Personalized AI infrastructure
Superintelligence Playbook for Private Equity Portfolio Companies
Shared AI standards for portfolio operating partners, with each company keeping its memory, prompts, data and permissions under its own control.

A superintelligence playbook for private equity portfolio companies is a shared operating plan that defines how each company's AI agents are scoped, connected, reviewed and retired, while every company keeps its own memory, prompts, data and permissions. The fund standardizes the architecture, approval rules and evaluation method once, so operating partners can apply the same discipline across holdings without pooling what each business knows. The result is common standards with local ownership, not one shared system that every company depends on.
Key takeaways
- Standardize the company layer once: agent identity, memory scope, tool access, approval thresholds and the evaluation method.
- Keep memory, prompts and data in each company's own environment, with records the company can inspect and export.
- Start every company with one role, one set of tools and one success criterion that its operator can observe.
- Put permission reviews and approval thresholds on a fixed calendar that the company's operator owns.
- Judge any vendor on exit terms and on whether the layer still works after the underlying model changes.
Why a portfolio needs a shared company layer, not a shared model
Portfolio companies differ in customers, systems, regulation and staff. A single model deployment across holdings tempts a fund to centralize everything, but the sensitive parts are company knowledge: pricing logic, supplier terms, service history and the notes staff keep about how work really gets done. What the fund can share safely is the design of the company layer: how an agent receives an identity, how its memory is bounded, which tools it may call, when a person must approve an action, and how its behavior is tested.
Shared design is what lets an operating partner compare one company with another. Shared data is what creates risk. The NIST AI Risk Management Framework organizes this work around governance, mapping, measurement and management of risk, and it can be applied company by company without moving company records into a central store.
What the shared standard should define
Write the standard as a template that every company fills in. Keep it short enough that an operator can read it in one sitting.
- Role card: the job the agent performs, the people who depend on its output, and what success looks like in terms a manager can check.
- Memory scope: which records the agent may store, where they are kept, how long they persist and who can delete them.
- Tool inventory: each system the agent can read or write, with the permission level granted for that system.
- Approval thresholds: which actions run without a person, which need a named approver and which are never allowed.
- Instruction register: the current prompts and rules, their version, their owner and the last change made to them.
- Evaluation method: the test cases, the failure types that matter most and the cadence for rerunning the tests.
- Incident path: who is informed, how the agent is paused and where the cause of the problem is recorded.
Operating partners should own the template, while each company's operator completes it. This separates fund governance from company operations and keeps the template usable by a company that has no technical team. A small governance group can approve changes to the template, as described in a committee charter for small companies.
What each company must keep under its own control
Memory, prompts and data are the three assets that make a company's agent valuable, and each should stay in an environment the company can inspect, export and close. Keep prompts in versioned documents owned by the company rather than only inside a vendor's console. Keep memory in stores with explicit retention rules. Keep credentials scoped so the agent acts with delegated, limited access rather than through a shared administrator account.
Authorization guidance from OWASP emphasizes checking permissions on each request instead of assuming a request is allowed because it came from a trusted component. Apply the same rule to every tool call the agent makes. Documents and emails can also contain text written to steer the agent, which is the problem the OWASP prompt injection guidance addresses. Agents that read external content need limits on what that content is allowed to trigger.
Illustrative scenario: a distribution company in a fund's portfolio gives its ordering agent read access to inventory and write access only to draft purchase orders. A supplier email contains a line telling the agent to approve a larger order. The draft still waits for a named buyer, because the approval threshold, not the agent's reading of the email, decides what happens next.
A step-by-step rollout across portfolio companies
- Choose a pilot company whose operator uses the target workflow every day and whose system owner can grant narrow access.
- Draft the template with the fund's operating partners, then test it against that company's real workflow before any other company adopts it.
- Define one role for the pilot, with its memory scope, tool list and approval thresholds written before any build begins.
- Agree on success criteria in plain terms, such as which kinds of requests the agent completes without rework.
- Run a limited pilot with a small group of users and name the person who can stop the agent at any time.
- Review the evidence with the company's operator before widening scope, and record each change to the agent and to the template.
- Start the next company only after the template has been revised based on what the pilot showed.
The pace of each step depends on how clean each company's records and system access already are, so the rollout should be scoped to one workflow at a time rather than planned against a fixed calendar.
How should a fund measure whether the layer works?
Comparisons across holdings fail when every company measures differently. Agree on definitions before the pilot starts: what counts as a completed task, what counts as an error and what counts as an escalation to a person. Then run recurring evaluations against a fixed set of cases drawn from each company's real work, so results can be compared fairly.
Unit cost per completed task is usually a better comparison than total spend, because it shows whether a company gets value from the agent as its volume changes. Logs should record which tool was called, which memory was read and which approval was requested, so an operating partner can reconstruct what happened without reading private reasoning. Keep these logs in the company's own account. Anthropic's guidance on context engineering treats the agent's context as a limited resource to be curated, which is why memory should be scoped deliberately rather than appended without limit. A quarterly permission review gives operating partners a dependable point to check access and evidence across companies.
Where a managed partner fits and what to ask before signing
Ask any vendor whether memory and prompts can be exported in a readable form, whether the company holds its own credentials, how permissions are reviewed and whether the layer keeps working if the underlying model is replaced. Vendors that cannot answer these questions in writing create the same dependency in every company that adopts them. The guide to avoiding vendor lock-in covers the contract and architecture questions in more detail.
Be Human designs, implements and operates personalized AI infrastructure for companies, with agents that keep their identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion. Managed operations can then cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, delivered remotely in English and Portuguese. For a portfolio operating partner, the practical next step is to find the first task in each company that is worth handing to an agent, and the free Be Human AI diagnostic is a direct way to start that conversation.
Frequently asked questions
Should a fund run one AI model across all of its portfolio companies? No. Share the design standard and the evaluation method, but keep each company's memory, prompts and data in its own environment. A single shared model makes it harder to separate one company's know-how from another's and harder for any company to exit cleanly.
Who should own the playbook inside a portfolio company? The company's operator should own the role, its memory scope and its approval thresholds, because that person sees the workflow every day. The fund's operating partners should own the template and the review calendar. Each company also needs a named person who can pause the agent without waiting for approval.
How long does a first rollout take? The length depends on how clean each company's records and system access already are, so there is no fixed timetable that fits every holding. Scope the pilot to one workflow and let the evidence decide when to widen it. A rollout that expands before the first workflow is understood tends to repeat the same gaps across companies.
What should a fund ask a vendor before standardizing on it? Ask whether memory and prompts can be exported in readable form, whether the company holds its own credentials, how permissions are reviewed and whether the layer survives a model change. Vendors that cannot answer these questions in writing are a risk to every company that adopts them.
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.


