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

Personalized AI infrastructure

The Cost of Waiting on Superintelligence Prep for Small Companies

Delaying memory, permissions and evaluation work costs small companies more than waiting for a new model. Here is what is lost and how to start now.

Be Human typographic cover with the title: The Cost of Waiting on Superintelligence Prep for Small Companies
Conceptual illustration created with AI · Be Human

The cost of waiting on AI adoption is rarely the price of a future model. For a small company, delay means the memory, permission and evaluation work that any capable agent will need stays undone, so when a useful role is finally ready, the team starts from zero. Every month without those foundations is a month of decisions made informally, knowledge left in people's heads and no way to tell whether an agent is getting better or worse.

Key takeaways

  • The main cost of waiting is the foundation work you postpone, not the capability you have not bought yet.
  • Memory, permissions and evaluations compound: each one is cheaper to build early and harder to reconstruct later.
  • Waiting for a stronger model does not remove the need to decide what an agent may see, change and approve.
  • Start with one role, one set of tools and one success criterion that a business owner can check.

What waiting actually costs before any new model arrives

Most small companies frame the decision as a bet on timing: buy now or wait for the next generation of models. That framing hides the costs that exist today. An operations lead who waits is not holding a neutral position. They are continuing to run the business with knowledge scattered across inboxes, spreadsheets and the memory of two or three long-serving employees.

The first cost is slower learning. Without a defined role and a place where its context lives, the company cannot see what an assistant gets right, what it gets wrong and where people quietly correct it. The second cost is risk that accumulates without being named. Tools get connected to a shared account, a contractor keeps access after a project ends, and nobody can say who approved what. The third cost is leverage lost to rework. When the company eventually does start, it spends most of its effort reconstructing context it could have written down a year earlier.

None of these costs depends on which model is the most capable. A weaker model with clear memory, narrow permissions and a measurable task usually produces more value for a business than a stronger model working blind.

Memory that never gets written down

An agent is only as useful as the context it can reliably reach. In practice, that context has several parts: the company's standard procedures, the customer or supplier conventions that people apply without thinking, the exceptions that were agreed by email, and the reasons behind past decisions. Most of this is never written in a form a system can use.

The Anthropic engineering guide on effective context engineering for AI agents describes how an agent works best when the information it receives is deliberately selected and kept small and relevant, rather than dumped in bulk. For a business, the practical translation is simple: decide which documents are authoritative, who owns each one, and how it gets updated. Retrieval architectures, such as the approach described in the Google Cloud reference on retrieval-augmented generation, depend on this same discipline. If the source material is contradictory or stale, a better model will simply retrieve the wrong answer more fluently.

Waiting makes this worse because the material keeps drifting. Policies change, price lists are revised in spreadsheets, and new staff learn workarounds that nobody records. Each delay adds another layer of undocumented practice that someone will eventually have to interview out of people.

Permissions that default to all or nothing

The second area of delayed cost is permissions. Many companies adopt AI in a way that gives the tool a broad login, because a narrow one takes longer to configure. That shortcut is easy to justify in a pilot and difficult to unwind once the tool is part of daily work.

The OWASP authorization guidance is useful here because it treats access as a property of each action, checked against the identity and the context of the request, rather than a single switch that is either on or off. For an agent, that means separating what it can read, what it can draft, what it can send or change, and what must always wait for a person. The NIST AI Risk Management Framework offers a broader vocabulary for the same concern, asking organizations to map and manage risks across the system's lifecycle rather than only at launch.

Delay makes permission decisions harder, not easier. Once a team has learned to rely on an agent with wide access, narrowing that access feels like a loss of service. Companies that define approval boundaries before the first deployment avoid having that argument with their own staff.

Permissions that were never designed tend to become permanent by default.

Evaluations you cannot run later

The third cost is the absence of evaluation. An evaluation is a repeatable set of cases that shows whether an agent performs the task it was given, under realistic conditions, at an acceptable level of error. Without one, the company relies on impressions, and impressions change with the mood of the person describing them.

Anthropic's guide on demystifying evals for AI agents explains why evaluations need to be built around real tasks and tracked over time. The OpenAI evals guide makes the same point from a different direction: a team needs a defined expected outcome before it can judge a result. For a small company, this does not require a research team. It requires a folder of past cases, the correct answers agreed by the person who owns the process, and a simple record of each run.

If a company waits to build evaluations until a new model arrives, it loses the most valuable moment to compare. A new model cannot be judged against a baseline that was never recorded. The company is left deciding on enthusiasm or on the vendor's claims, which is exactly the situation this discipline is meant to prevent. When reading vendor benchmarks, it helps to ask whether they measure the work your business actually does, a question covered in more depth in How to Evaluate AI Case Studies Without Being Misled.

A plan to start this quarter

The following plan is designed for a business owner or operations leader who wants to reduce the cost of waiting without committing to a large program. It works for a company with a few dozen people and a single operational area to begin with.

  • Choose one role that repeats weekly, such as answering standard supplier questions or preparing monthly order summaries, and write down the outcome that counts as success.
  • List every document, spreadsheet and rule the role depends on, mark the owner of each one and flag anything that is out of date.
  • Define what the role may read, draft, send or change, and name the person who approves each action that touches a customer, a payment or a commitment.
  • Collect twenty to thirty real past cases with the correct answers, so that any future version can be checked against the same baseline.
  • Record a short review date for permissions and evaluations, and assign it to a named person who can say yes or no to changes.
  • Only after these steps, connect the first tool, and keep the first period narrow enough that anyone can review what happened.

The AI Stack Architecture Map can help a leader see where memory, tools, permissions and evaluations sit relative to each other before any of this work begins.

How to decide what to do next

Illustrative scenario: a distribution company with about forty staff wants an assistant to answer routine questions from its sales team about delivery windows and product substitutions. The owner waits for a stronger model before starting. Meanwhile, the sales team keeps answering through chat threads, and the delivery rules live partly in a spreadsheet maintained by one logistics coordinator. When the owner finally starts, the first month goes to locating rules, agreeing on exceptions and discovering that the coordinator's spreadsheet contradicts the sales policy. Had the company started with the rules and a set of past cases, the same month would have produced a working baseline and a clear list of approvals.

Be Human approaches this problem by defining one useful role first, then the memory and tools it needs, the approval boundaries and an observable success criterion. Managed operations then keep that layer in working order through monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements. The aim is not to hand the company a stronger model. It is to keep the company's memory, permissions and evaluations under human control, so that whichever model is used next can be measured and adopted deliberately. Teams that want to see how the first months unfold can read What the First Months With a Managed AI Agent Look Like.

The useful question for a leader this week is not whether superintelligence will arrive on a particular schedule. It is which single task your team repeats often enough to be worth handing to an agent, and whether you could show someone the rules, the people and the evidence behind it today. Take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Is waiting for a better AI model ever the right decision? Waiting can be reasonable when your team has no repeated task worth automating and no internal owner for the work. In most small companies, though, the foundation work of memory, permissions and evaluation is useful with any model, so it does not need to wait for a new release. Building it now also makes a future model easier to judge.

What is the cheapest first step if we are not ready to start? Write down one recurring task, its success criterion and the documents it depends on. Ask the person who does the task to name the exceptions they handle by judgment. This takes little time, creates no access to systems and shows quickly whether the task is a good candidate.

How do we know whether our permissions are already too broad? Check which accounts or tools the AI system can reach, what each one can change, and who approved that access. If the answer to any of these is unclear, the permissions are too broad for a controlled pilot. A periodic review, as described in How to Run a Permission Review for Your AI Agents, Step by Step, gives you a repeatable way to check this.

Do we need technical staff before we can begin this work? A business owner or operations leader can define the role, the success criterion, the approval boundaries and the first set of evaluation cases. Technical work such as connecting tools and maintaining integrations usually needs a specialist, either internal or external. The business decisions should stay with a named person in the company, whoever does the implementation.

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 →