Personalized AI infrastructure
Company Superintelligence Delegation Tiers: What to Hand Off First
A tiered ladder for delegating work to company AI agents, from drafting to supervised action, with the evidence each step must show before you move up.

Company superintelligence works best when you delegate in tiers: start with drafting that a person reviews, move to summarizing and recommending, and only then allow supervised actions in your systems. Each tier is earned by evidence from the tier below it, not by a vendor demo or a good week. The first task to hand to an agent is one with a clear output, a reviewer who already knows what good looks like, and a failure that is cheap to catch.
Key takeaways
- Delegate in five tiers, from drafting to supervised action, and move up only one tier at a time.
- Choose a first task where a named person can judge the output quickly and confidently.
- Record the evidence for every tier change in writing, so the decision can be reviewed later.
- Widen permissions only after the agent has been reviewed at the tier below.
- Keep a rollback path and a named owner in place before any agent touches a live system.
Why delegate in tiers instead of all at once
Most companies make the same mistake with AI agents. They connect an agent to email, the CRM and the file store in the first week, then discover that nobody agreed on what it may do without asking. The problem is rarely the model. The problem is that the company skipped the steps where it would have learned how the agent fails.
A tiered approach slows the first decision down on purpose. Each tier gives you a narrower blast radius and a clearer signal. When an agent writes a draft that a manager rewrites, you learn about tone, facts and gaps in one place. When an agent takes an action in a live system, you learn about the same things, but a mistake now has a customer, a payment or a record attached to it. The further up the ladder a task sits, the more evidence you should want before you climb.
Tiers also protect the people who work with the agent. Staff who see the agent handled carefully at low stakes are more willing to use it at higher stakes. For guidance on getting people to use an agent in daily work, see How to Get Your Team to Actually Use an AI Agent.
The five tiers, from drafting to supervised action
The ladder below is a practical sequence. Each tier adds one new capability, and each capability brings a new kind of risk. Do not skip a tier because the demo looked convincing.
- Tier 1, Draft: the agent produces text, a plan or a first version that a person edits before anything leaves the company.
- Tier 2, Summarize and classify: the agent condenses documents, sorts incoming requests or tags records, and a person spot-checks the labels.
- Tier 3, Retrieve and recommend: the agent answers questions from approved company sources and suggests a next step, with the sources shown for each answer.
- Tier 4, Act on reversible steps: the agent creates drafts in systems, updates internal notes or schedules items, where every change can be undone and is logged.
- Tier 5, Supervised action: the agent executes changes in live systems such as sending messages to customers, issuing refunds or editing master records, under explicit approval rules and monitoring.
Notice that Tier 3 is where most of the useful value appears for many operations teams, because retrieval from trusted sources answers recurring questions without any write access. Tier 5 should be the exception, reserved for tasks where the rules are stable and the cost of an error has been measured.
How do you test whether a task is ready for the next tier?
Use a short, repeatable procedure. It works the same whether you are moving a drafting task from Tier 1 to Tier 2 or preparing a Tier 4 action for review.
- Step 1: Write the task as one sentence with an input, an output and a person who receives the output.
- Step 2: Build a set of real examples, including ordinary cases and at least a few awkward ones, and agree on what a correct answer looks like.
- Step 3: Run the agent on the set at its current tier and have the reviewer grade each output as usable, usable with edits, or wrong.
- Step 4: Group the failures by cause, such as missing context, wrong source, unclear instruction or a tool error.
- Step 5: Fix the causes that are inside your control, rerun the set, and compare results.
- Step 6: Decide, in writing, whether the task moves up one tier, stays where it is, or is stopped.
- Step 7: Name the person who owns the decision and the date of the next review.
The reviewer should not be the person who built the agent. Independence matters because builders tend to grade their own work generously. The same logic underlies the practice of evaluation sets described in Anthropic's guidance on [demystifying evals for AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents).
What evidence earns each tier?
Evidence should be specific enough that a second person could reach the same decision from the same records. Impressions do not qualify. A useful evidence pack for a tier change usually contains four things: the evaluation set and its date, the grading results broken down by failure cause, a log sample showing what the agent actually did, and the list of permissions the agent holds at the time.
Logs matter more than most leaders expect. You need to see which tool the agent called, what input it received, what it returned and which human approved it. Observability concepts from OpenTelemetry, including traces that connect each step of a run, give you a vocabulary for this. The [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/) is a useful reference for the people who will maintain the logs.
The NIST AI Risk Management Framework is also worth reading before you set thresholds. It frames risk management as a continuing activity of mapping, measuring and managing, which matches the way tiers should be revisited over time. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) gives your steering group language to describe what a tier decision accounts for.
A practical rule: do not move a task up a tier because the agent performed well on a few recent cases. Move it up when the evaluation set has been run after the latest changes, the failures are understood, and the people who will be affected by the next tier have agreed to it.
Permissions and approval boundaries at each tier
Each tier should come with its own permission set. Tier 1 agents need read access to the material they draft from and no write access anywhere. Tier 3 adds read access to approved sources. Tier 4 adds write access to a limited area, such as internal drafts. Tier 5 adds specific actions with approval rules. Grant access to the agent as a non-human identity with the narrowest scope that covers the task, and remove scopes that were added for a test and never taken away. The OWASP [Authorization Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) gives general principles for checking access on every request, not just at setup.
Retrieved content also needs care. A tool that reads an email, a web page or a support ticket may receive instructions hidden inside that content. The OWASP guidance on [prompt injection prevention](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) describes how to limit what untrusted text can cause the agent to do. At Tier 3 and above, treat any text pulled from outside the company as data, not as instructions.
Illustrative scenario, not a real case: a support operations lead wants an agent to answer routine questions about delivery dates. At Tier 3, the agent reads the order status and the published shipping policy, then drafts an answer with the source fields shown. A person sends each reply. After two review cycles, the lead sees that the agent confuses two shipping regions. The team adds the region field to the retrieval step, reruns the examples, and only then considers whether the agent should send replies on its own for a narrow category, such as tracking requests for orders already marked as shipped. Even then, refunds and address changes stay outside its permissions.
How a managed setup keeps the ladder honest
Be Human designs, implements and operates personalized AI infrastructure, so tier decisions are part of the operating model rather than a one-time launch choice. An engagement starts by defining one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion. Monitoring, recurring evaluations, permission review and controlled changes then run against that definition. When a team wants to raise an agent's tier, the change goes through the same request and approval path used for any other change, and the evidence pack is kept with the agent's records. Service is remote and available in English and Portuguese.
The practical benefit is continuity. A company does not have to rebuild its reasoning about permissions each time a person leaves or a new tool appears. For a closer look at how change requests move through approval, see How to Request and Approve Changes to a Managed AI Agent.
Choosing the first task worth handing over
Start with work your team already does repeatedly, where a reviewer can judge the result in minutes and where a wrong answer is inconvenient rather than damaging. Good candidates include first drafts of internal reports, classification of incoming requests, and answers to questions drawn from a fixed set of approved documents. Poor candidates include anything that moves money, changes a legal record or speaks for the company to a customer without review, at least until the lower tiers have produced clear evidence. Before you choose, it helps to check your permission model and your review capacity, as described in How to Run a Permission Review for Your AI Agents, Step by Step.
If you are not yet sure which task belongs on the first rung, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
How many tiers should a company use? Five is a useful starting point, but the number matters less than the discipline of moving one step at a time. A smaller company may collapse Tiers 2 and 3 into one stage. What matters is that every stage has a written boundary, a reviewer and a condition for moving up.
How long should a task stay at one tier before it moves? There is no fixed period. A task moves when the evaluation set has been rerun after the latest changes, the failure causes are understood and the owner agrees in writing. Some tasks stay at drafting for a long time because the draft is the right outcome.
Should the agent ever skip directly to Tier 5 for a simple task? It is better not to. A simple task can still touch a customer or a record in ways nobody anticipated, and a skipped tier removes the evidence you would have collected. If the task is truly low risk, it can still pass through Tiers 3 and 4 quickly with a short review at each step.
Who decides when a task moves up a tier? The decision should belong to a named business owner, with the operations lead or the technical maintainer preparing the evidence. Someone independent of the build should review the evidence before the change is approved. Record the decision, the date and the next review point so the choice can be examined later.
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.


