Two of the most advanced AI laboratories in the world have, in the same week, made the same operational admission.
OpenAI raised $4 billion for a new venture called The Development Company. Anthropic announced a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs.1 Both ventures share a single objective: embed engineers and consultants directly inside client companies to help them actually use AI.
The substance beneath the press releases is uncomfortable for the field. The most capable AI labs on the planet have looked at their own technology and concluded that it cannot be deployed at scale without putting humans in the room.
It is a story about an interface that has failed.
What follows is what I think is actually happening — and why this moment matters for strategic designers, transformation consultants, and the executives funding the next wave of AI investment.
What they built, and why they had to
Both ventures are explicit about the model. They are copying Palantir's forward-deployed engineer approach2 — sending people into client companies to sit with staff, map workflows, identify where AI fits and where it does not, design oversight and trust mechanisms, and ultimately rebuild how work happens.
The reason this approach is necessary sits in plain sight in the failure data.
of enterprise AI agent pilots never reach production.
Anaconda × Forrester research, replicated by a16z and the MIT Sloan CIO panel3
of executives report personal AI productivity benefit. Only 29% report meaningful organizational ROI.
Writer's 2026 Enterprise AI Adoption Survey · n=1,200 executives + 1,200 employees4
Most pilots die before they ship. Of those that reach production, a meaningful share generate negative ROI. The failure mode is consistent across surveys, and it is almost never about model capability.
The 97/29 split is the more revealing number. Individual productivity wins are everywhere; organizational transformation is rare. The gap between personal AI productivity and organizational AI value is the entire enterprise AI problem.
That gap is exactly what the OpenAI and Anthropic ventures are trying to close. Not by building better models. By putting humans in the room to bridge what the models alone cannot.
The shape of the work
Look at what the forward-deployed engineer actually does day to day. They sit with staff. They observe workflows. They identify where AI fits and where it does not. They design oversight, trust, and escalation patterns. They rebuild the interface between humans and AI on a per-organization basis.
This is service design. Workflow design. Change design. Interaction design.
It is being done — at scale, with billions in capital — by engineers, because the design profession has not claimed this work as its own. Yet.
A February 2026 Harvard Business Review article articulated the underlying point precisely: AI initiatives stall because employees develop industry-shaped anxiety about their relevance, identity, and job security.5 That produces surface-level use without real commitment. The failure is psychological and contextual, not technical. Treating AI rollout as a technology project rather than an organizational redesign produces exactly the failure rates above.
This explains the broader pattern. For every dollar enterprises spend on software, they spend approximately six on services. That ratio exists because translating capability into actual organizational change is the hard part. AI is not different. It is the same problem at a larger scale.
The forward-deployed engineer model exists because the design discipline that should be doing this work has not made the case for itself.
The interface — and why this consulting wave is transitional
The reason these consultant armies exist at all is that the current AI interface requires translation.
Every interaction with AI today begins with the user encoding their intent into structured language. Prompt engineering emerged as a profession because that translation is hard. Most people are not fluent in it. The forward-deployed engineer is, in part, a translation specialist embedded in an organization that cannot translate for itself.
This is solvable. It is being solved.
Look at where multimodal AI is heading. Google's Project Astra processes unified streams of video, audio, and text with sub-300ms latency. Google DeepMind has explicitly framed this as "a departure from the token-in, token-out architecture" of early language models.6 The system can see what you are looking at, hear the conversation in the room, remember context from days ago, and act on it. Apple's spatial computing direction, Meta's Ray-Ban glasses, the wave of agentic interfaces — these are all early experiments in AI that does not wait for a prompt because it understands context.
The interaction paradigm is shifting from requesting AI to being assisted by AI. That is not a feature change. It is a category change in what an interface even is.
Engineers will build the underlying capability. Consultants will sell the integration. But the interaction language itself — what trust looks like, what control looks like, what legibility looks like, what happens when ambient AI makes mistakes — has to be designed. And it does not yet exist.
The GUI moment, again
The shift from command-line interfaces to graphical interfaces in the late 1970s was not a technical breakthrough. The compute existed. What did not exist was a conceptual model for how a non-specialist should interact with a computer.
Engelbart, Kay, the Xerox PARC team — they invented a vocabulary. The mouse. The window. The icon. The desktop metaphor.7 None of it was inevitable. All of it had to be designed, defended against engineers who thought it was unnecessary, and pushed through against organizations that did not see why it mattered. That work unlocked computing for two billion people who would never have used a command line.
The post-prompt AI interaction paradigm is structurally the same moment. The capability is here. The conceptual model for how humans actually work alongside intelligent systems is not. Whoever defines that model will shape the next two decades of work.
The people who defined the GUI were not the engineers who could have. They were the designers, cognitive scientists, and HCI researchers who insisted that interaction design was a legitimate discipline. The post-prompt interaction language will be defined the same way. By whoever takes it seriously enough to do the work.
What this means
For strategic designers: the most leveraged work over the next five years sits at the intersection of AI capability and organizational interface. Not building the next AI product UI. Designing the interaction grammar that lets organizations actually use AI without breaking themselves. The 88% pilot failure rate is your brief.
For transformation consultants at BCG, McKinsey, Bain, Accenture, and the firms now building AI practices: the pattern is the same. Your clients' adoption failures are not engineering problems. They are interface and organizational design problems. Teams that bring design leadership to the table — not as visual support for the deck, but as a primary discipline — will deliver outcomes the others cannot.
For CXOs making AI investment decisions: stop measuring AI ROI by deployment count. Start measuring it by quality of integration into actual work. The gap between 97% personal benefit and 29% organizational ROI does not close with more licenses. It closes when someone designs how the work actually changes.
For everyone: the AI implementation gap is one of the most consequential design problems of the next decade. It is also one of the most open. The current consulting wave is solving the immediate problem at considerable expense. The interaction paradigm that will make that consulting wave obsolete has barely started being designed.
That work has to start somewhere. I am writing this because I would like to find the people who want to start it.