Personalized AI infrastructure
CFO Questions Before Funding a Company Superintelligence Layer
The questions a CFO should ask before funding a company superintelligence layer: who owns the gains, exit terms, and how spend builds reusable memory and tools.

A CFO should fund a company superintelligence layer only when the company can show who owns what the spend produces. The useful questions concern ownership of gains, the terms for leaving the arrangement, and whether each part of the spend builds memory and tools the company can reuse. Asked in that order, these questions turn a broad technology proposal into a budget line that can be reviewed, paused and measured like any other infrastructure commitment.
Key takeaways
- Fund the layer only if the company, not the vendor, owns the memory, tools, permissions and operating context it produces.
- Ask how gains will be measured and who signs off on the measurement before any spend is committed.
- Require exit terms that let the company export memory, instructions, tool definitions and logs in a format its own staff can read.
- Tie each budget line to an asset that persists, such as a documented procedure or a tool that more than one agent can call.
- Start with one role and one success criterion, then expand only when the evidence supports it.
Start with who owns the gains
The first question is not what the layer costs. It is who keeps what the layer produces. When a team builds an agent that drafts quotes, routes support requests or prepares a monthly reconciliation, the output is more than the work done that month. It includes the written procedure, the tool definitions, the permission rules, the evaluation cases and the operating memory the agent relies on. If those assets sit inside a vendor's account, part of the gain from the spend belongs to someone else.
Ask the proposing team to name each asset in writing, state where it is stored and state who holds administrative access. A proposal that cannot answer those three points is not ready for funding, however capable the underlying models look in a demonstration.
- Which assets will exist at the end of the first engagement, and which of them can the company copy without the provider's help?
- Who approves changes to memory, tools and permissions, and is that person named in the internal governance?
- If the agreement ends, what remains, and in what format?
- How will the gain be measured, and who confirms the measurement before it is reported to the board?
What exit terms should protect before the first invoice
Exit terms are where a company superintelligence proposal most often hides its real cost. A clean exit does not depend on goodwill. It depends on contract language and on technical facts about how the system is built. Ask for an exit plan written before work begins, and test it the way you would test a backup: can the company actually restore what it needs, in a usable form, without the provider in the room?
A reasonable exit package usually includes the following items.
- An export of memory stores, including the records, their sources and their update history.
- Tool and integration definitions, with inputs, outputs and the credentials they require held under the company's control.
- Instructions, prompts and evaluation cases in plain text.
- Logs of agent actions, kept for the period the company's own policy requires.
- A transition window in which the old and new arrangements can run the same task side by side.
For a deeper look at how dependence builds up over time, see how to avoid AI vendor lock-in with business agents.
Connect each budget line to an asset that persists
Model usage is consumed. Memory, tools and documented workflows persist. A CFO should separate these two kinds of spend before approving anything, because they age differently. Usage charges rise and fall with activity. Assets accumulate value only if the company keeps them and reuses them.
The FinOps Foundation describes unit economics as relating cost to the business output it produces. Applied here, the question is not how many tokens an agent consumed, but what completed task each part of the spend supported and whether that task's supporting material can serve the next task.
Illustrative scenario: a regional service company wants an agent that answers customer scheduling questions. Its spend falls into three lines: model usage billed each month, an integration that reads the booking system, and a set of service rules the agent consults. If the agent is switched off, only the first line stops. The integration and the service rules remain, and they can support a second agent that sends appointment reminders. The CFO's useful question is therefore which line creates an asset, not only what the scheduling agent costs each month.
Set approval boundaries before anything runs
Finance and operations often read the same agent proposal differently. Finance sees a budget ceiling. Operations sees a set of actions the agent might take on its own. The gap between those views is where unplanned spending and unplanned authority tend to appear. Approval boundaries close that gap by deciding, in advance, which actions the agent may take alone, which need a named approver and which it may never take.
OWASP's authorization guidance is a useful reference for checking that access is granted on purpose rather than by default. Apply the same discipline to spend.
- Define three lists: actions the agent performs alone, actions that need a named approver, and actions it never performs.
- Write the approval thresholds into the agreement so that finance and operations apply the same rule.
- Give the agent its own identity with scoped credentials, not a staff member's login.
- Schedule permission reviews at a fixed cadence and keep the evidence. A practical method is described in the quarterly AI agent permission review steps.
A funding check for the finance review
When the proposal reaches the finance review, the following sequence keeps the decision concrete.
- Step one: confirm the single role the layer will fill and the business task it completes.
- Step two: list the memory, tools and procedures that role needs, and name the owner of each.
- Step three: confirm the exit package and test a sample export before signing.
- Step four: set the approval boundaries and the named approvers.
- Step five: agree one observable success criterion and the date of the first review.
- Step six: decide what will be paused, reduced or stopped if the criterion is not met.
If any step cannot be completed on paper, the proposal is not yet a funding decision. It is still a discovery exercise, which is a reasonable thing to fund separately.
Measure what the company can verify
A superintelligence layer will be tempting to judge by impressions. A CFO should insist on criteria that someone outside the project can check. The NIST AI Risk Management Framework encourages organizations to define and measure AI risks and performance in terms tied to their own context, rather than relying on general claims. For a funding decision, that means a success criterion tied to a task the business already tracks, a named person who checks it and a record of the checks.
Avoid criteria that only describe activity, such as the number of agent sessions or the volume of documents indexed. Prefer criteria that describe an outcome the business can confirm, such as a reduction in the cases an employee must re-enter, or a documented decision that was made faster and reviewed without rework. The exact measure depends on the role, which is why the first role should be chosen narrowly. For guidance on deciding what each stage must prove, see what each AI stage must prove.
How Be Human approaches the funding decision
Be Human designs, implements and operates personalized AI infrastructure in which agents keep their identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools that role needs, the approval boundaries and one observable success criterion. Managed operations can then cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements. For a finance leader, this structure means the first funding decision covers a defined role with named assets, a written exit path and a review schedule, instead of an open-ended platform commitment. If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Should a CFO approve funding before operations has named a specific use case? No. A broad platform budget without a defined role makes it impossible to check ownership, exit terms or results. A small discovery budget that ends with a named role, its required assets and a success criterion is easier to justify and easier to stop.
What is the single most important question to ask a proposing team? Ask who will own the memory, tools, permission rules and logs at the end of the first engagement. If the answer depends on the provider's goodwill rather than on written terms and a tested export, the proposal needs more work before funding.
How is spend on an agent different from ordinary software spend? Ordinary software is usually a recurring license for a tool the company uses. An agent layer also builds persistent assets, such as procedures, tool definitions and operating memory, which can serve later agents. Those assets should be tracked separately so the company can see what remains if usage stops.
How often should a funded agent be reviewed by finance? Review the budget at the same cadence as the success criterion, and review permissions on a fixed schedule that operations and finance agree on in advance. The point is not to add paperwork, but to confirm that spending and authority still match the role the company approved.
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.


