Skip to main content
AI workflow agents

Let AI complete the step. Keep the approval.

An AI workflow agent uses specifically approved tools to complete a defined step in an operating process. Codefremics starts with bounded work, not open-ended autonomy.

The client decides what the agent may read, propose and do—and when it must stop for a person.

Action-taking AI

Agents take actions. Copilots advise.

Choose the interaction model before the technology. If a person should make every change, a copilot is usually the better and simpler fit.

Advisory AI

Operational copilot

Finds permitted information, prepares a summary and organises work. The user reviews the evidence and makes every change.

  • Reads approved sources
  • Prepares the review
  • The person decides and acts
Action-taking AI

Controlled workflow agent

Uses approved tools to complete a defined step, such as routing a case or preparing a follow-up, and stops where a person must approve.

  • Uses only approved tools
  • Completes one bounded step
  • Stops for approval where it matters
Defined operating work

Three bounded workflow patterns.

Each pattern is a single, repeated step with a clear owner, known records and an observable result.

01 · Intake

Case intake and routing

Check approved fields, identify missing information and route a complete case to the right queue.

02 · Exceptions

Reconciliation exception packs

Gather the approved records around a break and route review to the named owner.

03 · Follow-up

Supplier and field follow-up

Prepare an approved follow-up, log the outcome and return unresolved work to the team.

Production boundaries

Every action needs an operating rule.

These controls are agreed before an agent is allowed to act, and they are part of the delivered system, not a document beside it.

  1. 01Approved tools and permissionsList each system, record and action; grant the minimum access required.
  2. 02Approval before consequential actionPut a named person where money, access, customers or compliance are affected.
  3. 03State, retries and stop conditionsKnow where work is, limit retries and return unresolved tasks safely.
  4. 04Audit and traceabilityRetain the inputs, tool calls, approval and outcome needed to investigate a run.
  5. 05Evaluation and cost reviewTest representative work and monitor quality, volume, latency and cost.
  6. 06Failure handlingDesign for unavailable tools, incomplete data, uncertain results and incidents.
  • Readiness firstAn unsettled workflow, data path or approval point is resolved in the AI readiness assessment.
  • Evaluated before useRepresentative work is tested first. See evaluating an AI workflow.
  • Security reviewHow access, data and assurance are handled is set out in the Trust Center.
  • Delivered AI workSee the document extraction delivered for a client on the Applied AI page.
Before production

Prove the boundary.

If the workflow, data path or approval point is unsettled, begin with the fixed-scope readiness assessment. It settles what the agent may touch and where it must stop before any action is proposed.

What to bring

  1. 01The step to automateOne repeated task, where it starts and what counts as done.
  2. 02The tools and recordsThe systems the step touches and the access each one would need.
  3. 03The approverThe person who must sign off before a consequential action.
WhatsApp us