the difference is a single discipline: you only embed in order to generalize. A consultancy sells the same bespoke hours again and again and keeps the learning trapped inside each engagement. The forward-deployed model treats each deployment as raw material for a product every future customer can buy without the custom work. If I built a dozen deployments and could not point to a growing core product that absorbed the common patterns, I would admit I had drifted into an agency. The test is whether your cost of delivering the same result falls with each new customer. If it is flat, you are consulting. If it is dropping, you are building a company.
How to keep margins from collapsing
I treat the low early margin as a loan, not a permanent state. The first deployments should run near the 50 percent AI-product median, and I plan for that openly rather than pretending the number will be eighty on day one (Growth Unhinged, 2026). The way the margin recovers is paving: every time a custom build becomes reusable product, the next customer gets it at almost no marginal delivery cost, and the blended margin climbs. I would watch the margin on the second and third deal in each account, not the first. If expansion revenue is not carrying a visibly higher margin than the land deal, the generalization engine is broken and needs fixing before I add more customers.
When should you stop embedding and go self-serve
The moment the workflow becomes legible enough that a new customer can get value without a human in the room. That is a real, observable threshold: it is when your deployment engineers find themselves doing the same three steps every time with no judgment required. When the custom work stops teaching you anything new, the sensing job is done, and continuing to embed is just paying a premium for learning you already have. At that point I would push hard toward self-serve for the now-understood segment while keeping the embedded motion alive for the next unformed workflow. Most durable companies run both at once, aimed at different stages of the same market.
What kind of engineer makes a good forward-deployed hire
Not the strongest pure coder on the team. The person who thrives in this role is comfortable sitting with a confused customer, asking the dumb question, and tolerating an ugly first solution that works over an elegant one that does not ship. They need enough product instinct to notice when a custom request is actually a general pattern in disguise, and enough spine to say no to work that will not generalize. I have found that people who have done real operating jobs, not only engineering, tend to be unusually good at it, because they already know that the stated process and the real process are different things.
Does this work for a solo founder or a tiny team
It works unusually well, with one caveat. As a solo founder you are the first forward-deployed engineer, and that is the best possible use of your time, because the learning goes straight into the person making every product decision. The caveat is capacity: you can embed in one or two workflows at a time, not ten, so you have to be ruthless about picking the one whose workflow generalizes widest. The danger for a small team is saying yes to a second and third custom engagement before the first has been paved, at which point you are a two-person agency with no time left to build the product. Depth on one workflow beats breadth across five.
How do investors react to services-heavy revenue
Mixed, and you should be ready for it. Some investors still reflexively discount any revenue that smells like services, and will compare you to pure SaaS multiples you were never going to hit. The ones worth taking money from understand that AI-native companies already draw around a third of their revenue from services budgets and that this is a feature of the category, not a flaw in your plan (Growth Unhinged, 2026). My advice is to show the generalization curve directly: the share of revenue coming from reusable product rising quarter over quarter, and the delivery cost per customer falling. That chart turns a services story into a software story, and it is the honest version of what is actually happening.
I’m a four-time founder and reinvention specialist. I build AI-native operating systems that run businesses, and I write about reinventing yourself and your business for the AI age at vikasmalpani.com.