Custom AI agents

Connect a request to the work that follows.

Some tasks need more than a written answer: a record must be checked, information compared and an action prepared. We build agent workflows around those steps, using established models for interpretation and deterministic code for fixed business rules.

REQUEST → CONTEXT → ACTION

Where this capability helps.

01

Prepare a business action

Read a request, check the relevant records and prepare an update for an authorised person or system to approve.

02

Coordinate an exception

Identify missing or conflicting information, gather the available context and route the case to its owner.

03

Work across sources

Combine document content and permitted system records into a structured result with a trace of the decisions taken.

A delivery change with a review point.

This example design starts with a request to change an address. The workflow checks the order and the applicable rule before preparing a change. The sample below shows the decisions; it does not connect to an order system.

ILLUSTRATIVE DESIGN / SAMPLE INFORMATION

REQUEST

“Please change the delivery address on order EX-104.”

  1. 01

    Identify the request

    Extract the order reference and ask for the new address if it is missing.

  2. 02

    Read permitted records

    Retrieve the order status and the applicable address-change rule.

  3. 03

    Check the fixed rule

    In this sample, only orders awaiting dispatch can proceed.

  4. 04

    Prepare the change for approval

    Show the existing and proposed address to the authorised reviewer.

    Awaiting approval
  5. 05

    Apply the approved action

    Use the agreed write operation. Preserve a failure or an unconfirmed result for recovery.

  6. 06

    Report the confirmed result

    Completion depends on a confirmation from the order system.

What changes if the order has already shipped?

The fixed rule stops the address update. The workflow prepares a handoff with the request, order status and policy context. Approval does not bypass a rule unless that exception has been explicitly designed.

Explore the editable workflow prototype

A successful lookup is not a successful change. The workflow reports completion only after the destination system confirms the authorised action.

See this in a customer support workflow. When the request starts as a conversation, an agent can work behind a chatbot.

A workflow with a verifiable end state.

  • A defined route from incoming request to confirmed result, including exceptions.
  • Connections with agreed permissions, approval points and recovery behaviour.
  • A task evaluation and an operating plan for review, monitoring and handover.

Decisions that shape the system.

  1. 01

    Separate interpretation from permission

    A model can identify what someone is asking for. The permitted actions, required fields and approval rules are defined in the application. We agree those boundaries before connecting a tool.

  2. 02

    Make each tool specific

    A tool should have a clear input, output and failure response. Reading an order and changing its address are separate operations, with separate permissions. Available integration mechanisms determine what can be connected.

  3. 03

    Plan for incomplete work

    Missing records, rejected approvals and temporary failures need explicit paths. Where a system supports them, action references help prevent duplicate writes during bounded retries. Persistent failures remain visible for review.

  4. 04

    Keep enough context to investigate

    A useful trace records the sources consulted, proposed action, approval and confirmed result. The detail retained, access to it and retention period are agreed alongside the workflow.

How we would evaluate it.

Criteria are agreed for the intended task and tested on representative material. The items below describe the evaluation approach.

Task completion
Check the intended end state in the relevant system, including cases that should stop or be handed over.
Action correctness
Inspect tool arguments, rule checks and approvals. Include requests that are ambiguous or outside the agreed scope.
Recovery
Exercise timeouts, repeat requests and failed dependencies. Check that retries stay within limits and intended actions are not duplicated.
Operating effort
Measure review time, corrections and cost per completed task under the agreed workload.

Practical questions.

Can an agent make changes without approval?

That is a project decision, not a default capability. We define which actions may run independently, which require review and which remain outside the system. Early pilots can prepare actions while people retain approval.

Does every workflow need several agents?

No. One bounded workflow is often easier to inspect and operate. Specialised agents can help when tasks need different context or can run independently, but they also add coordination and failure paths.

Is the workflow designer a deployed automation platform?

The R&D designer is a browser-based planning tool for editing a graph and walking through sample data. A client solution needs separately implemented model calls, integrations, permissions and operating arrangements.

What would make this useful for you?

Bring one process, its source systems and a few difficult cases. We can define what the agent should do and how to judge the result.

We build client solutions using established models. Integrations, data access and operating arrangements are agreed for the project. Our proprietary model programme explores future deployment options.

Discuss your projectExplore this in a walkthrough