Retire Legacy Work Before Adding Superintelligent Agents

To prepare for an AI agent, retire the legacy process it would sit on. Map each manual step that repeats across people and tools, choose one source of truth for every fact, and name one owner for each step that survives. An agent built on top of duplicated work performs the waste faster and leaves no one clearly accountable for the result.

Key takeaways

• Automating a process nobody owns turns hidden duplication into faster output that is harder to trace.
• Map the work as people actually perform it, not as the procedure document describes it.
• Choose one source of truth for each fact and write down the reason for the choice.
• Retire duplicate steps in a fixed order, keep a log, and remove the access that the old steps used.
• Hand a task to an agent only when its inputs, owner, decision rights and success check are written down.

Why duplicated manual steps become a risk for agents

A team that has run the same workflow for years usually holds several versions of it at once. One spreadsheet tracks status, a project tool repeats it, and an email thread confirms it again so the person who started the work feels safe. Each copy made sense when it was added. Together they form a process that nobody designed as a whole.

An agent does not see this history. Given access to the tools, it treats each copy as valid, reads whichever one it reaches first, and writes back to several of them. It cannot tell which approval is binding, which reminder exists only because a former manager liked reminders, or which field is a leftover from a system that was switched off. It will repeat every one of these habits consistently and at scale, including the ones that should have ended long ago.

Duplication also hides ownership. When two people update the same customer status in different places, each assumes the other holds the record. The error stays invisible until someone asks who is accountable, and the honest answer is that nobody is. Anthropic's engineering guidance on building effective agents makes a related point: simple, well-defined workflows are easier to make reliable than open-ended autonomy. Retiring legacy work is the moment to make the ownership answer explicit.

How do you find duplicated steps before anyone builds?

Start with the work as it is performed. Procedure manuals describe what should happen, while the people who do the work know what actually happens. Ask them to walk through the last few real cases of one task, step by step, naming every tool they open and every person they message.

Use a short sequence for the walkthrough:

• Pick one workflow with a visible outcome, such as preparing a quote, following up an unpaid invoice or completing onboarding paperwork.
• Trace three recent real cases with the person who runs them, and note every tool, file and message involved.
• Label each step by its purpose: it makes a decision, moves information, repeats a check already done, or exists only because an earlier process needed it.
• Mark every fact stored in more than one place, and note which copy other teams treat as the truth.
• Count the handoffs, and record who receives each one and what that person does with it.

Drawing the surrounding systems on one page makes the duplicates easier to see. The method in Draw Your AI Stack: A One-Page Architecture Map for Leaders suits this step well, because it shows where each fact enters the business and where it is read.

Decide which copy is the source of truth

For each duplicated fact, choose one system of record. The choice should follow the people who depend on the data downstream, not the tool that is newest or most pleasant to use. Write the decision down with its reason, its owner and the date it was made. If the group cannot agree, treat the disagreement as the most important finding of the exercise. An agent would otherwise have to make that choice silently, every time it runs.

Use these rules as a starting guide:

• If the fact drives a payment, a contract or a customer commitment, the system of record is the one finance or legal already reconciles against.
• If the fact only coordinates people, a shared task list is enough, and the duplicate tracker can be retired.
• If two teams need different views of the same fact, keep one record and give each team a read view rather than a second copy.
• If nobody can name a reliable copy, pause the workflow and resolve ownership before any automation begins.

Illustrative scenario: A small services firm records each new client in a shared spreadsheet, a project tool and an accounting folder. Someone updates two of the three copies by hand each week. The team agrees that the accounting record is authoritative, runs the old spreadsheet in parallel for a short, defined period, and then retires it. The project tool becomes a read view. Nothing has been automated yet, and the weekly reconciliation disappears.

Retire steps in the right order

Retirement should move from the lowest risk to the highest. Start with steps that produce nothing anyone uses, such as a status report that goes unread or a copy of a file kept in case it is needed someday. Leave steps that people rely on for last, and only after the replacement has produced matching results for a defined period.

Keep a retirement log with four fields for each step: what was removed, why it was removed, who approved it, and how it could be restored. The rollback path matters. Teams often switch off a step by ignoring it, which leaves the old path technically alive and quietly used. A step is retired only when its inputs, outputs and access are actually disabled.

Changes like these also affect people. The habits built around the old work need to be explained, and the new record needs to be easy to find. The adoption side of this transition is covered in How to Get Your Team to Actually Use an AI Agent.

Assign an owner to each surviving step and close old access

Every step that remains needs one named person accountable for its result. A team name is not an owner. The owner approves changes to the step, reviews its exceptions and can say, in one sentence, what the step is for and what happens if it fails. When an agent later takes over part of the step, ownership does not move to the agent. The agent works inside the boundary the owner has set, and the owner keeps the decision.

Retiring a step should also retire the access it used. Old service accounts, shared logins, forgotten integrations and automations that still hold a token often keep working after the process they served has ended. These are exactly the paths an agent would inherit by default. The principle in OWASP's authorization guidance, granting only the access a function needs, is a good test. For each surviving step, list the permissions it requires and remove the rest before the agent is connected.

What should you check before an agent takes over a task?

Be Human's managed approach begins with this inventory before any agent is configured. A first engagement defines one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion. The retirement work sets those boundaries, because the agent should receive only the surviving steps and their owners. Managed operations then keep the agent's permissions, channels and operating context under review as the process changes, so the record of who owns what stays current. The NIST AI Risk Management Framework treats mapping context and assigning responsibility as core activities, and this inventory is a practical way to carry them out.

Confirm these points before connecting the agent:

• Every input the agent reads comes from a named system of record.
• Every output has one owner who reviews exceptions.
• Each retired step is logged, and its old path is switched off rather than left to be ignored.
• The agent's permissions match the surviving steps and nothing beyond them.
• The success criterion fits in one sentence that someone outside the project can check.

The permission side of this check is covered step by step in How to Run a Permission Review for Your AI Agents, Step by Step. If the process is still unclear after this inventory, stop and clarify it first, because an agent will not resolve the ambiguity for you. Take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Can we let the agent handle the duplicate steps and clean them up afterward? Clean-up after automation is harder than it looks, because the agent will have written to every copy it could reach and the owners may be unclear. Retiring duplicates first keeps the agent's first writes aimed at one record. It also means the team relearns the process from its real steps rather than inheriting its old habits.

How many people should take part in the mapping session? Keep the group small: the person who runs the workflow, one person who receives its output, and one person who can approve retirements. Add a specialist for tools or data access only when the walkthrough reaches a system nobody in the room can explain. Large groups tend to describe the official process instead of the real one.

What if two departments insist that their copy is the real one? Treat the disagreement as a decision to make, not a detail to hide. Bring the people who depend on the data downstream into one conversation, and decide according to what reconciliation, payment or customer commitment relies on. Record the decision and its reason so the same argument does not return once the agent starts running.

Does retiring legacy work delay the agent project? It usually moves the effort to the start, where it is easier to absorb. A short inventory, a clear source of truth and a small set of owners let the first agent role stay narrow and testable. Without them, the project often stalls later, when someone asks who approved the agent's last change.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.