A frontier model can pass the bar exam, write production code, and reason through a differential diagnosis. It still cannot tell you which of your four CRMs holds the real customer record, who signs off on a refund over $5,000, or why the accounts payable team keeps a spreadsheet that overrides the ERP. That gap, between raw capability and a working system inside your business, is the whole game now. Drew Bredvick put it cleanly in AGI: Some Assembly Required: the models are already superhuman on smarts, but smarts has never been the limiting factor on output. Action is. And action requires assembly.
This is the uncomfortable truth behind the AI adoption numbers. The capability is showing up faster than any enterprise can absorb it. The bottleneck moved. It used to sit inside the model. Now it sits inside your org chart, your data plumbing, and your security review queue.
The Assembly Problem
Bredvick's sharpest analogy is historical. The technology that defined the 20th century was not the electric motor itself. It was the 40-year build-out that rewired factories around it. Electricity was commercially viable in the 1880s. Most American factories did not see real productivity gains from it until the 1920s, because capturing the value meant tearing out steam-driven line shafts and redesigning the entire floor around distributed motors. The bottleneck was never the dynamo. It was everything around the dynamo.
AI is following the same curve. A capable model is the dynamo. Your business is the factory still laid out for steam. Dropping a brilliant agent into a process built for human handoffs, tribal knowledge, and undocumented exceptions does not produce a 10x outcome. It produces a confused pilot that technically works in the demo and quietly dies in production.
The data backs this up. The MIT Media Lab's State of AI in Business 2025 report found that roughly 95 percent of enterprise generative AI pilots delivered no measurable return. Not because the models were bad. Because the integration around them was missing. Gartner has separately projected that over 40 percent of agentic AI projects will be scrapped by 2027, citing cost, unclear value, and inadequate controls. The pattern is consistent: the model works, the deployment does not.
Key takeaway: Capability is no longer the constraint. The constraint is the multi-year, unglamorous work of rewiring a business around that capability, and that work does not happen on its own.
Why the Labs Are Hiring Forward Deployed Engineers
Here is the tell that should end the debate. If smart models deployed themselves, the companies building the smartest models would not be staffing up armies of people whose entire job is deployment.
They are. OpenAI runs a dedicated forward-deployed engineering team with dozens of open roles. Anthropic is hiring aggressively into applied AI and solutions functions. Joe Schmidt's a16z piece, Trading Margin for Moat, documents how category-defining AI companies are deliberately trading early gross margin for an implementation moat. Decagon staffs Agent PMs. Sierra and Harvey put engineers on-site. This is the same playbook Salesforce, ServiceNow, and Workday ran during the cloud transition, when professional services revenue was the price of getting software to actually stick.
The forward deployed engineer concept did not start with AI. Palantir built its business on it: send engineers to live inside the customer's problem, against the customer's real data, rather than shipping a product and hoping. The reason the role is now the hottest job in startups is that the AI deployment gap is wide and the people who can close it are scarce.
Read the signal plainly. The foundation-model labs do not believe enterprises will deploy themselves. They are standing up services arms and partnering with private-equity-backed integrators because they have seen what happens when you hand a powerful model to a buyer with no assembly plan. It sits on the shelf.
The Five Ways Unassisted Enterprise AI Stalls
When an internal AI initiative dies, the postmortem rarely blames the model. It blames one of these five failure modes, and usually several at once.
- Data is fragmented across systems of record. The customer lives in three CRMs, the order history is in the ERP, and the support context is in a help desk that nobody has API access to. An agent is only as good as the data it can reach, and reaching that data is an integration project before it is an AI project.
- Workflows are tribal and undocumented. The "process" exists in the head of a senior operator who knows the seven exceptions that matter. You cannot automate a workflow nobody has written down, and the act of documenting it is half the work.
- Compliance, audit, and security review block rollout. A pilot that works in a sandbox hits a wall the moment it touches production data. Without audit trails, access controls, and a security review that the agent can actually pass, legal stops the launch.
- Middle managers resist agents that absorb their team's scope. This is the failure mode nobody puts in the deck. When an agent does the work of three analysts, the manager of those analysts has every incentive to find reasons it is not ready. Change management is people management.
- Nobody owns evals, observability, or escalation after the consultants leave. The pilot ships, the contractors roll off, and six weeks later the agent is drifting with no one watching. There is no eval suite, no dashboard, no defined escalation path. The system rots.
Klarna is the public cautionary tale. After loudly replacing support staff with AI, the company walked it back and rehired humans when quality slipped. The lesson is not that AI support fails. It is that AI support without the assembly - the guardrails, the escalation paths, the human handoff - fails.
Key takeaway: None of the five failure modes is solved by a smarter model. Every one is solved by integration, documentation, controls, and an owner who stays past launch.
Your Three Options, and Their Tradeoffs
Once you accept that AI needs assembly, the question becomes who does the assembling. There are three real paths, and they are not close on the dimensions that matter.
| Dimension | In-House FDE Team | Big Four / Systems Integrator | AI-Native Partner |
|---|---|---|---|
| Time to first agent | 3 to 6 months (hiring) | 2 to 4 months (scoping) | 2 weeks |
| Cost model | Loaded salaries + ramp | High day rates, large teams | Flat monthly retainer |
| AI-native depth | Depends entirely on who you hire | Generalist, retraining onto AI | Built on production agents |
| Change-management coverage | You own it | Process-heavy, slow | Embedded, workshops bundled |
| Ownership at handoff | Native, but you built the team | Often locked to their stack | Full client ownership |
| Velocity | Slow until staffed | Low, governance-driven | High, ships weekly |
The in-house route is the right answer if you are large enough to staff a permanent team and patient enough to spend two quarters hiring forward deployed engineers in a market where OpenAI and Anthropic are outbidding you for the same people. Most companies are not.
The Big Four route buys you audit credibility and scale. It also buys you high cost, low velocity, and a partner that is retraining a generalist consulting workforce onto AI in real time. IBM and Accenture are rebuilding their services arms around AI delivery precisely because the old model did not fit, but you are paying for that transition.
The AI-native partner route exists because of the gap between those two. The premise: a small senior team that has already shipped production agents, embeds with your operators, ships a working pilot in two weeks, hardens it through production, and hands you a system you own. It is the forward deployed model without the multi-quarter hiring problem.
What an AI-Native Partner Actually Does, Week by Week
"Services partner" is a vague phrase, so here is what it looks like concretely. The cadence is embed, build, harden, go live, and it runs in about four weeks for a focused workflow.
- Week 1, Embed. Sit with the operators who actually do the work. Map the process, find the seven exceptions, identify the systems of record, and surface the compliance constraints early. This is where the tribal knowledge gets written down.
- Week 2, Build. Stand up the agent against real data, with guardrails and human-approval checkpoints designed in from the start, not bolted on later. Ship something an operator can use, not a demo.
- Week 3, Harden. Add evals for your specific failure modes, observability so someone can see when the agent drifts, and defined escalation paths to a human. Pass the security review.
- Week 4, Go live. Production release with a real ownership handoff: documentation, CI/CD integration, and an internal owner trained to run it after the partner steps back.
The part most engagements skip, and the part that decides whether the system survives, is re-skilling. If the people whose work the agent touches do not understand it and cannot operate it, you have bought a dependency, not a capability. A real partner bundles education and workshops so the transformation is org-wide, not a black box.
How OpenNash Can Help
OpenNash is built for the assembly problem. The team brings eight years of data and AI engineering from Silicon Valley, runs a deliberately limited client roster so the work stays senior-led, and operates on the embed-build-harden-live cadence above. The pilot runs the first two weeks at no charge, billing is a flat monthly retainer rather than per hour, and every engagement targets three measurable outcomes: 10x more revenue, 10x lower cost, or 10x time saved.
What that looks like in production:
- A 30-attorney firm deployed a voice agent for intake and routing. Booked-consult rate moved from 41 to 67 percent, with conflicts checks and an audit trail built in.
- A $40M services business automated accounts payable. Invoice cycle time dropped from 11 days to 3, with human approval on anything over threshold.
- A 50-person ops team replaced a manual monthly board pack. Nine hours of analyst work became 12 minutes, with the underlying data fully owned by the client.
The honest version of the buying decision: if you have the scale and patience to build an in-house forward deployed team, build it. If you need audit theater for a board, the Big Four will give you that. If you want a custom system that fits your actual workflow, ships in weeks, and leaves you owning a maintainable asset, that is the AI-native case, and it is the one OpenNash makes. See AI agent consulting and AI agent pricing for the full model, or CX if the workflow is customer-facing.
The runway here is long. Bredvick is right that AI is a continuation of the connected-computing mega-trend, a multi-decade build-out rather than a weekend install. The capability will keep arriving on schedule. Whether your business captures it depends entirely on the assembly, and the partner you choose now sets how much of that runway you get to keep.
Book a call to map one stalled workflow to a two-week pilot.