The future of enterprise AI isn't build vs. buy

A firm and its opposite numbers

Latham & Watkins — by revenue the world's second-largest law firm (approximately $8.3 bn reported for 2025, per Legal IT Insider) — signed an enterprise license for firmwide rollout of Harvey in August 2025, making the platform available to all of its 3,600+ attorneys (the firm's own press release). In September 2026, the Financial Times and Bloomberg Law reported something pointed in the other direction: the firm has spent roughly three years buying its own Nvidia servers — multiple H200 GPUs in a third-party data center facility that only Latham employees can access — to fine-tune open-weight models and run selected AI workloads on infrastructure it controls.

The apparent contradiction is the interesting part. Why does a firm that bought the category-leading legal AI product also build a private AI stack? And the firm answered it directly: Rubin, who chairs Latham's internal AI strategy committee, told Bloomberg Law the server buildout is not intended to lessen reliance on Harvey — it provides “capabilities that Harvey simply doesn't offer … They're in the cloud and this is on premise.” The strategy isn't build versus buy. It is build and buy and integrate and train, layered against what each workload demands.

Why the two strategies are one strategy

Read the reporting as a stack and the logic becomes ordinary. Commercial AI products deliver mature capability immediately — research, drafting, document analysis, purchased as a subscription with no engineering payroll behind them. Proprietary infrastructure addresses the other half of the ledger: client data that the firm decides should not sit with “any cloud vendor” (CIO Rene Mendoza's words to the Financial Times), integration of AI into internal systems the way the firm wants them integrated, and tools that encode its own workflows rather than a vendor's generic ones.

Both halves are rational at the same time. Buying is faster and cheaper where the vendor's product already does the job. Owning is justified where confidentiality, integration depth, or a differentiated process makes the generic product the wrong shape. The decision question moves from “which AI product should we buy?” to what should our business actually be capable of doing? — and the answer determines which layers to own.

What Latham's stack actually shows

The lesson for smaller businesses is not “buy your own GPUs”

Copy the reasoning, not the invoice. No law firm, and no plumbing company, has Latham's capital, 900-person technology department (~100 dedicated to AI, per Bloomberg Law), or in-house ML engineers. Penn State's Daryl Lim made the necessary caveat to Bloomberg Law: owned infrastructure only makes sense if it stays busy, and buying, maintaining, and securing it requires technical staff. The transferable lesson is the decision discipline: know what your business must be capable of, then assemble the cheapest safe combination of layers that delivers it.

For a small or mid-sized business, the workflow worth owning rarely involves a GPU cluster. It looks like this: a lead arrives → AI responds → the lead is qualified against your criteria → your CRM gets the full record → an appointment is offered from your calendar → follow-up is triggered → a human is alerted when judgment is needed. That is an integration and automation problem, not a model-training problem. The customer in that business doesn't need to understand models, APIs, vector databases, or orchestration layers. They need the workflow to work — and the owner needs to know what it earns. KAVO's ROI method is how we keep that honest: cost savings and recovered revenue measured from your data, reported separately.

The implementation gap

KAVO works in the gap this story exposes every day. On one side sits off-the-shelf software — fast, cheap, and generic, yours until it hits a wall your business can't go around. On the other sits an internal AI engineering organization — the Latham option, which for most companies is neither fundable nor worth funding. In between is the actual work: start from the workflow, then decide the appropriate mix of existing software, AI models, automation, integrations, custom development, private infrastructure where the data justifies it, and human oversight at every step that carries business risk.

The operating principle follows directly from the Latham pattern, and KAVO claims no invention in it: the workflow determines the technology, not the other way around. If an existing product solves the problem well, use it. If several systems must connect, integrate them. If a workflow needs a capability nobody sells, build the missing layer. If confidentiality or control demands private deployment, evaluate that architecture honestly — KAVO's Private tier exists for exactly that case. Model-agnostic, tool-agnostic, and deliberately so: the thing being bought is the operational outcome, and good engineering should not care whose brand is on the component.

You don't need Latham's AI budget to think like Latham

What Latham has is not a secret — it is a technology function that knows the firm's work, decides layer by layer what to own and what to buy, wires it together, and trains people to operate it. The scale is impossible to imitate; the structure is not. A growing business can get the same structure without the payroll by having a technical partner whose job is exactly that decision process: what should be bought, what integrated, what automated, and what — if anything — custom-built.

That is KAVO's business model: an implementation layer between a company's operations and a fast-shifting AI ecosystem, held together by someone accountable for the outcome. To keep the comparison honest about what is and isn't the same: Latham's clients pay for legal work; KAVO's clients pay for workflow outcomes measured before and after (how we work). The discipline — own the business outcome, treat AI capabilities as components — is the part that scales down.

Who sells what

1. An AI SaaS vendor sells you its product. The roadmap decides what you can do.
2. Traditional automation connects predetermined applications. Useful plumbing, but the workflows are drawn by someone else.
3. Consulting can hand you the right strategy deck — and leave you holding it.
4. KAVO identifies the operational problem, designs the system, integrates the appropriate technology, deploys it, and operates the result — reporting what it saved in your dollars, per the contract.

The sentence worth stealing

The businesses that benefit most from AI may not be the ones that buy the most AI products. They will be the ones that integrate AI most effectively into how their business actually operates. Latham's servers are a striking illustration, at a price only Latham can pay; the principle underneath them costs nothing to adopt. So before the next AI tool purchase: name the workflow you want to improve, and what improvement is worth. If the answer is real, talk to KAVO — you get a written diagnosis with your numbers, yours to keep either way.