We’re Common Works.
We empower PE sponsors and their portfolio companies to leverage AI to fundamentally improve how they run their businesses. Unlike traditional consulting firms, we don’t focus on delivering decks and high-level strategy. We’re a full life-cycle partner. We work directly with teams to identify core operational friction and build the precise AI tooling and workflows to solve it.
We tell our clients that we operate as their AI General Contractor: we help them create the vision, design the implementation blueprint, and execute the build.
Moving forward, we’ll use this platform to share our active learnings and build and scale in public. We firmly believe the industry is only at the beginning of this shift.
The Hidden Vendor Trap
Most AI decisions inside a private equity-backed company arrive dressed as a simple vendor question: Which model? Which platform? Which copilot? How many seats?
Unfortunately, that framing suits the vendor’s sales cycle, not the sponsor’s hold period. It defers the most critical strategic decisions. When the implementation work is finished, part of the new capability belongs to the company, and part belongs to a third party. That split is decided early and quietly – often by people who don’t realize they’re deciding it.
Let’s start at the exit, because that’s where this split gets priced.
Buyer diligence now rigorously tests “AI Readiness.” Institutional buyers increasingly distinguish between operational capability that transfers with the business and capability that depends entirely on vendor continuity. A buyer views an internal capability as proprietary IP that compounds enterprise value. Conversely, heavy vendor dependencies carry ongoing negotiation friction, rising renewal costs, and – especially with AI startups – acute business continuity risk.
Own What Compounds
Sovereignty doesn’t mean building everything from scratch, and it has nothing to do with running physical GPUs on-premise. In our experience, four distinct layers compound as a company integrates AI:
- Proprietary Data: The unique, historical records only your business generates.
- Workflows and Processes: The distinct operational logic of how the work actually gets done.
- Institutional Memory: The collective operational lessons captured systematically rather than re-learned by churned staff.
- Encoded Judgment: The nuance of your best people’s decisions, captured so the enterprise can replicate them at scale.
Conversely, we advise our clients to be deeply mindful of three things that don’t compound, where renting is unequivocally the better choice: model selection, raw compute, and generic tooling.
In practice, the line between what you own and what you rent falls in a very specific place. You must own the code, the data layer, the lineage, the connectors, the schemas, and the documentation. The model sits above that line as a highly swappable dependency. Changing it should be an optimization exercise, never an architectural rewrite.
Admittedly, what sits below that line is usually a mess, which is why teams avoid it. Most middle-market portfolio companies run on data that’s inconsistent, undocumented, or fragmented across legacy systems. The first stretch of value creation is always cleanup: naming fields, repairing data lineage, and forcing two departments to agree on what a “completed order” actually means. That foundational work dictates whether anything built on top is worth owning at all.
The model layer itself is genuinely rentable, largely thanks to the rise of open-weight models. For the high-volume production workloads that drive the most business value – meaning extraction, classification, routing, and summarization – open-weight models routinely do the job at a fraction of frontier API costs. Crucially, they can run entirely within a cloud environment the company controls, ensuring sensitive records never leave your security boundary.
The Real Switching Cost is the Eval Suite
We’re in the early days of a paradigm shift. Much like the search engine wars of the Internet 1.0 era, no single AI lab stays on top for long. The ground under any single-vendor bet is constantly moving.
Model economics are also moving too quickly to underwrite today’s pricing as permanent. It’s incredibly difficult to underwrite a reliable terminal value for a business when you can’t accurately forecast future token costs. While consensus says costs will plummet with data center build-outs, a sponsor must still underwrite when that efficiency materializes. It reminds us of 2015, when an Uber from Manhattan to Brooklyn cost $15 – a heavily subsidized rate that simply couldn’t last.
While cost is a variable exposure, the true danger is architectural dependency.
Anyone who’s moved a production workload between models knows that changing the API call is the trivial part. The real cost is re-validating behavior: regression-testing prompts, rebuilding evaluation sets, and hunting down the quiet places where output quality depended entirely on one model’s specific habits.
Because of this, we build the evaluation suite first.
An institutional eval suite ensures that changing a model provider is seamless. Without one, switching models requires weeks of nervous, manual checking – completely negating the efficiency gains that justified implementing AI in the first place.
When an AI implementation firm standardizes its practice on a single LLM lab, your switching cost becomes their moat. Common Works holds no equity in any model provider and takes zero reseller margins. That structure costs us the recurring revenue line much of our category is built upon, but it buys us the radical independence required to tell a client when a cheaper, open-source model is good enough.
Navigating Organizational Realities
The live decision in most portfolio companies isn’t which niche AI lab to standardize on. The reality is that the business already runs on the Microsoft stack, IT security has cleared Copilot, and the CFO refuses to onboard a second vendor.
That resistance is organizational, not technical. Arguing with it is a waste of everyone’s time. So, don’t.
Seat licenses are perfectly fine for what they do. The far more important architectural question is: Where did the assistant get its facts?
If the underlying records and operational rules live inside a proprietary vendor subscription, the company is renting its core workflows. But when those records live in a data layer the company completely controls, the AI assistant is merely an interface – one that can be replaced on a standard procurement cycle.
Five Questions for Your Next Board Meeting
- Which of our operational advantages compound as we use AI, and which are we quietly renting back from vendors?
- If our primary model provider changed its pricing, data policy, or availability tomorrow, how many weeks would it take us to switch?
- When we terminate a vendor contract, does our data, fine-tuning, and workflow knowledge come with us, or does it stay with them?
- Are our operating metrics measuring actual business outcomes, or are we simply tracking token consumption?
- What exactly did we build this year in an architectural layer that the company completely controls?
Sovereignty isn’t a top-down corporate program that a company launches. It’s a value creation metric that accumulates, one build at a time, and resolves into a single form at exit.
One company hands the next buyer a permanent, proprietary capability that easily survives a change in vendors. The other hands over an expensive dependency and hopes the vendor economics hold.