The Software Factory Is Coming for Your $40 Million SOW

By Errin O'Connor — Founder & Chief AI Architect, EPC Group · Microsoft Solutions Partner across all six designations · G2 Leader in BI consulting

Ethan Mollick recently profiled something that should be circulating in every procurement department in the Fortune 500, and as far as I can tell, it is circulating in none of them. A three-person team at StrongDM, a security software company, built what they call a Software Factory: human product leaders write the roadmap, coding agents build the software, testing agents attack it inside a simulated customer environment the testing agents themselves construct, the two agent populations loop against each other until the output holds — and humans review the finished product and ship it without anyone ever touching, or even reading, the underlying code.

In the same essay cycle, Mollick cited an agent running autonomously for fourteen hours and producing a software package a human engineer would have needed weeks to build — at a token cost of roughly $251. Weeks of engineering labor. Two hundred fifty-one dollars. I will give you a moment with that ratio, because the rest of this article is about what it does to a number you are far more familiar with.

The 100,000-hour question

Enterprise blended rates on large-integrator engagements sit in the $300-to-$500-per-hour range once the partners, managers, architects, offshore delivery, and specialty subcontractors are averaged together. I am not theorizing — I have reviewed these proposals and redlined these statements of work for the better part of three decades, and public GSA schedules corroborate the band. A 100,000-hour engagement — completely ordinary for a two-year AI, Fabric, or Microsoft 365 transformation at a Fortune 500 — prices between $30 million and $50 million.

Now hold the two numbers next to each other. The integrator model prices software delivery at hundreds of dollars per human hour, across hundreds of humans, for years. The factory model just demonstrated weeks of equivalent output for the price of a nice dinner. I am not claiming those two numbers describe the same work — they don't, yet, and I will be honest about the gap in a moment. I am claiming something more uncomfortable: the large-integrator business model is a bet that they stay different, and every quarter of agent progress is a vote against the bet. Meanwhile Anthropic has disclosed that roughly two-thirds of the code on its own product team is now written by its agents. The factory is not a demo. The factory is how the frontier labs already work.

What the SOW is actually selling you

Be precise about what a 100,000-hour engagement is made of, because this is where the disruption lands unevenly. Some of those hours are judgment — architecture decisions, regulatory interpretation, stakeholder navigation, the gray-haired call that saves a program. Agents are not close on those, and anyone who tells you otherwise is selling something. But an enormous share of the hours in every large SOW I have ever redlined are not judgment. They are production: scaffolding code, migration scripts, test suites, documentation, status decks, the fourth round of formatting on the steering-committee readout. That is exactly the layer the factory just industrialized. When the production layer collapses in cost, the pyramid that priced it — armies of associates billed out at hundreds an hour, learning on your dime — collapses with it. What remains valuable is the top of the pyramid: the named senior people. Which raises the question every CFO should now put to every incumbent: how many of my 100,000 hours were judgment, and how many were production billed at judgment rates?

And there is a second cost the SOW never itemizes: the perimeter. Every additional human in the delivery pyramid is another credential against your tenant, another laptop, another subcontractor chain, another offshore handoff, another token provisioned in 2023 and never rotated. I wrote at length about what that perimeter looks like when it fails — the analysis is here: the delivery-partner vendor-risk breakdown. The factory model and the specialist model share a property the pyramid can never have: a small number of named humans, and a heavily automated production layer that carries an identity, a log, and a kill switch instead of a badge and a plane ticket.

The honest caveats — because the hype will skip them

Now the fairness counterweight, because I have zero interest in selling you the inverse hype. The StrongDM factory works because three unusually strong engineers wrote unusually good roadmaps and built unusually good verification loops — the humans moved up the stack; they did not leave the building. The $251 build has not been independently replicated, and nobody publishes the failure rate on fourteen-hour autonomous runs. Agent-written code still needs security review, licensing review, and architectural coherence over years, not sprints. And an unsupervised agent fleet inside a regulated enterprise without identity, logging, and autonomy boundaries is not a factory — it is a liability with a burn-down chart. The factory does not eliminate senior judgment. It concentrates everything on it.

Which is exactly why this is not a story about big firms versus small firms. It is a story about pyramids versus specialists. The pyramid's economics require the production layer to stay expensive. The specialist's economics improve every time it gets cheaper — because the specialist was always selling the judgment, not the hours.

What I tell clients to do

One. Take your largest active SOW and sort the hours into judgment and production. Ask the incumbent to do it with you. Their enthusiasm for the exercise is itself a data point.

Two. Demand agent-leverage transparency in every new proposal: which workstreams will use agentic delivery, with what human verification, priced how. A firm using agents and billing pyramid rates is charging you for headcount it no longer deploys.

Three. Pilot the factory pattern on one internal workstream — roadmap in, verified software out — under your own tenant, your own identity model, your own logs. Learn the economics before your vendor's pricing team does.

Four. Re-read your delivery perimeter. Fewer humans with credentials is not just cheaper — it is safer, and the specialist-versus-global comparison walks the whole argument.

The SOW audit framework

For each workstream in your engagement, evaluate: total hours, judgment percentage (architecture, regulatory interpretation, stakeholder navigation), production percentage (scaffolding, scripts, tests, documentation, status decks), whether the workstream is agent-leverageable, and whether the billed rate reflects the actual role performing the work.

The outputs of that exercise reveal three things: the dollar figure represented by production hours billed at judgment rates, the renegotiation targets for your next contract cycle, and the workstream that is the best pilot candidate for an internal factory pattern (roadmap in, verified software out, under your tenant identity and logging model).

Where I land

Twenty-nine years ago I watched enterprises pay integrators to hand-build what platforms were about to make configurable. The winners were not the companies that hired the biggest army. They were the ones that noticed the ground moving and repriced their contracts before their vendors repriced their expectations. The software factory is that moment again, at ten times the speed. The 100,000-hour SOW on your desk was priced for a world where production hours were scarce. That world has roughly the shelf life of your current contract term. The firms that thrive will be small, senior, named, and heavily agent-leveraged — with every agent carrying an identity, a log, and a kill switch. That is not a prediction about EPC Group. That is a description of it.

Multiple models. One truth.

The data behind this (sources and verification)

• StrongDM Software Factory (via Mollick, “The Shape of the Thing”) — Roadmaps in, agent-built and agent-tested software out, humans review, no human touches code. Reported figures; attribute as Mollick’s.
• Mollick-reported autonomous build figures — 14-hour autonomous build ≈ weeks of engineering at ~$251 in tokens; a quarter of OpenAI staff running 4+ agents weekly. Reported figures; attribute as his.
• Anthropic product-team disclosure — Approximately two-thirds of product-team code is agent-written.
• U.S. GSA Multiple Award Schedule filings — Published senior-category integrator rates into the mid-$300s/hr and higher for principals; blended $300–$500 corroborated by three decades of EPC Group proposal review (directional).
• EPC Group delivery-pyramid credential exposure analysis — Perimeter risk from large delivery pyramids: credentials, subcontractor chains, offshore handoffs. See /insights/delivery-partner-vendor-risk-accenture-2026.

Third-party figures above are attributed to their named sources as of the Last verified date. EPC Group audit figures are directional findings from client engagements. Items marked [VERIFY] must be confirmed before external quotation.

Frequently asked questions

Multiple models. One truth.