Tesseract Academy Reports
Tesseract Research · October 2026

Who Chooses What AI Gets To Do

Takeaway. Who chooses what AI gets to do is an authorization test: named chooser, scoped tools, and a stop owner. Blank rows are unowned action, not a capability.

Labs are publishing swarm-scaling math, biology watermarks, and voluntary industry pledges. Inside the enterprise, those stories collapse to one operating question: who chooses what AI gets to do. If that chooser is unnamed, you did not deploy a capability. You deployed unowned action.

First test
Who may act?

Name the role that can authorize the agent to take a consequential action.

Second test
What is scoped?

List the tools, data classes, and irreversible actions the agent may touch.

Third test
Who can stop it?

Name the kill-switch owner and the condition that fires it.

Boardroom accountability gap: agent fleets deployed without a named owner who can authorize or stop action
Everyone deployed agents. Authorization is missing when no one can say who may act, what is scoped, and who can stop it.

The chooser is missing

Public research, private gap: On 5 October 2026, Import AI opened with a single line: who chooses what AI gets to do. Jack Clark was writing about swarm scaling, biology watermarks, and a proposed market for AI scientists. The sentence also names the enterprise failure mode. Most boards can list the vendor. Fewer can name the person who may let an agent change a record, send a message, or move money.

Governance gap: principles documents without inventory, risk tiers, or a named shutdown owner
A principles document is not an authorization system. Inventory, scoped tools, and a shutdown owner come first.

Swarms are not free parallelism: Toby Ord, writing on 21 September 2026 and quoted in that Import AI edition, treated swarms as a new form of inference scaling. A four-agent swarm needed about twice the total tokens for the same performance, and only half as many tokens per agent, so wall-clock time can theoretically fall by half. Scaling the swarm ten times buys about three to five times the capability, not ten. Coordination has a tax. Paying for more agents without a named chooser multiplies unowned action, not just speed.

Voluntary pledges are not a control: Import AI also reported polling from the Center for Shared AI Prosperity. In a sample of 2,498 people, 61 percent said it is not enough for companies to agree to self-police AI development, and 54 percent said the government should set and enforce rules for AI. You do not need to adopt that poll as your policy. You do need to notice the gap it describes. A vendor pledge, a principles slide, and a Slack channel called AI Champions are not an authorization system.

Loyalty math is not a substitute: Harvard Business Review recently asked Chewy CEO Sumit Singh how he justifies customer-obsession spend when precise ROI is missing. His answer was that some acts cannot be traced to a clean return in the moment, and still need a leadership rule. Agent fleets have the inverse problem. Leaders often cannot prove a precise return either, and they still need a rule for who may act. Intuition without an owner is not a standard of evidence. It is a story told after the action already happened.

Five authorization tests before the next agent goes live

Write these five rows on one page before you widen an agent from a sandbox to a live workflow. If any row is blank, you are not ready for production. You are ready for another demo.

Tesseract AI Coach

Work the authorization rows in the AI Coach

The live tool is the AI Coach inside the free course. Use it to name the chooser, the scoped tools, and the stop condition for one workflow. It is not a calculator on this page.

Enter your volume and labour rates there. Paid certification saves the result to your Coach as an Opportunity Map.

Open the AI Coach in the free course
Tesseract AI Coach executive hub: learning path, CV audit, and Coach tools

Detection is not permission

Watermarks track origin, not authority: On 30 September 2026, Google DeepMind introduced SynthID Bio, a family of watermarking methods for synthetic biology. In wet-lab tests across VEGF-A, the SARS-CoV-2 spike protein RBD, and PD-L1, watermarked protein designs matched the hit rate, binding affinity, and sequence diversity of unwatermarked versions. DeepMind presents the work as a verification layer inside a wider biosecurity stack, not as a single lock. Provenance is useful. It does not answer who may synthesize, who may deploy, or who can refuse the order.

Do not automate chaos: scoped tools and kill switches before agent fleets go live
Do not automate chaos. Scope the tools and name the stop owner before the swarm reaches a live workflow.

An agent can notice a breach and still lack a mandate: On 5 October 2026, Stratechery described an always-on Mac Mini that ran coding agents and was compromised through a screen-sharing flaw. The persistent agent stopped executing commands after it saw passwordless admin access, then helped isolate the intrusion. That is a useful detection story. It is not a substitute for a permission design. The same essay notes that Apple is tightening Full Disk Access because agents can expose mail, messages, and browsing history, while the Mac permission prompts sit at the wrong layer for software that writes new programs all day.

Vendor controls are not your chooser: Apple can decide what an app may see on a consumer Mac. Your vendor can decide what a seat may call. Neither decision names the executive in your firm who may let an agent bind a customer, a ledger, or a regulated file. If the only stop button lives in a vendor dashboard, you rented a remote, not a control system.

Unit economics follow the same page: Swarm scaling makes the cost of unowned action visible. Twice the tokens for the same answer is a finance line, not a curiosity. If you cannot name the chooser, the scope, and the stop condition, faster parallelism is just a more expensive way to guess. The budget committee and the authorization test are the same document viewed from two chairs.

Boardroom Takeaway

A watermark, a vendor pledge, and a clever agent that notices a breach are detection. Authorization is a named chooser, a scoped tool list, and a stop condition that does not wait for the next steering group.

Build capability: practice the authorization rows before the next rollout

Structured practice over more seats: You cannot buy authorization with another licence pack. Leadership teams need a shared vocabulary for who may act, what is in scope, and what stops the workflow. The Free AI Fluency Course is the open starting point for that practice. It is built for department heads and operators who must govern AI work, not for people collecting tool badges.

Coach as the practice desk: Inside that free course, the AI Coach is the place to work one live workflow: the chooser, the scoped tools, and the stop condition. It is not a calculator on this page. Open the course, then use the Coach as the practice desk before the next agent leaves the sandbox.

Then widen the ladder: When a team can repeat the five rows without the original champion, the free courses hub and the project-based AI Fluency Certification are the next rungs. Certification is completion-based and assessed through work, not a claim of statutory accreditation. Corporate advisory stays the later conversation for organisations that need a scoped operating design after the team can already fill the rows.

What to bring to the next operating review

One page is enough: Before you approve another agent workflow, require the five authorization rows in writing. Keep blank cells visible. A missing chooser is information. A hidden blank is risk that will show up as an incident, a cleanup invoice, or a regulator question.

If the stop owner is missing, halt: If a sponsor cannot name the chooser, the scoped tools, the evidence path, the repeatability test, and the person who can stop the agent without a meeting, you are not buying speed. You are underwriting unowned action. Capital and permission should move together, or not at all.

FAQ

What is the executive mistake this report targets?

Treating more agents, watermarks, or vendor pledges as permission. Authorization needs a named chooser, a scoped tool list, and a stop owner.

Why does swarm scaling matter to a board?

A four-agent swarm can finish faster only by spending more total tokens. Speed without a chooser multiplies unowned action and cost.

Is a watermark enough?

No. A watermark can help prove origin. It does not name who may act or who can refuse the action.

What should leaders do next?

Require the five authorization rows on one page, keep blank cells visible, then practice the rows in the Free AI Fluency Course.

Sources cited

Practice, then a shared standard

Start with the Free AI Fluency Course and the AI Coach inside it. AI Fluency Certification is a four-week, project-based credential when you want a shared standard. Advisory is for organisations that need named owners, not another licence pile.

Tesseract Academy · Who Chooses What AI Gets To Do · Planning estimates only · More reports