The promise of frontier AI has all the time seemed like a utility: plentiful intelligence, obtainable on demand, as straightforward to entry as electrical energy, water, or cloud computing. The metaphor is highly effective, and for good motive. Utilities scale as a result of they summary complexity away. You don’t want an engineer from the facility firm sitting in your workplace each time you activate the lights.
And but, essentially the most refined AI corporations on the planet are more and more doing one thing very completely different: They’re sending individuals.
OpenAI recently announced the OpenAI Deployment Company, explicitly designed to embed forward deployed engineers (FDE) inside organizations engaged on advanced issues in demanding environments. These engineers, in keeping with OpenAI, will work with enterprise leaders, operators, and frontline groups to determine the place AI could make the most important impression, redesign workflows, and switch these beneficial properties into sturdy methods. Anthropic is hiring FDEs for its applied AI team, individuals who embed immediately with strategic clients to drive enterprise adoption and ship real-world functions. And Google is doing the same. Is {that a} coincidence?
That’s revealing. As a result of if intelligence have been already a real utility, this is able to not be mandatory. You wouldn’t have to ship your personal engineers to each buyer to make the tap work.
The paradox of AI as a utility
That is the paradox on the coronary heart of the present enterprise AI mannequin: The business speaks the language of scale, abundance, and platforms, however its supply mannequin more and more resembles high-end consulting.
That doesn’t imply the work is unimportant. Fairly the other. Ahead deployed engineers are sometimes fixing the true drawback: taking frontier fashions out of the demo surroundings and making them perform inside messy, regulated, fragmented organizations. They cope with permissions, legacy methods, compliance, knowledge high quality, workflows, operational constraints, and all of the issues that make corporations completely different from benchmarks.
However that’s exactly the purpose. The necessity for these individuals will not be merely a business innovation. It’s a symptom. It tells us that the product, as presently packaged, will not be but sufficient.
Within the earlier articles on this collection, I argued that large language models were never built to run a company, that enterprise AI must move from tools to systems, and that the systems that finally work will not look like chatbots or copilots, but like intelligence embedded into the organization itself. The FDE phenomenon confirms that argument from the seller aspect. If the AI lab has to ship engineers to reconstruct context, redesign workflows, and make the system function beneath actual constraints, then the lacking layer will not be imagined. It’s sitting there, being provided manually.
The preplatform sample
Each main expertise business goes by means of an artisanal part earlier than it turns into industrial.
Earlier than enterprise software program grew to become packaged, implementation was bespoke. Earlier than cloud platforms grew to become mature, corporations wanted armies of specialists to configure infrastructure. Earlier than the net stabilized round browsers, requirements, internet hosting suppliers, content material administration methods, analytics, and design conventions, constructing an internet site required much more customized work than it later would.
Ahead deployed engineering belongs to that very same historic sample. Palantir popularized the model years ago. Its personal description of the ahead deployed software program engineer function relies on engineers working immediately inside buyer environments to make software program perform in operational actuality.
That mannequin made sense for Palantir as a result of its clients typically had extraordinarily advanced, high-stakes, extremely particular environments. However when OpenAI and Anthropic begin to converge on similar patterns, the sign is completely different: the frontier AI business is discovering that fashions alone don’t cross the enterprise hole.
That doesn’t make FDEs a failure. It makes them a transitional kind. They’re what seems earlier than a class has discovered its true platform layer.
SAP doesn’t ship SAP staff to each buyer
That is the place the comparability with mature enterprise software program turns into helpful.
SAP doesn’t scale by sending SAP staff into each buyer. It has a vast partner ecosystem. Salesforce doesn’t implement each buyer itself. It has AppExchange, now evolving into AgentExchange, and a big ecosystem of companions, unbiased software program distributors, and methods integrators. The platform firm creates the substrate; the ecosystem industrializes supply.
That distinction issues. When the seller itself has to provide the scarce human experience required to make the product work, the class remains to be immature. When companions, integrators, templates, requirements, and repeatable architectures take over, the class begins to scale.
That is why the present FDE wave ought to be learn rigorously. It’s not proof that frontier AI has turn out to be a platform. It’s proof that it has not but turn out to be one.
A real platform reduces the necessity for bespoke intervention. A preplatform product will depend on it.
The enterprise mannequin lure
There may be one other drawback, and it’s extra delicate: As soon as ahead deployed engineering turns into a income, status, buyer lock-in, and strategic proximity, it turns into more durable for the seller to get rid of it. The very individuals fixing the product’s incompleteness can turn out to be a part of the enterprise mannequin that will depend on that incompleteness.
That is basic innovator’s dilemma territory. Clayton Christensen’s argument was that profitable corporations typically battle not as a result of they fail to spot the long run, however as a result of their present enterprise fashions make the long run unattractive or cannibalistic. On this case, the dilemma is straightforward: If a frontier AI firm builds the layer that makes deployments repeatable, modular, and partner-scalable, it could undermine the bespoke, high-touch mannequin that presently brings it near the biggest clients.
That’s the reason the true platform could not come from inside the businesses coaching the fashions. It might come from one other layer.
The lacking layer will not be one other mannequin
The temptation, as all the time, is to imagine that the reply is a greater mannequin. A bigger mannequin. A extra agentic mannequin. A mannequin with longer context, extra instruments, extra reminiscence, extra reasoning traces, and extra autonomy.
However the FDE mannequin suggests one thing else. If engineers are being despatched into clients to map workflows, perceive constraints, join methods, construction context, govern entry, and switch AI outputs into operational outcomes, then the lacking piece will not be merely intelligence. It’s structure.
Extra particularly, it’s the layer that turns firm actuality into one thing AI methods can function inside:
- persistent context,
- course of construction,
- permission fashions,
- constraint administration,
- suggestions loops,
- workflow state,
- enterprise semantics, and
- consequence monitoring.
Right this moment, that layer is usually reconstructed manually by professional engineers on every deployment. Tomorrow, it should turn out to be infrastructure.
That’s the actual alternative.
Why that is actually BPR with brokers
This additionally connects on to the return of enterprise course of reengineering (BPR).
In 1990, Michael Hammer’s well-known Harvard Enterprise Assessment article, “Reengineering work: Don’t automate, obliterate,” argued that corporations shouldn’t use expertise merely to hurry up outdated processes. They need to redesign the processes themselves. The thought was proper, however in lots of circumstances the expertise of the time was not but able to supporting the ambition.
AI adjustments that, nevertheless it additionally makes the issue extra demanding. If corporations merely insert AI into present workflows, they get quicker variations of out of date processes. If distributors merely ship engineers to customise every deployment, they get artisanal transformation that doesn’t scale.
The true breakthrough comes when the redesign itself turns into systematized: when enterprise processes are usually not simply automated, however represented, ruled, tailored, and optimized constantly.
That’s the level at which enterprise AI stops being a consulting engagement and begins changing into a platform.
The FDE is the clue
That is why the ahead deployed engineer is so attention-grabbing. The FDE will not be the way forward for enterprise AI. The FDE is the clue that the long run has not totally arrived.
The function exists as a result of present methods nonetheless require people to bridge the hole between normal AI functionality and particular organizational actuality. Somebody has to translate the corporate into the machine. Somebody has to interpret constraints. Somebody has to find out which workflows matter. Somebody has to attach knowledge, course of, motion, and consequence.
However historical past suggests that after a repeatable layer seems, the artisan turns into much less central. Internet consultants didn’t disappear after the net matured. However “construct me an internet site” stopped being a mysterious customized engineering drawback for many organizations. Enterprise useful resource planning (ERP) consultants didn’t disappear after SAP matured. However the ecosystem grew to become standardized sufficient that the seller didn’t have to personally deploy the product in every single place. Cloud architects didn’t disappear after Amazon Internet Companies (AWS) grew to become a platform. However infrastructure grew to become programmable, repeatable, and scalable.
The identical factor will occur right here: Ahead deployed engineers won’t vanish. But when enterprise AI turns into an actual platform class, they’ll turn out to be distinctive relatively than foundational.
The true check of a platform
The check is straightforward: Can the system work with out sending the lab? Can it perceive the corporate with out a bespoke mapping train each time? Can it function beneath constraints with out guide reconstruction? Can it adapt to workflows with out a group of engineers sitting contained in the buyer? Can companions construct on it? Can clients configure it? Can it scale past the handful of enterprises that may afford white-glove deployment?
Till the reply is sure, we ought to be trustworthy about what’s being bought. It’s not AI on faucet. It’s AI on faucet, with plumbers included.
And that’s positive, for now. Each class has its artisanal part. The error is complicated that part with the vacation spot.
What comes subsequent
The subsequent stage of enterprise AI won’t be outlined by who has essentially the most spectacular mannequin or the biggest deployment group. It is going to be outlined by who builds the layer that makes these deployment groups much less mandatory.
That layer won’t merely reply questions. It would characterize the corporate. It would encode processes, constraints, permissions, reminiscence, and outcomes in ways in which AI methods can truly use. It would enable fashions to function contained in the enterprise relatively than hover above it. It would flip bespoke deployment into repeatable structure.
When that occurs, the present FDE growth will look apparent looking back—not as the ultimate type of enterprise AI, however because the bridge between demos and platforms.
And when the true platform layer seems, the business will change in a short time. As a result of utilities don’t scale by sending engineers to each sink. They scale when the plumbing is already there.
