It has been three years since the AI wave started reorganising enterprise budgets, and the picture has finally clarified enough to draw a map. Five distinct delivery models are competing for the same dollars. None of them is yet the dominant winner. The question of who actually delivers production AI to the businesses that need it most is still open.
The five delivery models, named
The five models are worth naming explicitly, because most conversations in this space proceed as if there are only two or three.
The first is Big-4 transformation. McKinsey QuantumBlack, Accenture's AI practice, Deloitte AI & Data, BCG X. The shape of the engagement is familiar: a partner-led roadmap, opportunity prioritisation workshops, a pilot proposal at month six. The deliverable is a recommendation. The customer they sell to is a global enterprise with an in-house technology function and an offshore engineering capacity that can take the recommendation forward.
The second is the AI boutique. Faculty in the UK, Mosaic, Plan A Technologies, a growing class of fifteen-to-fifty person firms staffed by ML engineers who will build something custom. The model is bespoke. You describe the problem precisely, they spec a system, six months later they hand it back. The customer is a tech-forward leader who can specify the problem precisely, has internal engineering to maintain the result, and can wait six months.
The third is the bolted-on SaaS copilot. Microsoft Copilot for Microsoft 365. Salesforce Einstein. Glean, Writer, Notion AI. The model is software-shaped: per-seat fee, generic AI capability layered on top of an existing application. The customer is one whose primary work happens inside the application that gets the copilot. This works well for documents, decks, and spreadsheets. It works less well for the work that happens between systems, which is most of the work that needs automating.
The fourth is the in-house AI team. Hire two ML engineers and an MLOps person, give them eighteen months, deploy something. This is the model that actually works at scale, which is why the largest companies all default to it. It assumes a customer who can hire ML engineers at four-hundred-thousand-dollar fully-loaded compensation, can offer them a problem interesting enough to choose your insurance brokerage over a frontier lab, and can wait eighteen months for the first thing to ship.
The fifth and newest is the AI-enabled roll-up. Buy a fragmented services market at services multiples, deploy AI to compress costs, exit at software multiples. General Catalyst earmarked $1.5B from an $8B fund for the strategy in late 2024. Thrive Holdings committed $1B+ in April 2025. Lightspeed and Bessemer are in. Total capital deployed across the strategy now exceeds three billion dollars.
None of the five fits the largest segment
Each of these five models is well-shaped for a customer it serves well. None of them is well-shaped for the business that needs production AI but does not have an AI team, which is the largest unaddressed segment in the market.
How the Big-4 firms are restructuring under AI pressure
The Big-4 firms know their model is breaking, and they are restructuring under the pressure faster than most observers credit them for. McKinsey now reports 25 percent of projects priced against outcomes rather than time. Lilli, the firm's internal LLM-powered tool, is in daily use by 70 percent of consultants. Five thousand support roles have been eliminated, replaced by AI doing the work those roles used to do. Forty percent of revenue is now classified as tech-enabled services. PwC cut graduate hiring by 30 percent over three years. EY has hired 61,000 technologists since 2023, roughly fifteen percent of the workforce, and is openly exploring what it calls service-as-a-software. The transformation industry is transforming itself, and the directional read on the data is unambiguous.
The boutiques are scaling but staying small relative to the demand. The SaaS copilots are succeeding at the easy twenty percent of horizontal use cases (document Q&A, meeting summarisation, drafting assistance) but have not penetrated the operations layer where most of the actual work sits. The in-house teams remain a Fortune-500 phenomenon, because the labour-market math does not work for everyone else. The roll-ups are early enough that the financial thesis is unproven, and the natural-experiment from Concentrix's deep AI deployment (still trading at single-digit EBITDA multiples after deploying gen-AI to a thousand-plus customers) suggests the multiples may not transfer the way the playbook assumes.
The customer in the middle, between every model
What none of these five models has yet shown is a path to delivering production AI to non-tech operating businesses at the scale the market needs. The integrators handle large volumes of standardised deployments inside enterprises with technology teams. The boutiques handle small numbers of bespoke builds for the most sophisticated customers. The copilots handle the easy horizontal cases. The in-house teams handle the customers who can hire them. The roll-ups handle the customers who got bought. The customer in the middle, the one running on QuickBooks or NetSuite or a legacy ERP with a manual process layer that consumes a meaningful share of employee time, falls between every model.
This is not a complaint about the existing models. Each of them serves a customer well. It is an observation that none of them serves the largest segment, and that the segment in question is the one with the most to gain from the technology actually being installed.
The next several posts work through what the shape of a delivery model has to be to reach this customer. The argument starts from a thirty-five year old paper, runs through the economics of consulting under AI pressure, picks up Diogo Santos's framing of the third layer, and ends with the operational model that serves the segment the existing five do not. If you only read one post in the sprint, the next one is the one to read. It is the foundational piece. Everything else is downstream of it.