Personalized AI infrastructure
Five Superintelligence Buying Mistakes Owners Make in Year One
Five costly first-year mistakes when buying superintelligence infrastructure: unclear memory ownership, missing approval rules, and how to avoid each one.

The most common superintelligence buying mistakes in year one are rarely about the model itself. They are unclear ownership of company memory, tool access granted too broadly, approval rules that exist only in people's heads, success criteria nobody can observe, and no named owner after launch. Each one can be prevented by a decision the owner makes before the first agent goes live.
Key takeaways
- Decide who owns company memory, source records and tool credentials before any build starts.
- Give each agent a narrow tool list and write down the actions it may never take.
- Turn approval rules into explicit thresholds that a person can check quickly.
- Attach one observable success criterion to each role and review it on a fixed rhythm.
- Name an internal owner for the system after launch, or buy a managed operation that keeps watch.
Mistake one: Nobody decided who owns the company memory
Many first-year buyers assume the platform owns whatever it learns. In practice, an agent's memory is a working copy of customer history, pricing notes, process documents and staff knowledge. The terms under which that copy is stored decide what you can export, correct, delete or move later. A vendor that keeps memory in a closed format can make switching expensive even when the product works well.
The fix is a short ownership record written before the build starts. It names the source system for each kind of memory, the person accountable for accuracy, the retention rule, and the exit path if the relationship ends. Ask for export formats in writing, not as a verbal promise.
- List every memory source the agent will read or write, such as the CRM, shared drive, ticket queue or chat history.
- Name one accountable person per source who can confirm whether the content is current.
- Write the retention and deletion rule for each source, including what happens to summaries the agent derives from it.
- Confirm in the contract that you can export memory in an open, documented format.
In a Be Human engagement, the memory a role uses is defined as part of the role itself, along with where that memory is authoritative. Context stays under the client's control, so the ownership map exists before any agent holds company knowledge.
Mistake two: Tools were connected because connecting was easy
Integrations are quick to switch on. An agent that needs to read an inbox often receives send rights as well, and an agent connected to the accounting system can end up with write access to every ledger because that was the default setting. Within a few weeks the owner may find the agent can do things nobody intended, and taking those permissions back takes longer than granting them did.
Least privilege means each tool is granted for one role and one purpose, with read and write access separated. Credentials should be scoped, separate for each agent and rotated on a schedule. Every tool call should be logged so you can answer who or what changed a record and when. Untrusted text inside emails or documents can also try to steer an agent toward a tool it should not use, which is why security guidance treats tool scope as a control rather than relying only on filtering the input.
- Start each agent with read-only access and add write rights one tool at a time.
- Give each agent its own credentials so one can be revoked without stopping the others.
- Log every tool call with the time, the action and the record it touched.
- Keep a written list of excluded actions, such as deleting records, issuing refunds or changing prices.
Illustrative scenario: a distributor sets up an order-status agent with full access to its order system because the setup wizard offered that option. A few weeks later, a customer message containing unusual instructions leads the agent to attempt a bulk update. The update is stopped only because the database rejects it, not because any policy prevented the attempt. The owner then splits read and write credentials and adds a rule that bulk changes always require a person.
Be Human handles tool access as part of the role definition. Each tool is named with its permission level, excluded actions are written down, and permission changes are reviewed rather than silently inherited.
Mistake three: Approval rules existed only in people's heads
Owners often say that a person should check anything important. That sentence cannot be enforced. An agent needs a rule it can apply and an auditor can verify: a category of action, a threshold, a type of customer or a kind of external message that always goes to a named approver. Without that, the first unfamiliar case forces a bad choice. The agent either stops for everything, which kills adoption, or acts on everything, which creates exposure.
Treat thresholds as a one-page policy. They should start conservative and be raised only when logged outcomes show the agent handles routine cases reliably. The sequence for loosening them is covered in approval thresholds for superintelligent agents.
- Name the actions that always need approval, such as external messages to new customers and any refund.
- Set a clear threshold for the remaining actions and write it into the agent's runbook.
- Define who approves, how long approval may take, and what the agent does while it waits.
- Record every approval and rejection so the threshold can be reviewed against real cases.
In Be Human's managed approach, approval boundaries are set in the first engagement. Later changes move through a recorded request, as described in how to request and approve changes to a managed AI agent, instead of a message in a chat thread.
Mistake four: Success was never defined in a way anyone could observe
"It should save time" is not a success criterion. Year-one purchases often launch without a baseline, so after a while nobody can say whether the agent helped. The remedy is one criterion per role, measured from logs or records the company already keeps, with a baseline captured before launch. Observability guidance generally describes the same principle: make the system's behavior visible so you measure it rather than guess at it.
The criterion should be something a business owner would recognize, such as how long a customer request waits for a first useful reply or how many invoices pass matching without manual correction. Decide in advance what improvement justifies continuing and what result triggers a pause.
- Pick one outcome the business already tracks and write the exact definition.
- Record the baseline for the same measure over a period before launch.
- Decide the minimum improvement that justifies continuing and the result that triggers a pause.
- Review the measure on a fixed schedule and write down each decision.
Be Human writes the observable success criterion into the first engagement for each role. Monitoring then tracks that measure, so the review conversation starts from evidence rather than impressions.
Mistake five: Nobody owned the system after it went live
Launch is the easy part. Integrations change, staff move on, vendors update their interfaces, and the permissions granted in month one drift by month twelve. Owners who bought a build without an operating plan often end up with an agent that nobody trusts and nobody maintains, especially after the internal champion who understood it has left the company.
Post-launch ownership needs four pieces: a named internal owner with authority to pause the agent, a recurring review of tool access and approval rules, an incident path that says who is informed and how the agent is stopped, and a change log for every prompt, tool or rule update. A regular permission review keeps access aligned with the roles people hold today, and the steps are laid out in quarterly AI agent permission review: steps and evidence. Companies that want the governance layer written down can also use a committee charter for small companies.
Be Human's managed operations cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements. The internal owner still decides what the business needs, but the watch duties do not depend on one person remembering to check.
A year-one checklist before you sign
Use these questions as a pre-contract checklist. A buyer who can answer them before the first build is in a much stronger position than one who answers them after the agent is live.
- Who owns each memory source, and how is its content exported?
- Which single role comes first, and which decision does it support?
- Which tools does that role need, at what permission level, and which actions are excluded?
- Which actions always require a person, and who is that person?
- What is the baseline, and what result would justify continuing?
- Who inside the company can pause the agent this week?
- How are changes requested, reviewed and recorded?
- Who runs monitoring and incident response after launch?
If you are not yet sure which role should be first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Who should own the company memory when we buy superintelligence infrastructure? The owner or a named executive should hold ownership, with a day-to-day steward for each memory source. The vendor can operate the system, but its terms should not prevent you from exporting, correcting or deleting what the company has stored. Put these rights in writing before signing.
How many tools should an agent be allowed to use in its first year? An agent should only have the tools its single role needs. Start with read-only access, add write permission one tool at a time, and keep a written list of excluded actions. Revisit both lists at each permission review.
What is the simplest approval rule to start with? Require a named person to approve any external message to a new contact, any refund or payment, and any change to a record that affects money or legal status. Raise the thresholds only after logged cases show the agent handles routine work reliably.
Is a managed operation worth it for a small company? It depends on whether someone inside the company can monitor the system, review permissions and respond to incidents. If that capacity does not exist, a managed operation can cover those duties while the owner keeps control of the decisions that matter.
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.


