Shipping into ninety products, not one
The standardize-across-the-portfolio thesis is why a surface gets built centrally at all, and it moves the definition of good: what ships into ninety products is judged on how cheaply it survives the ninetieth, not on how well it fits the first.
Everything so far has assumed one product. Now change the assumption to the one that actually defines the job: the surface you are proposing does not ship into a product, it ships into a portfolio. Vista describes an “In-house team of AI engineers and specialists that rapidly designs, builds, and monetizes Agentic AI products”, and an Agentic AI Factory described as “a platform purpose-built to scale Agentic AI across our enterprise software portfolio.” Both are Vista describing its own model, which is worth holding in mind for the whole lesson.
You already know how to build a component that many teams consume. What changes here is not the technique. It is the economics, the failure mode, and what the word “good” refers to — and the value case you write for a shared surface is a fundamentally different document from the one you write for a feature.
Why the thesis exists at all
The economic logic is simple enough to state in one paragraph, which is why it is so persuasive at the fund level. If a capability has to be built once for each of ninety companies, ninety teams each pay the full build cost and produce ninety inconsistent results. If it is built once centrally and integrated ninety times, the build cost is paid once and the recurring cost per company collapses to the integration cost. In the income statement terms from the P&L lesson, this is a research and development route: the claim is not that any one product earns more, but that the portfolio spends less to reach the same capability.
Recall the scale of that line. In Snowflake’s fiscal 2026 results, research and development was $1.97 billion against $4.68 billion of total revenue — about 42%. In a software business, engineering capacity is not a rounding error being optimised for tidiness; it is one of the two largest cost lines in the company. That is the number the standardisation thesis is aimed at.
Vista frames the portfolio itself as the advantage, publishing that with “85+ companies and over 250 million users, Vista’s portfolio is a proving ground for agentic products at scale”. The joint announcement with Google Cloud on 22 April 2026 puts the reach differently again — “90+ companies serving more than 2.5 million enterprise customers and more than 750 million users worldwide” — and describes one worked instance: Duck Creek building a claims first-notice-of-loss orchestration agent through the Agentic Factory using Google’s models.
Where people get burned
Vista’s own page also forecasts “5-10 AI agents per user” and the eventual deployment of billions of autonomous agents across the portfolio. That is a projection published by the firm making the investment, in a document written to explain why the investment is smart. It is legitimate as a statement of intent and worthless as evidence, and repeating it in your own case would be the exact overclaiming this course is against. Cite it as “Vista’s stated ambition,” never as a market fact.
What actually changes when N is ninety
Four things, and each one changes what your value case has to contain.
The break-even is a count, and you have to name it
A shared surface has a build cost paid once and an integration cost paid per product. That means the case has an arithmetic core that a feature case does not: how many integrations before the central build is cheaper than everyone doing it themselves? If the honest answer is “eleven,” then the case is that eleven products will adopt it, and the falsifier writes itself — adoption stalls below eleven.
This is enormously freeing, because it converts an argument about taste into an argument about a number, and it is a number you can be wrong about in public without embarrassment. It also disciplines scope: a surface with a high per-integration cost needs many more adopters to pay for itself, which is a design constraint expressed as arithmetic.
Adoption is a sale, not a deployment
Each portfolio company is a separate business with its own roadmap, its own release train, its own customers and its own engineering leadership being measured on its own plan. Shipping a component into that environment is an internal sale to a buyer who has a full backlog and no obligation to care about portfolio-level savings that do not appear on their own P&L.
Whether adoption is ever mandated is not something this course knows. Vista publishes that it provides portfolio companies with “early access to next-generation agentic tooling, engineering collaboration and go-to-market channels” — the language of an offer, not an instruction. Nothing public says how that plays out inside a given company, and the honest posture is to assume you must earn adoption and be pleasantly surprised if you do not.
“Good” moves from the first integration to the ninetieth
A component evaluated on the first product it lands in will be shaped by that product’s specifics and will fight every product after it. The quality that matters for a portfolio surface is how cheaply the next integration goes — how little the adopting team has to understand, how few of their assumptions it breaks, how well it tolerates a brand, a data model and a release cadence it was not designed against.
This is the one place your existing expertise transfers almost directly: token-driven theming exists precisely so that visual identity is a configuration rather than a fork. What does not transfer is the governance assumption underneath most design systems — that consumers share a roadmap and can be asked to migrate. Across separately owned companies, a breaking change is not a migration guide. It is dozens of independent negotiations with teams who each have something more urgent on.
The failure mode is the average product
A surface designed to fit ninety products with equal weight fits none of them well, and the symptom is configurability: every disagreement gets resolved by adding an option, until the integration cost per product exceeds what building it locally would have cost. Then the break-even count silently rises past the number of companies willing to adopt, and the whole thesis inverts — the shared component is now more expensive than the duplication it replaced.
The defence is unglamorous and worth writing into the case explicitly: name the first three products the surface is genuinely for, design against those, and state that the other eighty-seven are a hypothesis to be tested rather than a promise being made.
Retrieval check
You are asked to ship a “human review gate” pattern — the confidence-scored approve-or-correct surface you already built once — across the portfolio. What is the strongest honest claim, and what is the tempting dishonest one?
Check your answer
The dishonest one is that standardising human review across the portfolio will improve trust in AI features and therefore retention. Every word of that is plausible and none of it is checkable, which is why it will get nodded at and not funded.
The honest one is an R&D-route claim with a count in it: several portfolio products are independently building human-in-the-loop review for AI output; each build is some number of engineer-weeks; a shared surface costs one build plus an integration cost per product; here is the number of adopters at which that is cheaper, here are the three products that would go first, and here is the signal — adoption stalling after the first two, or integration taking longer than the local build would have — that would tell me the thesis is wrong for this component.
The second claim is smaller and vastly more fundable, because everything in it can be checked by someone who does not trust you yet.
Hands on
Convert one case into a portfolio case
Done when: One entry in VALUE-CASES.md is rewritten as a shared-surface case containing a break-even integration count, three named first adopters, and an explicit adoption number that would falsify it.
- Take the case whose mechanism is most reusable — almost certainly the review gate rather than the advisory chatbot — and copy it into a second entry marked as the portfolio version.
- Estimate two costs, in engineer-weeks, and write both down even though both are guesses: the one-time build of the shared surface, and the per-product integration. Label them as estimates. A stated estimate you can be corrected on is worth more than a confident number nobody can interrogate.
- Divide to get the break-even integration count, and write it as a sentence: “this is cheaper than local builds once N products adopt it.”
- Name the three products it is actually designed for first. If you do not know real portfolio products, use three shapes you do know — a document-heavy workflow tool, a claims or intake system, a reporting product — and say they are shapes, not names.
- Write the falsifier as an adoption number and a date: the count of adopters below which, by a stated point in time, you would call the shared build a mistake.
- Bring it into the chat. I will press hardest on the integration-cost estimate, because that is the number that quietly rises with every configuration option and takes the whole case down with it.
What this does not cover
You now have the structure — the clock, the P&L routes and the portfolio arithmetic — and none of the vocabulary. Everything so far has carefully avoided saying ARR, net revenue retention, CAC payback or the Rule of 40, because each of those is defined differently by different companies and a metric half-learned is a metric misquoted. The metric-set module defines them precisely, names where the definitions conflict, and ends on the subtractive discipline this course actually turns on: saying out loud which of them your change cannot move. After that comes the one-page case itself.
Until those land, the translations reference page is the working shorthand — design-language claims paired with the operating- language version, and the mechanism that makes each pairing honest rather than spin.
Read this next — primary source
Vista’s Agentic AI FactoryVista Equity Partners — free; a vendor documenting its own operating model
This lesson takes the structure from it — a central team, a platform, portfolio companies receiving tooling and go-to-market help. What the page adds beyond that is the firm’s own theory of why the portfolio is the asset rather than the constraint, plus a worked case of one portfolio company deploying agents into an existing product. Read it the way you would read a framework vendor’s architecture page: first-hand and authoritative about what the firm intends, silent about what has been hard, and written to persuade limited partners and prospective employees alike. The gap between what it claims and what it declines to quantify is itself the useful part.
Stuck, curious, or think this lesson is wrong? Ask your teaching agent. The lessons are the scaffold; the conversation is where the learning gets unstuck.