Every AI Giant Just Built an Army of Engineers. None of Them Are Coming to Your Business.
If you run a company of two or three hundred people, the last ten weeks have been a strange thing to watch.
Almost every week, another announcement: a frontier AI lab, or one of the largest cloud platforms, committing extraordinary sums to put its best engineers inside customer organisations — to sit next to the business, redesign the workflow, and stay embedded until the AI produces a result. Billions of dollars. Thousands of engineers. All of it pointed at one admission: buying the technology was never enough. Someone senior has to do the hard, unglamorous work of making it deliver.
None of this is news if you have tried it. You licensed the tools, ran a pilot, got a demo that worked and a production system that stalled, and waited for the help that would close the distance. The past ten weeks answer whether that help exists. It does, and it works. It is also being built for companies several sizes larger than yours.
The whole industry moved in a single quarter
Look at the shape of what was announced. Four of the biggest names in AI and the largest cloud platforms committed billions of dollars to one idea: embed senior engineers in the customer, and get paid on outcomes rather than licences. One committed six thousand people to it. One built a standalone company around it with private-equity money. One stood up pods of five or six engineers to live inside a single customer at a time. The term of art is “forward-deployed engineering,” borrowed from a defence-software firm that has embedded engineers in client sites for two decades.
Every serious player moving at once, inside a quarter — several announcements visibly rushed to avoid landing second — is the market repricing where AI value gets created. A structural correction, not a trend to wait out.
This is the industry conceding, in capital, something it spent three years implying was already solved. The model was never the hard part. The last mile is: the messy integration into real data, real workflows, real change management inside an organisation that has other things to do. That is where measured value is won or lost. I have argued exactly this here before — deployed is not delivering, and the gap between the two is a design problem, not a technology one. It is oddly satisfying to watch the people who sell the technology arrive at the same conclusion. The largest of these ventures even brands itself “outcome-driven,” the language of measurable business results turning up at last in the marketing of firms that spent three years selling seats.
I take the point seriously. It is the correct diagnosis. My only quarrel is with who gets the cure.
The long tail of benign neglect
Once you accept the diagnosis, you notice who the economics are built for.
An embedded pod of senior engineers — people who cost well into six figures each, on an engagement measured in months — only makes sense against a contract large enough to absorb them. So the line gets drawn, cleanly, by arithmetic. The largest enterprises get the first-party armies. The private-equity-owned mid-market gets the venture built specifically for it, because there the distribution channel is portfolio ownership: the fund already owns the customers and can route engineers to them at scale. The independent small and mid-sized company — no sponsor, no fund, no systems-integrator budget — sits where it has always sat. Outside the line.
This is the long tail, and its treatment has a name worth using plainly: benign neglect.
The pattern has been sized before. In capital, the same long tail carries an estimated $5.2 trillion in unmet financing across formal small and mid-sized firms in developing markets, with roughly 40% of them credit-constrained, and it falls hardest across the emerging markets where I work: East Asia and the Pacific alone accounts for close to half of the gap. Too small to serve at a profit, so left underserved. AI delivery is repeating the shape of it.
The small and mid-sized business has always been the long tail of enterprise technology, too small to justify white-glove attention and too numerous to ignore entirely. The industry’s answer was self-serve product: here is the tool, here are the docs, good luck. The neglect was tolerable because it was roughly equal. The big enterprise got software and figured it out; so did you, at your scale. Nobody was getting hand-delivered outcomes.
That equality just ended. The giants have declared, with their wallets, that self-serve product was never sufficient for AI, and that you need senior humans embedded in the work. Then they built that capability for everyone except the long tail. The long tail is now denied the exact thing the whole industry just agreed is the thing that matters.
The cost of that exclusion is not marginal. Small-business productivity runs at roughly half that of larger firms, a gap near 70% in emerging markets and 40% in advanced economies. Closing it is exactly what embedded, outcome-focused delivery is built to do. The giants just built the machine to do it, for everyone above the line.
I wrote recently about enterprise AI splitting into two tiers on the model side — high fashion and fast fashion, frontier models for genuine complexity and efficient open models for everything else. The delivery layer is now splitting along the same fault line. The model split at least had an answer for the long tail: cheap, efficient, good-enough models exist so a smaller company can play. The delivery split has no such answer yet. There is couture — bespoke, embedded, human-intensive — and beneath it, for the long tail, nothing built on purpose.
A smaller army is the wrong answer
The reflexive fix is to imagine a scaled-down version of the same thing: the FDE army, cheaper, for smaller companies. That instinct is wrong, and the reason matters.
The forward-deployed model is capacity-heavy and validation-light. You embed a lot of engineering and discover the value as you go: solve it by hand for the first handful of customers, then productise what works. That is sound when the contract is large enough to fund the exploration. The long tail cannot fund exploration. For a company of a few hundred people, one wrong six-figure build is not a learning; it is the year.
The embedded-army model exists to supply engineering capacity. What the long tail lacks is judgement — knowing what is worth building at all.
Judgement at the tip, capacity behind it
So the shape has to invert. Lead with senior validation — a real KPI, an owner who can absorb the change, a way to measure the result — and commit build capacity only to the use cases that survive it. Most will not, and killing them early is the point. This is the discipline I call Solution-Outcome Fit, and it matters more the smaller the company, because a smaller company has less room to be wrong.
That inversion also dissolves the cost structure that prices the long tail out. Leading with judgement means no standing army of engineers on the payroll — and that overhead is what sets the minimum contract size the long tail can’t clear. A senior translation layer sits at the front; lean, specialist build capacity is pulled in behind it for validated work only. A boutique can run that. Embedded-pod economics cannot.
The honest objection is that the giants’ partner ecosystems will close this gap themselves; hundreds of millions are already committed to enabling partners. It doesn’t change the arithmetic. That money flows to the same global integrators whose economics also require large contracts — it speeds up the top of the market and leaves the minimum viable contract exactly where it was. You don’t reach the long tail by discounting couture. You reach it with a model built differently from the ground up.
Two weeks, not twenty
Picture the company from the opening. A few hundred people, one overworked IT lead, an AI rollout that stalled after the pilot, and five ideas for “where AI could help.” The forward-deployed answer is a pod and a multi-month engagement they will never sign, aimed at all five at once. The smaller-firm answer is two weeks spent establishing whether any of the five has a measurable outcome behind it — most will not survive that fortnight, which is the value — then a focused build on the one or two that do, delivered by a small specialist team engaged for that exact scope. They reach a measured result for the price of one validated build, not an embedded army.
If your AI programme has more licences than measured outcomes, you are the long tail. The operative question is not which model to buy or which pod to hire. It is who validates the outcome before you spend.
The neglect isn’t benign anymore
The giants just spent billions of dollars telling the market that the model was never the hard part, that value lives in the last mile, and the last mile takes people. They built the armies for the customers who could pay for them, and drew the line exactly where the arithmetic told them to.
The long tail obviously needs the same thing. The open question is whether anyone builds a delivery model that fits a company that will never sign a twenty-week engagement, or whether “too small for couture, too numerous to matter” quietly hardens into a permanent tier of the AI economy. The neglect was benign while the tools didn’t really work either. It is not benign anymore.
Sources
-
OpenAI, Introducing the OpenAI Deployment Company (May 2026). openai.com
-
Anthropic, Enterprise AI services company — with Blackstone, Hellman & Friedman, Goldman Sachs (May 2026). anthropic.com
-
AWS — $1B Forward Deployed Engineering unit (June 2026), reported by CNBC, 30 June 2026.
-
Microsoft, Microsoft Frontier Company — $2.5B / 6,000 engineers (July 2026). blogs.microsoft.com
-
Google Cloud — forward-deployed engineering expansion, Cloud Next ‘26 + hiring push (April–May 2026). cloud.google.com
-
World Bank / IFC — SME share of business, employment and GDP; MSME finance gap (~$5.2T formal). worldbank.org · ifc.org
-
McKinsey Global Institute (2024) — small-business productivity gap (~half that of larger firms; ~71% emerging / ~40% advanced markets).