Goldman Sachs now deploys thousands of autonomous AI agents alongside its developers. McKinsey finds only 5.5% of companies see real results — because the gap isn't about tool access, it's about the infrastructure behind the tools. The good news for mid-market: that infrastructure is available today.
The Number Worth Reading Twice
In spring 2025, Goldman Sachs CEO David Solomon said something at a conference that barely made a ripple in the financial press — but deserved a bigger headline: AI can now produce 95% of an IPO prospectus in minutes. The same work used to keep a six-person team busy for two weeks.
"The last 5% now matters," Solomon said, "because the rest is now a commodity."
That's not a vague statement about AI's future. It's a description of what's happening at Goldman Sachs right now. And Goldman isn't alone. Bank of America hit 90% AI adoption across its workforce by April 2025 and reports more than 20% efficiency gains for developers. Citigroup rolled out GitHub Copilot to 40,000 developers, measuring productivity improvements between 2x and 20x on agentic tasks. In July 2025, Goldman CTO Marco Argenti announced plans to deploy thousands of autonomous AI agents alongside the firm's roughly 12,000 human developers, with an expected productivity lift of 3x to 4x.
Klingt interessant?
These aren't experiments anymore. This is production.
What Enterprises Actually Built
The difference between large corporations and everyone else isn't access to AI tools. ChatGPT, GitHub Copilot, and similar products are available to any company with a credit card. The difference is the infrastructure that makes those tools actually productive.
McKinsey's 2025 State of AI report — based on responses from nearly 2,000 organizations — puts a number on it: 79% of organizations say they're using generative AI, but only 5.5% are generating measurable financial returns. What separates top performers from the rest? They've fundamentally redesigned their workflows. Not with ChatGPT that someone queries step by step, but with AI agents that autonomously handle multi-stage tasks — embedded in real development processes.
Goldman Sachs, JPMorgan, Citigroup: over the past 18 months, these organizations built what it actually takes to get there:
- Multi-agent pipelines — not individual Copilot seats, but networks of specialized agents that write, review, test, and deploy code
- Quality gates — automated validation that checks AI output before it reaches production
- Audit trails — complete traceability of every AI action, for compliance, regulation, and internal oversight
- EU data residency and security — in regulated industries, not a nice-to-have but a hard requirement
- Human-in-the-loop — processes where developers stay in control and actively approve AI output
Building this infrastructure took time, budget, and specialized teams. That's the real competitive advantage enterprises hold — not the AI access itself.
The Gap Mid-Market Companies Are Underestimating
According to IDC, 83% of organizations with more than 5,000 employees already have AI in production — compared with 42% of companies with 50 to 499 employees. That sounds like a problem. It's actually an opportunity.
Mid-market companies aren't trapped in the same legacy structures that make AI transformation so expensive for large organizations. Shorter decision cycles, less organizational resistance, more genuine willingness to implement new processes — those are real structural advantages.
The core challenge for mid-market CTOs is rarely a lack of appetite. It's the missing infrastructure. A GitHub Copilot seat gives individual developers 10–30% productivity gains — but no systemic impact on delivery as a whole. Comparing that to what Goldman Sachs is running today is like comparing a torch to a floodlight. Both emit light. Just at completely different orders of magnitude.
McKinsey's data is unambiguous on this point: companies that genuinely scale AI agents and redesign workflows from the ground up are a small minority — and they represent the 5.5% that see real results. Fewer than 10% of surveyed companies deploy AI agents at scale in any function at all.
The Head Start You Don't Have to Build Yourself
Here's the good news: what took enterprises 18 months and significant internal resources to build is available today as ready-made infrastructure.
That's the model behind Nopex: multi-agent development workflows that are production-ready from day one. No need to hire internal AI specialists, no months-long pilot phase, no building compliance structures from scratch. Data processing in the EU, full audit trails, human-in-the-loop as the default process — not as a retrofit.
A mid-market company can start its first sprint with infrastructure that a large corporation spent 18 months assembling.
The gap between companies that use AI as a tool and those that run it as infrastructure will show up in hard numbers in 2026 — in delivery speed, team capacity, and cost. That gap widens with every quarter that agentic workflows are refined on one side or the other.
The question is no longer whether. It's when.
Start Now, Don't Watch from the Sidelines
The companies generating measurable results from AI today don't have more tools than others. They built their processes differently. For mid-market companies, that means not waiting for the next generation of models — but starting now with what enterprise organizations have already validated.
If you want to know which development processes in your organization are best placed to scale first, reach out. No sales pitch — concrete insights from projects already running.
