AI agent for company-specific work
The Custom Agent is scoped to work that the standard agent categories do not fully cover.
A Custom Agent is designed for an internal or customer-facing workflow with a recurring task, identifiable data, definable operating limits and a goal.
- Where it helps
- Company-specific work
- What to improve
- One clear improvement with an agreed metric
- Where people decide
- Access and action limits are defined per use case.
Summary
- Where this fits
- The workflow is scoped to a single use case. The operating logic, access rights and metric are defined accordingly.
- What the agent does
- Tasks are scoped by use case. Below are examples of possible responsibilities.
- Where a human approves
- Process ownership, operating limits and exceptions are always defined per use case.
- What to measure
- Measures are examples to track, not promised results.
Where the workflow needs its own solution
Standard agent categories cover only part of the work, and the rest is left to manual work.
- Work moves between several systems and people
- Information is repeatedly checked and combined manually
- The next step depends on rules, exceptions and approvals
Show more (1)
- A standard agent category covers only part of the work
What the agent can do in practice
Tasks are scoped by use case. Below are examples of possible responsibilities.
- Receiving and structuring information
- Checking agreed conditions or rules
- Preparing a document, task or proposal
Show more (2)
- Identifying and escalating exceptions
- Updating agreed systems after approval
How the work moves forward
Defined use case
Inputs and data sources
Rules and operating logic
Approvals and exceptions
Measurable outcome
What data and systems the agent can use
The systems and data sources required by the use case are agreed and scoped case by case.
Data examples
- Customer or employee data for an onboarding use case
- Received documents and attachments
- Order or delivery data
- Inputs for internal reporting
- Checklists approved for the use case
Example systems
- Agreed systems required by the use case
The agent uses only the approved data sources required for the task.
Human control and agent limits
Process ownership, operating limits and exceptions are always defined per use case.
The agent can
- Analyze agreed data
- Combine agreed context
- Prepare a proposal
- Suggest the next action
- Create an internal task
- Execute a limited action, if it has been separately approved
The agent does not do independently
- Does not make final decisions without approval
- Does not execute actions outside agreed limits
- Does not replace the owner of the customer relationship
- Does not handle high-risk exceptions independently
Human responsibility
- Process ownership
- Operating limits
- High-risk decisions
- Handling exceptions
- Goal and ongoing development
Possible measures
Measures are examples to track, not promised results.
Possible measures
- Process turnaround time
- Number of manual steps
- Identification of gaps or exceptions
- Work queuing
- Number of approved actions
- A quality metric defined for the use case
When this fits, and when it doesn’t yet
This fits when
- The work is recurring
- The start and outcome can be defined
- Data and access rights can be scoped
- Exceptions have an owner
- Success can be measured
Not a fit yet when
- The work keeps changing and has no clear owner
- The goal is only to "use AI"
- The required data or access rights cannot be defined
- The agent is expected to make unsupervised high-risk decisions
This agent's setup:
Custom AI Agent
Do you have work that standard agent categories don’t cover?
The agent assessment helps scope the use case, data and measurable goal.