Personalized AI infrastructure
Superintelligence Investor Updates: What to Report Without Hype
How to write an AI investor update with observable metrics, decision logs and approval records, instead of superintelligence claims or unverifiable promises.

An AI investor update should state what the system did, how the work was checked, and what changed because of it. Founders who report this way list each agent in use, the tasks it completes, the approval limits it works under, and the measures that show whether results are improving. Claims about superintelligence add nothing an investor can test, and they make the next update harder to trust.
Key takeaways
- Report outcomes for each workflow: volume handled, outputs accepted on review, escalations to people, failures, and cost per completed task.
- Separate what runs in production, what is in a pilot, and what is still a hypothesis.
- Put permission and approval evidence next to performance, because investors will ask who can do what.
- Name the owner of each agent and the decision you need from the board, if any.
- Tie every forward-looking statement to a measurable test and a scheduled review, not to general trends in model capability.
Why superintelligence language weakens an investor update
Investors read updates to judge risk and evidence. The word superintelligence describes a future state that no company controls, so it cannot be checked against this quarter's work. It also invites the question of what the company will actually do differently, and the answer is often missing from the page.
A more useful update replaces the category with a system you can describe in operational terms: which agent, which role, which memory it can read, which tools it can call, and which actions still require a person. Once those details are written down, the progress is visible without any claim about the future.
Which metrics show the system is working
A small set of measures covers most of what an investor needs. Adoption shows which teams send work to the agent and whether that number is growing for the right reasons. Quality shows the share of outputs a reviewer accepted without edits, measured on a sample that you record along with its size. Reliability shows failed runs, incidents, and the time taken to recover from them.
Cost deserves its own line. Report cost per completed task rather than cost per model call, since the second figure hides whether the work was finished. The FinOps Foundation's guidance on unit economics describes this kind of per-unit view in general terms, and it translates directly to agent work. For incident and run tracking, the OpenTelemetry observability primer explains the basic idea of recording traces and events so that a result can be traced back to what happened.
Control is the fifth measure, and most updates skip it. Count permission changes, approvals granted, approvals refused, and any action the agent attempted outside its boundary. A quiet quarter with zero recorded exceptions is a result worth stating, but only if the logs exist to support it.
A step-by-step plan for the next update
Write the update from records rather than memory. The following sequence takes one working session for a small company and needs no special tooling.
- Inventory every agent or automated workflow, with its role, its named owner, and the channels it uses.
- For each one, write the single task it is allowed to complete and the task it must hand to a person.
- Pull the period's counts from logs: runs started, completed, escalated, and failed.
- Sample outputs for review, record the accepted share and the sample size, and note what the reviewer rejected and why.
- Export the current permission list and the approval log, then list every change made during the period.
- Calculate cost against completed tasks for the period, and state any cost that cannot be attributed to a workflow.
- Write the single decision you need from investors, if one exists, and the evidence that supports it.
The result is a page that a skeptical reader can check. Each line points to a log, a sample, or a recorded change.
Illustrative scenario: a support agent that escalates too often
The following is an illustrative scenario, not a client case. A services company runs an agent that drafts replies to customer emails. In one quarter, the update shows that the agent completed a large share of routine questions, but reviewers rejected many drafts about refunds, and the escalation log points mostly to that category.
A superintelligence-style update would describe the agent as a step toward autonomous operations. The evidence-based update says that refund questions were escalated because the approval rule was set too broadly, that the team will narrow the rule for small amounts only after a sample review, and that the change will appear in the next permission log. Investors can evaluate that plan. They cannot evaluate the first sentence.
What to leave out of the update
Leave out capability adjectives such as powerful, transformative, or near-human, because they describe nothing you measured. Avoid productivity claims that no one timed, and avoid anecdotes presented as results. If a single customer story is useful, label it as one example and keep it out of the metrics section.
Do not attach dates to model releases or to predictions about what a vendor will ship. These details change outside your control and make the update look dependent on news rather than on your own operation. If a statement cannot be checked by someone outside the company within a week, either remove it or mark it clearly as a hypothesis with the test that would confirm or reject it. The guidance on benchmarks in our article on benchmarks for business covers which measures tend to predict business results and which only look impressive.
How Be Human handles the reporting layer
Be Human designs and operates AI agents with their memory, tools, permissions and operating context kept under the company's control. In managed operations, the same records that run the agent also feed the investor update: run logs, sampled reviews, scheduled evaluations, and a permission history that shows every change with its approval. Permission reviews run on a regular schedule, and changes to the agent go through a recorded request and approval step, so the update reports what was actually changed and why. This does not replace the company's own judgment about what to tell investors, but it means the evidence exists before the update is written.
For a fuller treatment of the review cycle, see quarterly AI agent permission review, and for a short map of the components that produce these records, see drawing your AI stack.
Choosing the first workflow to report
Most companies should not start reporting AI progress across every function at once. Pick one workflow with a clear owner, a task that can be checked by a reviewer, and an approval boundary that someone already understands. Once that workflow produces a clean monthly record, add the next one. A narrow update that is fully supported by records earns more trust than a broad one that needs explanation. If you are unsure which workflow should be the first one reported, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Should an AI investor update mention superintelligence at all? Only if the company can state what the term means in its own operations, which is rarely the case for a small business. Most investors will accept a precise description of the agents, their boundaries and their results without any reference to the long-term category. If the topic comes up, answer with the current system and the test that would show any future change.
What is the minimum an investor update on AI should contain? It should list each agent in production, its owner, the task it completes, the approval limits it operates under, and the measures for completed work, failures and escalations. It should also report permission changes made during the period and state any decision needed from the board. Anything beyond that is useful only if the records support it.
How do I report AI results when the sample of reviewed outputs is small? State the sample size plainly and avoid rounding it into a trend. A small sample supports a cautious statement, such as that the reviewed outputs were mostly accepted, and it does not support a claim about overall performance. Say what you will sample next period and how that will change the confidence of the figure.
How often should the update be written, and who should approve it? Quarterly is a common cadence for investor updates, with a monthly internal record kept between them so the quarter is not reconstructed from memory. The executive sponsor for the AI work should approve the content, and the agent's named owner should verify the operational facts before the update is sent.
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.


