Shared AI Standards for Portfolio Companies: Memory and Approvals

Superintelligence for portfolio companies means preparing each operating company so its AI agents can take on more capable work over time, while the holding group keeps the rules for memory, approvals and observability consistent across the portfolio. Without shared standards, every company buys or builds its own agent, stores context in its own way and records activity differently, so the useful learning stays trapped inside one business. A holding group can avoid that outcome by agreeing on a short set of standards before the first agent goes live.

Key takeaways

• Decide who owns each company's agent memory before the agent stores anything, and write that ownership into contracts.
• Use one approval matrix across the group, with local thresholds only where the risk genuinely differs.
• Require every agent to keep the same minimum observability record so performance can be compared on equal terms.
• Keep the group standard short enough to fit on one page, and review it on a fixed schedule.
• Let the group own the rules and each operating company own the agent's daily work and data.

Why portfolio companies start from zero

Most portfolio companies adopt AI because a department lead finds a tool that saves time. The result is a set of agents that each hold their own instructions, documents and customer context, configured by whoever had time that month. When the next company wants a similar agent, its team begins again, and the group cannot see which prompts, evaluation methods or escalation rules actually worked.

Starting from zero is rarely a question of model quality. It is a question of missing structure. Nobody decided who owns the memory an agent accumulates, which actions it may take without a person, or how anyone would notice that it had gone wrong. These gaps are cheap to close at the start and expensive to close after several companies have built around different assumptions.

Compounding requires a shared frame. The holding group does not need to run every agent. It needs to define the parts that should be identical across companies, so that an improvement in one business can be carried into the others without moving confidential company data.

Which standards a holding group should set

Keep the list short. Four or five standards, written on one page, are easier to enforce than a thick policy that nobody opens. A workable set covers:

• Memory ownership: who owns what an agent stores, where it lives, and how it is exported or deleted.
• Approval boundaries: which actions an agent may take alone, which need a named person, and which are never allowed.
• Observability: the minimum record every agent keeps, so that its behavior can be reviewed and compared.
• Tool and permission access: which systems an agent can reach, under which identity, and who reviews those rights.
• Change control: how prompts, tools and permissions are altered, tested and approved.

Each standard should name an owner at group level and an owner inside the operating company. The group owns the rule. The company owns the agent's daily work and the data it touches. This split keeps control at the top without pulling operations up to the holding company.

How to set memory ownership without exposing company data

Memory is the part most often misunderstood. An agent accumulates instructions, reference documents, customer and supplier facts, past decisions and human corrections. Some of that belongs to the company and must never leave it. Some of it is a reusable pattern, such as a scoring rubric for drafts or a template for escalation notes, and the group may want that pattern in a shared library. A practical sequence looks like this:

• Inventory what each agent stores: instructions, documents, facts about customers or suppliers, past decisions and corrections.
• Classify each item as company-owned, shared pattern or disposable, and write the classification down.
• Write export and deletion rights into each company's agreement with the agent provider, so the company can take its memory elsewhere if it changes vendors.
• Move only shared patterns into the group library, with names, contract terms, pricing details, customer records and personal data removed.
• Test the arrangement once: take one pattern, apply it to a second company's agent, and confirm that the second agent improves without any access to the first company's records.

Keep the ownership rule simple enough that a new operations manager can apply it in a week. The distinction between what a company keeps and what a vendor can reuse is also central to avoiding dependence on one platform. The question is explored in How to Avoid AI Vendor Lock-In With Business Agents.

Approval boundaries that travel across companies

Approval rules tend to drift. One company lets its agent send replies directly, another requires a person to read every message, and a third never considered the question at all. A single matrix, applied the same way everywhere, prevents that drift. A workable version classifies actions by consequence rather than by software:

• Reading and summarizing internal material: the agent acts alone, and the action is logged.
• Drafting external messages or documents: the agent drafts, and a named employee sends or signs.
• Changing records in a customer, supplier or finance system: the agent proposes the change, and an approver confirms before the write.
• Committing money, signing terms or changing prices: the agent never acts alone, and a named person with authority decides.
• Granting or widening tool access: the change goes through the change process and is reviewed by the group owner.
• Using personal data beyond its original purpose: the agent stops and escalates to the company's data owner.

Local thresholds are allowed when the risk differs. A company that handles regulated customer data may require approval for actions that another company can automate. Record each local exception with its reason, so the group can see why two agents behave differently and can challenge exceptions that no longer make sense.

Illustrative: two operating companies each run an agent that answers supplier questions. In the first, the agent looks up order status and replies directly. In the second, the agent also looks up order status, but any reply that mentions payment terms goes to a finance lead. Both follow the same matrix, and the difference is one local exception, written down with its reason. At the next review, the group can confirm that the exception is still needed. For the steps of that review, see How to Run a Permission Review for Your AI Agents, Step by Step.

Observability: the record every agent must keep

Observability means being able to answer four questions about any agent action: what happened, why it happened, which information was used, and what a person changed afterwards. Observability guidance from OpenTelemetry describes telemetry in terms of traces, metrics and logs. The same logic applies to agents, where a single request may call a model, retrieve documents and invoke a tool, and each step can fail in its own way.

For a holding group, the goal is a minimum record that is identical across companies:

• The request, the user or system that triggered it, and the time it occurred.
• Each tool call, with the identity used to make it and the result returned.
• The documents or memory items retrieved to produce the answer.
• Every approval requested, granted, refused or skipped.
• The final output and any later human correction.

Use the same fields in every company. The group can then compare agents on the same measures instead of on impressions. Store the record in a place the group can query, and set retention periods in line with each company's data obligations, since some of these logs will contain personal data.

The record shows what happened. Evaluations show whether a change made the agent better. A recurring test set of realistic tasks, run after every prompt, tool or permission change, reveals whether an agent improved or only behaves differently. Without that test set, a group will argue about anecdotes, and the agents will drift without anyone noticing.

Running the standard without a central bureaucracy

A group standard fails when it becomes a document that nobody uses. Three habits keep it alive. First, hold a short group review each quarter where each company shows its permission list, its exceptions and its incidents, using the same template. The steps for that review are set out in Quarterly AI Agent Permission Review: Steps and Evidence. Second, share incident summaries across the group with company identifiers removed, so the lesson travels even when the data does not. Third, route every change to an agent through one path, so no company quietly widens its own access.

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. For a holding group, a first engagement can define one useful role in one operating company, the memory and tools that role needs, its approval boundaries and an observable success criterion. The same structure can then be applied in the next company with the group standard in place. Managed operations can include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, which keeps the standard working after it is written.

If you are not yet sure which operating company should go first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

Does the holding group need to own each portfolio company's agent? Not necessarily. The group owns the standard, while the operating company owns the agent's daily work, its data and the people who approve its actions. The group's role is to define the rules, review the evidence and approve exceptions. Where a company already uses a vendor-hosted agent, the group can still make export rights and the common observability record conditions for continued use.

Will shared standards slow down the operating companies? They can, if the standard is long or changes often. A short set of rules, reviewed on a fixed schedule, usually creates less friction than each company debating the same questions alone. Local exceptions remain possible when the risk differs, and each one is recorded with its reason so that teams know exactly where they stand.

Can companies learn from each other without sharing confidential information? Yes, if the group separates patterns from data. Evaluation rubrics, tool descriptions, escalation templates and prompt structures can be shared once names, prices, contract terms, customer records and personal data are removed. Facts specific to one company, such as the terms a particular supplier agreed to, stay in that company's memory and never enter the shared library.

What should a holding group check first in a company that already runs an agent? Start with four questions: what the agent stores and who can export it, which actions it can take without a person, which systems it can reach and under which identity, and whether anyone can see what it did last week. If any answer is unclear, the agent is not ready for group comparison yet. A short permission review answers the first three questions, and the observability record answers the last.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.