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

Personalized AI infrastructure

Superintelligence Gap Analysis: A Worksheet for Your Current Stack

A worksheet to compare your models, memory, tools, permissions and logs with what a company superintelligence layer will need as models improve.

Be Human typographic cover with the title: Superintelligence Gap Analysis: A Worksheet for Your Current Stack
Conceptual illustration created with AI · Be Human

A superintelligence gap analysis is a structured comparison between what your company runs today and what a company-wide AI layer will need once models are much more capable. It checks five areas: the models your workflows depend on, the memory they can read, the tools they can call, the permissions that govern them, and the logs that show what happened. The result is a short, ranked list of gaps, each tied to a decision and an owner.

Key takeaways

  • A gap analysis compares your current stack with what a company AI layer will require, not with what a model can do in a demonstration.
  • Score five areas separately (models, memory, tools, permissions and logs), because each one fails in a different way.
  • Every gap should end in a decision with a named owner: fix it now, schedule it with a trigger, or accept it in writing.
  • Start from the workflows people already run, since they reveal the memory and permissions a superintelligence layer would inherit.
  • Repeat the worksheet on a fixed schedule, because both your stack and the models underneath it will change.

Why start with your current stack instead of the models

Most planning conversations start with the models: which one, how capable, and how soon. Those questions move quickly, and no company controls the answers. What a company does control is the surface that any model will touch: the records it can read, the systems it can act in, and the approvals that stop it from acting. A gap analysis shifts attention to that surface.

The useful question is not "when will a stronger model arrive?" It is "if a much more capable agent were connected to our workflows tomorrow, what would break first?" The answer is almost always found in the existing stack. Customer context may be split across three inboxes. Shared logins may let anyone issue a refund. Nobody may be able to say which model produced last month's quote. Those weaknesses exist today, and a stronger model would only make them more expensive. If you want a simple picture of how these parts connect before you score them, Draw Your AI Stack: A One-Page Architecture Map for Leaders gives a useful starting frame.

The five columns of the worksheet

Create a table or a spreadsheet with five rows, one per area, and three columns: current state, required state and gap. Keep the wording plain. The worksheet is only useful if a manager can read it without a glossary.

  • Models: which providers and model versions your workflows call, whether you can switch them without rewriting processes, and whether outputs are checked against a fixed set of tasks with known good answers.
  • Memory: where business context lives (CRM records, shared drives, chat history, spreadsheets), who is allowed to change it, and whether a reader can tell a current fact from an outdated one.
  • Tools: every action an agent could take, such as sending a message, changing a record, or issuing a credit, with the system behind each action and the exact scope granted.
  • Permissions: who approves which kinds of action, which identity each agent uses, and whether access is granted per task or left open indefinitely.
  • Logs: whether each agent action is recorded with its input, the tool called, the output, the approver and the time, so that an incident can be reconstructed afterward.

The logs row deserves extra attention. Observability guidance such as the OpenTelemetry primer describes traces, metrics and logs as signals that must be correlated to explain a system's behavior. A company without that trail cannot prove what an agent did, and cannot improve it in a disciplined way.

How to run the worksheet in one working session

The analysis works best as a single meeting with the right people in the room. Invite the person who runs the workflow, not only the technical lead, and include someone who can approve spending or access changes. Plan for the meeting to end when the list is complete, not when it is perfect.

  • Name the workflow you are examining first, then keep the discussion anchored to it.
  • For each of the five areas, write the current state in one sentence that names the actual systems, not the category of software.
  • Mark each cell as ready, partial or missing, using the same definitions for every area so that scores can be compared.
  • Trace one real request from start to finish through the workflow, noting every point where information is read, changed or approved.
  • Turn each gap into a sentence that begins with a verb and names an owner, such as "Assign the discount policy owner and move the rule into the shared policy document."
  • Close the session by choosing which gap to examine first in the decision step, not by fixing anything yet.

The trace step matters most. Abstract lists tend to rate every area as partial. A single traced request usually shows exactly where an approval is missing or where two systems disagree.

How to score each gap: a decision guide

A gap is not a score. It is a decision waiting to be made. Use the following guide to decide what each gap requires, and record the decision next to the gap so that it can be reviewed later.

  • If the gap blocks a workflow that already runs, fix it before any new agent work begins, because the existing process is where the damage would happen first.
  • If the gap matters only when a more capable agent is connected or its scope widens, record it with a clear trigger, such as a new integration or a new approval level, and review it at the next scheduled check.
  • If an agent that can write to a system has no log of its actions, treat that as a blocker for the agent, regardless of how capable the model is.
  • If two sources give different answers for the same business fact, assign one system of record before anything else is built on top of either source.
  • If you decide to accept a gap, write down who accepted it, why it is acceptable for now, and when it returns for review.

The last rule protects you from quiet risk. Accepting a gap is a legitimate choice for a small company, but an accepted gap that nobody revisits becomes an unowned gap. For the permission side of these decisions, Quarterly AI Agent Permission Review: Steps and Evidence shows how to keep accepted risks visible over time.

Illustrative scenario: a distribution company's first pass

Illustrative scenario, not a client account: a wholesale distributor with a small sales and support team uses a shared inbox, a spreadsheet of quotes and a helpdesk tool. The operations lead runs the worksheet with the support coordinator and a sales manager.

The models row is partial. The team uses one provider through a chat tool, and nobody can say which model version drafted a given reply. The memory row is missing. Discount rules live in an email from the sales director, and the spreadsheet uses a different figure for two product lines. The tools row is open: the helpdesk assistant can reissue invoices. The permissions row is ready on paper, but the team still shares one login for the helpdesk. The logs row is missing for invoice actions.

The ranked output is short. First, replace the shared login and start logging invoice actions, because an agent with write access cannot be trusted without them. Second, move the discount rule into one document with a named owner. Third, schedule model switching as a later item with the trigger "before a second provider is added." None of these decisions depends on predicting how capable future models will be. They depend only on what the company does today.

Where the gaps usually sit, and how Be Human handles them

In most small and mid-sized companies, the largest gaps sit in memory and permissions, not in models. Business context is scattered across tools that were never designed to act as a single source, and access grew by convenience rather than by design. Logs are the third common gap, because many tools record outcomes but not the reasoning path that led to them. The practical lesson is that a company can prepare most of its superintelligence layer with its existing models, provided its records, approvals and traces are in order. For a company that wants to see the data side of this work first, Data Readiness Before Superintelligent Agents: A Six-Point Audit covers the records that should be clean before any agent relies on them.

Be Human designs, implements and operates personalized AI infrastructure in which an agent keeps its identity, memory, channels, tools, permissions and operating context under human control. In a first engagement, the team turns a worksheet like this one into a single useful role, documents the memory and tools that role needs, sets approval boundaries and defines an observable success criterion. Managed operations can then include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, with changes requested and approved through the same boundaries.

Turning the worksheet into a first decision

A finished worksheet is valuable only if it leads to one action within the next few weeks. Choose the gap that blocks the workflow you care about most, assign its owner and write down the trigger for each deferred item. Then review the list again on a fixed schedule, adding new rows whenever a new tool, channel or agent is proposed. Keep the logs of each review, because the evidence of what changed is what makes a later decision defensible.

If you are not yet sure which workflow should be the first one handed to an agent, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

How often should a superintelligence gap analysis be repeated? Repeat it at least once every quarter, and also whenever you add a new tool, channel, data source or agent. Each change can create new memory, new permissions or new actions that the previous worksheet did not cover. Keep the earlier versions so you can see which gaps were closed and which were accepted.

Does the worksheet require technical staff to complete? It requires people who understand the workflow and people who can approve access, and a technical person should confirm the systems named in each cell. A non-technical manager can usually complete most of the rows if the definitions of ready, partial and missing are agreed at the start. Technical input becomes essential when the trace step reveals an integration that nobody can describe.

Can a small company run this analysis without an outside partner? Yes, a small team can complete the worksheet internally, and doing so is useful even if a partner is later hired. An outside partner adds the most value when the gaps touch many systems or require ongoing operation, such as monitoring and permission review. Whoever runs the analysis should be able to explain every accepted gap in writing.

What should we do if almost every cell in the worksheet is marked missing? Treat that result as useful information rather than as failure, since it shows the company is at an early stage. Start with the logs and permissions rows, because they protect everything else and are usually the cheapest to improve. Then choose one workflow, make its records and approvals clear, and rerun the worksheet after that single workflow is in order.

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 →