The 4 C's of Enterprise AI: The Only Scorecard That Survives the Model Churn

Open any serious business publication this week and you will find a variation of the same headline: enterprise AI is not paying off. The MIT follow-up study says the data was never ready. Gartner has quietly downgraded its agentic-AI forecast for the third quarter in a row. A16z has spent the last month arguing on X that the real bottleneck is "the context economy." Every take is confident, and every take is partial.

The reason the discourse feels muddled is that everyone is holding one leg of the same elephant. The pilot that failed at your bank failed because the semantic layer was thin, and because the guardrails were absent, and because the token bill blew the budget, and because switching providers would have taken a quarter. All four things are true at once. Any framework that only names one of them is going to keep producing surprise post-mortems.

We think the honest scorecard has exactly four columns. Context, Control, Cost, Choice. It is the frame the whole of The Context Advantage is organised around, and after two years of watching teams succeed and fail with it, we have not yet met a serious enterprise AI failure that does not fit inside those four boxes. This essay is the plant-a-flag version: what each C means, how to know if you have it, and where to go deep in the book when a specific pillar is your bottleneck.

Context is the meaning of your business, encoded well enough that a machine can reason with it. It is not the size of the model's context window. It is not the number of tokens you can stuff into a prompt. It is the honest, canonical answer to the question: what does "active customer" mean at your company, and where does that definition live?

For most enterprises the answer is embarrassing. "Active customer" lives in the head of one analyst in finance, one dashboard in product, and a 2023 Slack thread nobody can find. The model, no matter how large, cannot resolve that. Every agent you point at your business is either guessing, or asking the humans, or quietly making it up.

The teams pulling ahead have started to treat this as the first-class problem it is. They are building a semantic layer — a single, versioned, machine-readable set of business definitions the models resolve against. They are encoding an ontology of their five to ten core entities. They are grounding retrieval in that meaning rather than in raw similarity search.

If you want the long version: https://bricksnotes.com/context-advantage/book/chapter-3 is the working definition of context we keep coming back to. https://bricksnotes.com/context-advantage/book/chapter-6 and https://bricksnotes.com/context-advantage/book/chapter-7 are the semantic layer and ontology playbook. https://bricksnotes.com/context-advantage/book/chapter-12 and https://bricksnotes.com/context-advantage/book/chapter-13 are how to make retrieval reflect the business, not the shape of your documents.

If you take one thing from this section: the context layer is the part of your stack a competitor cannot copy by swiping a credit card. Everything else on this list can be bought.

Every governance regime we have inherited from the last twenty years assumes a human at the keyboard. Approvals, audit trails, role-based access, code review — all of them were designed around the pace and blast radius of a person. Agents violate every one of those assumptions. They are faster, cheaper, and willing to do the same wrong thing a thousand times before anyone notices.

Control is the discipline of building a governance surface that matches the new pace. It has three layers. Guardrails at the edge — what an agent is allowed to touch, when, and with whose approval. Evals in the middle — a running record of what "good" looks like on your actual traffic, not a public benchmark. Audit at the base — the trace log that lets you reconstruct, after the fact, exactly what the agent saw and why it decided what it decided.

The eval piece is the one most teams underestimate. Evals are the dashboards of the agent era. If you cannot see regressions before your customers do, you have no idea whether your last prompt change made the product better or worse.

The chapters that go deep here: https://bricksnotes.com/context-advantage/book/chapter-14 is the eval harness playbook. https://bricksnotes.com/context-advantage/book/chapter-15 is the shadow-eval-on-production pattern. https://bricksnotes.com/context-advantage/book/chapter-18 is the guardrail architecture — the explicit gates on writes, sends, and spends that turn an agent from a liability into a colleague.

Capability without control is not an asset. It is a lawsuit waiting to happen. And it is why the second C matters more than the first for anyone deploying to customers.

The dirty secret of most agentic pilots is that nobody knows what they cost. Token bills arrive at the end of the month, latency budgets are guessed at, and cache hit rates are a line item on a slide nobody has revisited since the demo. Then the loop scales, and the economics collapse in a fortnight.

Cost, as a discipline, is the choice to treat tokens, latency, and cache hits as first-class product metrics. Instrument them. Budget them. Review them with the same seriousness a data team reviews query costs on a warehouse. It is not about being cheap. It is about being durable.

The cheapest AI product on the market next year will be the one whose team knew, per feature and per customer, exactly what an agent turn cost — and had a routing layer to keep it there.

https://bricksnotes.com/context-advantage/book/chapter-20 is the deep dive on token math, caching, and model routing. https://bricksnotes.com/context-advantage/book/chapter-21 is the field guide to silent budget leaks — retries, tool-call storms, and runaway agent loops — that eat pilots alive when the traffic pattern shifts. If your finance business partner cannot see your AI cost per action on a dashboard today, that is where the third C starts.

The frontier model market is, healthily, a leapfrog market. Three labs will trade the crown every quarter for the foreseeable future, and the cost curve will keep falling. That is good news for buyers, and bad news for anyone whose product is architecturally anchored to a single provider.

Choice is the fourth C — the architectural discipline of keeping your product portable across models. It is a thin abstraction between your application and the model call. It is prompts stored as data, not baked into vendor-specific SDKs. It is a tool interface that any provider can implement. It is the explicit decision that switching from Provider A to Provider B, when the frontier moves, is an afternoon of work rather than a quarter of rewrites.

Most teams treat this as a nice-to-have. It is not. It is the single architectural choice that decides whether your product compounds through the next three model generations or gets stranded on a checkpoint that stops being competitive.

https://bricksnotes.com/context-advantage/book/chapter-24 is the portability chapter. https://bricksnotes.com/context-advantage/book/chapter-25 is the case study of what a real cross-provider migration looks like when Choice was designed in versus retrofitted under deadline.

The 4 C's overview lens is https://bricksnotes.com/context-advantage/book/chapter-4. The reference architecture that ties all four together is https://bricksnotes.com/context-advantage/book/chapter-27. Between them, they are the shortest path to a shared vocabulary for your next AI planning meeting.

Here is the part that turns this essay from a manifesto into a tool. Print this table. Score your team 0, 1, 2, or 3 on each pillar. Be honest. The exercise is only useful if the score reflects reality, not the pitch deck.

Context — 0: definitions live in people's heads. 1: some definitions are written down, in scattered docs. 2: a semantic layer exists for the top entities and one team uses it. 3: every agent in production resolves through a single, versioned semantic layer that finance, product, and legal all sign off on.

Control — 0: no evals, no guardrails, no audit trail. 1: manual spot-checks and a wiki page called "AI guidelines." 2: a golden-set eval that runs on every deploy, guardrails on the highest-blast-radius actions, traces logged. 3: shadow evals on production traffic, an explicit approval gate on writes/sends/spends, and a weekly human review of the tail of the trace distribution.

Cost — 0: nobody knows the token bill until it arrives. 1: monthly reporting, no per-feature attribution. 2: per-feature cost and latency dashboards, a caching layer, and a routing policy between models. 3: cost per action budgeted per feature, alerting on drift, and a finance partner who can read the dashboard without a translator.

Choice — 0: prompts and tool calls are baked into one provider's SDK. 1: a thin wrapper exists but is untested against another provider. 2: at least one non-trivial feature has been run end-to-end on two providers in the last quarter. 3: switching primary providers is a documented, rehearsed, half-day operation.

Add the four numbers. Twelve is world-class. Eight is a healthy, compounding programme. Four or below is the reason your last board deck about AI landed with a thud. The point of the score is not the number; it is the shape. A team scoring 3-3-0-0 has a very different next quarter than a team scoring 1-1-1-1, even though the totals are close.

To make the framework concrete, walk through a fictional but very typical pilot: a customer-service agent at a mid-sized SaaS company. Same team, same model, same six-week timeline. Four versions of the same pilot, each failing at a different C.

Context-mode failure: the agent confidently tells a customer their subscription is "active" when finance would call it "in grace period." There are three definitions of active in the company; the agent picked the wrong one. No amount of prompt tuning fixes this, because the problem is that meaning was never encoded. This is the semantic-layer problem in https://bricksnotes.com/context-advantage/book/chapter-6.

Control-mode failure: the agent, given tool access to the CRM, updates a thousand records to the wrong status overnight because a subtle prompt change regressed one edge case. No eval caught it because there was no eval. No gate stopped it because writes were ungated. This is the guardrail and eval story in https://bricksnotes.com/context-advantage/book/chapter-14 and https://bricksnotes.com/context-advantage/book/chapter-18.

Cost-mode failure: the agent works beautifully. The bill in month three is nine times the pilot budget because a retry loop, triggered by a rare error, quietly issues seventeen tool calls per user turn on ten percent of traffic. Nobody notices until finance escalates. This is the silent-leak pattern in https://bricksnotes.com/context-advantage/book/chapter-21.

Choice-mode failure: the pilot succeeds and the team ships. Six months later, the provider raises prices thirty percent and deprecates the model version everything was tuned against. Migration is estimated at a quarter. The competitor, whose Choice score was higher, migrates in a week and undercuts on price. This is the portability argument in https://bricksnotes.com/context-advantage/book/chapter-24.

The same pilot, the same model, four failure modes. Any framework that names only one of them was going to miss three of these post-mortems.

Most "AI strategy" documents we have read this year score, if we are being honest, a 1 on every pillar. They have a paragraph about data, a paragraph about governance, a line about budget, and a note about vendor risk. That is not a strategy. That is a checklist that has learned to sound like a strategy.

The teams pulling ahead are doing something different. They pick one C at a time — usually Context, sometimes Control — and they compound it for a quarter. They ship a real semantic layer, or a real eval harness, or a real cost dashboard, before they move on. A year in, they are at 3-3-2-2 and the next model release is a tailwind rather than a fire drill. Their competitors are still at 1-1-1-1 and reading trend pieces about why enterprise AI does not pay off.

If you take one action from this essay: run the scorecard on your team this week. Not next quarter, this week. It is a twenty-minute exercise, and the disagreements it surfaces — inside a room of people who all thought they were aligned on the AI plan — are worth more than the score itself.

The Context Advantage is the long-form playbook behind every column of that scorecard. Start with https://bricksnotes.com/context-advantage/book/chapter-3 if Context is your bottleneck. Jump to https://bricksnotes.com/context-advantage/book/chapter-14 for Control. https://bricksnotes.com/context-advantage/book/chapter-20 is Cost. https://bricksnotes.com/context-advantage/book/chapter-24 is Choice. https://bricksnotes.com/context-advantage/book/chapter-27 is the full reference architecture that ties them together. If you have read this far and want the whole thing, https://bricksnotes.com/context-advantage/buy is the buy page.