Skip to content
be-human.ai
← Back to the journal

Personalized AI infrastructure

Superintelligence Layer Sizing: Estimate Users, Agents and Memory

Estimate users, agents and memory volume for a company superintelligence layer before you commit to infrastructure, using a plain sizing method.

Be Human typographic cover with the title: Superintelligence Layer Sizing: Estimate Users, Agents and Memory
Conceptual illustration created with AI · Be Human

Superintelligence layer sizing is the work of estimating how many people will use a company AI layer, how many agents it needs to run, and how much memory those agents must store and search, before you commit to infrastructure. Start from the work your team already does, count interactions and records instead of headcount alone, and size for one useful role with a margin you can review. Treat the result as a working document that you revisit after measured use, not as a forecast to defend for years.

Key takeaways

  • Size from work units such as requests, handoffs and records touched each week, not from headcount alone.
  • Plan one agent per role, with one memory scope and one permission set, before counting how many agents the company may run later.
  • Memory volume is records, times average size, times versions kept, plus index overhead and growth over the review period.
  • Peak concurrency and approval capacity set real limits more often than the total number of named users.
  • Re-run the estimate after the first measured month and keep it as a document the operations leader owns.

Start with the work, not the headcount

Headcount tells you who might log in. It does not tell you how much work reaches the layer. A better starting point is a list of the recurring tasks an agent would touch, with a rough count of how often each one happens in a normal week and in a busy week. Those counts drive the rest of the sizing, and they are easy to check with the people who do the work.

Use this sequence to build the first estimate:

  • List the three to five roles or processes where an agent would act first, and name the person who owns each one.
  • For each process, count the requests, messages or records it handles in a typical week and in its busiest week.
  • Note which channels carry the work, such as email, a chat tool, a web form, a shared folder or call transcripts.
  • Mark each task as read only, draft, or act, because actions need more permissions and more approval capacity than reading.
  • Record how many records each task must consult and how long the answer must be kept or reviewed.
  • Write the totals on one page, with the date and the assumption behind each number, so the next review can see what changed.

Illustrative scenario: a 40-person services company wants an agent for quote follow-up. Its sales coordinator handles about 30 quote threads a week, most of them read from a shared mailbox and checked against a price list. The first estimate therefore counts 30 threads, two channels, one price list and a set of past proposals. That is a much smaller sizing problem than treating the whole company as one group of users, and it is the problem the first decision should rest on.

How many people will use the layer?

Count users in three groups. Direct users type, click or approve inside the layer. Indirect users are people whose records the agent reads or writes, such as customers held in a CRM. External users reach the layer through a channel you run, such as a customer portal or a messaging number. Each group produces different load and different risk, so do not add them into one number.

For direct users, the important figure is concurrency, not membership. Ten named users who work in the same hour create a different load from ten users spread across a full day. Estimate the busiest hour for each role, then add a margin for the early months, when people test the layer more often than they will later. Write the margin down so that you can remove it deliberately.

For external users, estimate contacts per day and the typical length of a conversation. Long conversations hold more context and therefore more memory per session, which matters for the memory step below.

How many agents should the first phase plan for?

Agent count should follow roles, not ideas. A useful rule is one agent for one role, with one memory scope and one permission set. A role may use several tools, but it should not share its authority with another role just because the two overlap. Two roles that touch the same customer record can read the same memory with different permissions, which is often cleaner than one agent that does both jobs.

One agent per role, one memory scope per agent, one permission set per agent.

Plan the first phase with one or two agents. Count the others as candidates, each with a short description, an owner and a success criterion. This list is the company's pipeline of agents. It changes as you learn, and it is not a commitment to build everything at once. Teams that describe each candidate as a role, not as a feature, usually find it easier to decide what comes next.

When you add agents later, check whether they share a memory store. Sharing reduces duplication but enlarges the area that one permission mistake can expose. The same review applies when a second location or a second brand joins the layer. For a worked example of shared memory with separate permissions, see Superintelligence for Two Locations: Shared Memory, Separate Permissions.

How much memory does the layer need?

Memory volume is the amount of information the layer must store and search so that agents can answer with the right context. Estimate it in plain units before choosing any technology. A practical formula is records per entity, times average record size, times the number of entities, times the number of versions you keep, then plus index overhead and expected growth over the period you plan to review.

Each term needs a source. Records per entity can come from your CRM, your ticketing tool or your document system. Average size can be measured by exporting a sample and checking the text or file length. Versions matter because policies, prices and procedures change, and an agent that retrieves an outdated version creates a real error. Index overhead is the extra space a search structure needs. It varies with the tool you choose, so measure it in a test load rather than assuming it.

How the agent gets knowledge also affects the size. Retrieval keeps information in a searchable store and pulls relevant pieces into each request, while fine-tuning changes model behavior through training. For most business roles the store is the thing you size, because it grows with your records and can be corrected without retraining. The [AWS guidance on comparing retrieval and fine-tuning](https://docs.aws.amazon.com/prescriptive-guidance/latest/retrieval-augmented-generation-options/rag-vs-fine-tuning.html) sets out the trade-offs in more depth.

Finally, separate memory that must be kept from memory that can expire. Conversation logs, drafts and working notes usually have a shorter life than contracts or approved procedures. A sizing sheet that treats all of them as permanent will overstate storage and understate cleanup work. Before loading any store, check the source material itself. The six-point audit in Data Readiness Before Superintelligent Agents is a useful gate.

Where permissions and approvals set the real limit

Storage is rarely the constraint that surprises companies. Approval capacity is. Every action that needs a person to approve, such as sending a quote, changing a record or issuing a refund, consumes someone's time. If an agent produces forty approval requests a day and one coordinator can review twenty, the layer is undersized even when its memory and compute are ample. Estimate approval volume per role and compare it with the people who will review it.

Permissions also change the count. Each agent should have the least authority it needs, and access should follow the task rather than convenience. The [OWASP Authorization Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) offers a sound basis for deciding who may do what to which resource. A read-only agent is simpler to size and to review than one that can write to three systems.

Log every approval with the request, the decision and the reason. A short record of what was asked, approved and rejected gives you the evidence to raise or lower thresholds later. Without it, sizing human effort becomes guesswork.

How a managed approach keeps the estimate current

The numbers in any sizing exercise are assumptions until the layer runs, so they should be written as claims that can be tested. Observability practice helps here. The [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/) describes traces, metrics and logs as the basic signals for understanding how a system behaves. For an agent, the useful measures are requests per role, tool calls per task, retrieval results per request, approval waits and the number of tasks that finish without correction. Cost per completed task is a useful companion measure, and the [FinOps Foundation guidance on unit economics](https://www.finops.org/framework/capabilities/unit-economics/) explains how to tie cost to output rather than to raw usage.

Be Human designs, implements and operates this kind of layer as a managed service. A first engagement defines one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion. Managed operations then cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, so the original sizing is compared with real usage each month. When volume grows, the change goes through a documented request and approval rather than an unplanned expansion. Service is remote, in English and Portuguese.

Where to start this week

Pick the role with the clearest weekly workload and the cleanest records. Write its request, record and approval counts on one page, and mark every assumption you cannot yet verify. Review that page with the person who approves the work, because they know the real peak hours better than any export does. Keep the page beside your architecture notes so that the next decision starts from evidence. If you are not sure which role should come first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

How many users should I count for a superintelligence layer? Count direct users who work inside the layer, indirect users whose records it reads, and external users who reach it through a channel you run. Size the direct group by concurrency in the busiest hour, not by total headcount. Keep the three groups separate so that load and risk stay visible.

How many agents does a company need at the start? Most companies should start with one or two agents, each built around one role with its own memory scope and permission set. Keep other candidate roles on a written list with an owner and a success criterion. Add agents only after the first role has produced measured results.

How do I estimate memory volume without a large technical team? Multiply records per entity by average record size, the number of entities and the number of versions you keep, then add index overhead and expected growth. Measure average size by exporting a sample from your own systems rather than guessing. Mark which items must be kept and which can expire, because that difference changes both storage and cleanup work.

Should I re-estimate after launch? Yes. After the first measured month, compare the estimate with requests per role, tool calls, retrieval results, approval waits and cost per completed task. Update the assumptions on the sizing page, note why each one changed, and send any expansion through a documented change request.

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.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →