Personalized AI infrastructure
Portfolio Companies and Superintelligence: Standardize Once, Own Locally
Standardize the AI layer across portfolio companies and let each company own its data, approvals and results. A practical split for operating teams.

A shared operating layer for superintelligence across portfolio companies is a common set of standards for how AI agents are built, connected, evaluated, reviewed and paused, maintained by the operating team and applied by every company in the portfolio. Each company still owns its data, memory contents, approval thresholds and the results its agents produce. The shared layer removes repeated engineering, while ownership keeps the gains with the company that earns them.
Key takeaways
- Standardize what repeats across companies: interface contracts, logging fields, evaluation method, permission review cadence, incident handling and cost reporting.
- Keep data, memory contents, approval thresholds and business outcomes inside each company, each with a named owner.
- Measure every agent against a success criterion defined by the company that uses it, then compare patterns across the portfolio rather than raw activity.
- Write the shared layer down as a versioned product with a change process, not as a folder of templates nobody updates.
- Revisit the split every quarter, because what should be shared changes as companies mature.
Why portfolio companies rebuild the same AI work
When a portfolio has several operating companies, each one tends to start its first agent alone. One team wires a connector to its ticketing tool, another writes its own prompt library, a third invents a log format, and none of them agrees on what a failed action looks like. The result is duplicated engineering and, more seriously, lessons that stay trapped in one company. A permission gap found in one business is often the same gap in the next, yet nobody has a common standard to check against.
A shared layer answers that fragmentation. It gives the operating team one place to define how agents are shaped, so each company inherits sound defaults instead of rediscovering them. The goal is not uniformity for its own sake. A company should not pay to learn something a sister company has already learned.
What the operating team should standardize
Standardize the parts that stay the same regardless of industry. These usually include:
- Interface contracts for tools, so every agent calls internal systems through documented inputs, outputs and error messages. The Model Context Protocol introduction describes one open approach to connecting models with external tools and data sources.
- Logging and tracing fields, so every agent records what it was asked, which tools it used and what it returned, under consistent names. The OpenTelemetry observability primer gives teams a shared vocabulary without requiring a specific product.
- An evaluation method: how test cases are written, how a run is scored, and how results are reviewed before a change ships.
- A permission review template and cadence, including who can grant access and how access is removed.
- An incident scale with clear severity levels, a pause procedure and a named escalation path.
- A cost report per unit of work, so each company sees what an agent costs per completed task rather than a vague monthly total. The FinOps Foundation's unit economics capability is a useful reference for thinking about this.
Keep the standards short. A standard that takes an afternoon to read will be ignored within weeks, while a two-page version with a clear owner tends to be used.
What each company must own
Some things cannot be shared without damaging the company that depends on them. Each operating company should own its raw data and the records its agents read, the memory content that describes its customers, suppliers and processes, the approval thresholds that decide when a person must sign off, the success criteria that define a good outcome for its business, and the know-how that makes its work distinct.
Ownership also means the company has a named person who can change or pause its agents. The shared layer can recommend a configuration, but the portfolio company decides whether to adopt it. That separation matters most when an agent touches payments, contracts or personal data.
Illustrative scenario: two companies in the same portfolio both want an agent that answers supplier invoice questions. The operating team supplies the connector template, the log schema and the evaluation method. Company A sets a low approval threshold and routes disputed invoices to its finance lead. Company B pays suppliers on a different cycle, so it sets a higher threshold and keeps its supplier records in a separate memory store. Both companies follow the same standards, and each keeps the results of its own agent.
A step-by-step plan for the first shared standard
Start with one standard that already repeats in at least two companies, rather than designing the whole layer at once. A workable sequence looks like this:
- Inventory the agents and agent requests already in progress across the portfolio, noting the tools each one touches and the data it reads.
- Pick the single standard that causes the most rework, usually the tool interface or the permission review template.
- Write that standard as a short document with an owner, a version and an explicit list of what each company may change locally.
- Pilot it in one company against a clear success criterion, and record every point where the standard had to be adapted.
- Revise the standard using that evidence, then offer it to the other companies with their adaptations listed separately.
- Set a review date so the standard does not drift away from the systems it describes.
Be Human treats this as a managed operation rather than a documentation exercise. Each company's first agent is scoped to one useful role, with the memory and tools it needs, its approval boundaries and an observable success criterion defined before any build begins. Monitoring, recurring evaluations, integration maintenance and permission review then run against the same standard, with each company's data and approvals kept separate. That ongoing work is what keeps a shared layer accurate after the first version is written.
Keeping gains inside each company
Shared infrastructure carries a quiet risk: data and know-how drift into places where the owning company loses control of them. The operating team should set rules that prevent this. Memory stores should be separated by company, with the shared layer holding only schemas and interface definitions. Access decisions should be made per company and per agent, applying least privilege to every tool. The OWASP authorization guidance offers a practical checklist for asking whether each access decision is explicit and reviewable. For the know-how question specifically, the guidance on keeping company IP out of shared models is a useful companion.
Two more rules matter. A company's evaluation cases should be its own, because they encode what success means in that business. Exit terms should also be agreed in advance: a company must be able to export its memory, logs and configurations if it leaves the shared arrangement or changes provider. A shared standard should never become a dependency that a company cannot leave.
How the portfolio review should work each quarter
Treat the shared layer as something reviewed on a fixed calendar. A quarterly portfolio review works well when it covers three questions: which standards were used, which were bypassed, and which local adaptations proved valuable enough to promote into the shared layer. Permission reviews belong in the same session, not because access changes every quarter, but because the quarter is a reliable point to confirm that nothing has quietly expanded. The steps for that review are set out in the quarterly permission review guide. Companies with more than one location face the same questions in a sharper form, as covered in the two-location example.
Use this decision guide when deciding where an item belongs:
- Standardize when the same tool, log field or review step appears in at least two companies and changing it would not alter a company's business rule.
- Keep it local when the item encodes an approval threshold, a customer relationship, a supplier term or a success criterion.
- Share the schema but not the content when the structure is useful across companies but the values are sensitive.
- Pause the shared version when an incident shows that a standard failed in a way its owner did not anticipate.
If you are unsure which of your first agent tasks belongs in the shared layer and which belongs to one company, the free Be Human AI diagnostic is a plain way to find the first task worth handing to an agent.
Frequently asked questions
Who should own the shared operating layer? A small central operating team, usually reporting to the investment or operations lead, should own the standards, the templates and the change process. Each portfolio company should name its own agent owner, who decides adoption and can pause agents. The split is deliberate: the center keeps the layer consistent, and each company keeps accountability.
Is it safe to share memory between portfolio companies? Sharing a memory schema is generally safer than sharing memory content. Keep each company's records, customer details and process notes in separate stores with separate permissions, and let the shared layer define structure only. Check any arrangement against each company's own data handling obligations before data moves.
Should every portfolio company use the same AI model or provider? Not necessarily. The shared layer should define interfaces and evaluation methods that work across providers, so a company can change models without rewriting its tools. Avoiding lock-in depends on exit terms and exportable memory as much as on technical design.
How do we know the shared layer is paying off? Compare each company against its own success criterion and its own cost per completed task, then look for what the standards actually reduced, such as rework or incident handling time. Avoid ranking companies by raw activity, since volume reflects scope more than quality. If a standard adds friction without a measurable benefit, revise it or retire it.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


