When AI can perform the work: designing for bounded delegation
For most of the recent history of workplace AI, the person remained the engine of the process. They opened the tool, wrote a prompt, moved information between systems, checked the response and decided what to do next.
That pattern is changing.
Newer AI systems can carry a brief across several steps. They can research, work through files, use browser applications, write code and create finished documents. OpenAI describes GPT-6 Astra as a model for complex reasoning, computer use, research, coding and document creation. Its model guidance says it can carry out multistep workflows across code, browsers and professional software.
Astra is one current example of a broader development across AI platforms. The management question is becoming more immediate: how do you delegate work to a system that can act?
The answer starts with bounded delegation.
A prompt is becoming a work brief
Consider how a business development manager prepares for a new account.
The work might include researching the company, identifying relevant people, reading recent announcements, reviewing previous contact, preparing an account brief, updating the CRM and drafting an email. An employee using a chatbot can complete each step faster. They still carry the workflow in their head and move the work forward.
An AI agent can potentially take responsibility for a larger portion. It can gather the approved information, work through the sequence and return the completed brief with a draft email. With the right connection and authority, it may also prepare the CRM update.
This is a larger unit of delegation. It needs more than a clever prompt.
A useful work brief tells the system:
- The outcome it is responsible for
- The sources it may use
- The systems it may access
- The decisions it may make
- The actions that require approval
- The standard the output must meet
- What to do when the normal process does not apply
Managers already use these ideas when delegating to people. AI makes the gaps harder to ignore because it cannot rely on years of informal knowledge, team habits and corridor conversations.
What bounded delegation means
Bounded delegation gives an AI system responsibility for a defined piece of work within explicit limits.
The boundary matters because business processes contain different kinds of responsibility. Collecting public information carries a different consequence from changing a customer record. Drafting an email carries a different consequence from sending it. Flagging a contract clause is different from giving a legal conclusion.
A workflow can give AI considerable room to act while reserving particular decisions for a person. For example, an agent might:
- Research a prospective client using approved public sources.
- Compare the findings with the business's qualification criteria.
- Prepare an account brief and suggest a priority.
- Draft the CRM changes without submitting them.
- Pause for the account owner's review.
- Apply the approved changes and record what it did.
The boundary sits around the consequential actions. The person remains accountable for the relationship and the decision. The AI carries much of the preparation and administration.
This structure also makes testing possible. You can inspect the sources, compare the brief with the criteria, review the proposed record changes and measure the time saved.
The workflow has to become visible
Many business processes work because experienced people quietly compensate for them. They know which spreadsheet is current, which client needs a different format and who must approve an unusual request. Much of that knowledge has never been written down.
An agent encounters those gaps directly.
If 3 source documents disagree, the workflow needs a rule for precedence. If a missing field should stop the process, that needs to be clear. If a manager must approve any commitment above a certain amount, the approval belongs inside the workflow.
This does not require documenting every possible event before a pilot begins. It requires enough clarity for the normal path, the likely exceptions and the decisions carrying real consequences.
Our guide to choosing your first AI workflow covers how to identify a contained process and establish a baseline. The next step is describing that workflow well enough for responsibility to be shared deliberately between people and AI.
Give the agent a job description
A useful agent job description has 7 parts.
1. Outcome
Define the finished result. “Help with reporting” leaves too much open. “Prepare the weekly project report using the approved template and submit it to the project manager for review by 3 pm Thursday” gives the work a clear end point.
2. Context
Identify the information needed to do the work. This may include previous examples, client instructions, business rules, templates and definitions. Context also needs an order of authority when sources conflict.
3. Tools and access
List the systems the agent may use and the permissions it receives. Read access, draft access and permission to commit a change are materially different.
4. Decision rights
State which choices the agent can make within the process. A system may select the right template or classify a routine request. Pricing changes, contractual commitments and sensitive people decisions will usually need human authority.
5. Review points
Place approval where it protects the outcome. Reviewing every minor action can remove the benefit. Waiting until an irreversible action has occurred creates unnecessary risk. The right checkpoint usually sits immediately before the action creates a consequence.
6. Quality test
Define how someone can tell whether the work is acceptable. A report might need all required sections, figures traced to source data, correct comparison periods and commentary that follows the organisation's writing standard.
7. Exceptions
Tell the agent when to stop, ask or escalate. Missing information, conflicting records, unusual client instructions and low confidence are normal operating conditions. They need a route back to a responsible person.
The human role becomes more specific
As AI takes on more preparation and execution, people spend more time on judgement, exceptions, relationships and improvement.
This works well when the division is deliberate. A project director may want an agent to collect updates, compare them with the programme and draft a status report. The director still interprets the political context, challenges an optimistic forecast and decides what the client needs to hear.
The technology can also reveal work that should disappear. If an agent spends hours moving the same data between two systems, the better answer may be a direct connection or a change to the process. Workflow design should improve the work itself, rather than preserve every inherited step.
People who perform the current process need a central role in this design. They understand the exceptions and informal decisions that rarely appear in a procedure. Their knowledge determines whether the new workflow survives normal use.
Governance belongs inside the workflow
Policies provide useful boundaries. Day-to-day control comes from the design of the work.
Access should match the task. Sources should be approved. Sensitive actions should create a record. Approval should sit before commitments, payments, publication or changes that are difficult to reverse. A named person should own the workflow and review failures.
These controls can be proportionate. A research agent using public sources needs a lighter structure than an agent working with payroll or customer accounts. The consequence of an error should determine the level of oversight.
Security also includes the instructions and information an agent encounters. Browser-based work can expose an AI system to unreliable pages, misleading instructions and content designed to influence its behaviour. Clear source rules, limited permissions and human review help contain that risk.
Measure the finished work
AI adoption dashboards often report licences, active users and message counts. Those numbers can show participation. They cannot establish whether a business process improved.
Measure the workflow before and during the test:
- Total cycle time, including waiting
- Staff time required
- Completion rate
- Output quality against the agreed test
- Rework and corrections
- Number and type of exceptions
- Errors or incidents
- Experience of the people doing and reviewing the work
The comparison should use several real examples, including difficult ones. A polished demonstration proves that the workflow can work once. Normal use shows whether it can become part of the business.
A sensible first test
Choose one repeated digital workflow with a clear owner and a reviewable outcome. Map the current steps and record a baseline. Then write the agent job description.
Start with limited permissions. Let the agent research, prepare and propose. Add execution only after the team understands its failure patterns and can see a reason to grant more authority.
Run the workflow often enough to learn. Review every exception during the test. Update the instructions, sources and approval points as evidence accumulates. At the end, decide whether to adopt it, revise it or stop.
GPT-6 Astra gives the current conversation a useful sense of urgency. The lasting task for business leaders is building the management practice around delegation to AI.
The ChatGPT Business guide to GPT-6 Astra covers the model, its availability and its business capabilities in more detail.
Addaptive helps Australian businesses examine real workflows, decide what people and AI should each handle, and test the new way of working under normal operating conditions. Start with one workflow.