Enterprise AI Operating Capability

Piloting AI in Complex Workflows Start with One Workflow

Break multi-team handovers into testable steps. Compare delivery quality, waiting time and rework.

Discuss a Pilot

If a quotation stalls, find out whether the issue is missing information, failed review or unavailable approval capacity before deciding where AI can help.

The Operating Problem

Name the operating goal and the people who use the result. Choose one workflow with observable outputs and manageable risk.

Trace a real deliverable back through the workflow. Record its inputs, owner, acceptance criteria and downstream recipient.

Manage Handovers

Set a Human Fallback

Check two things at handover: whether the upstream output passes, and whether the next step is ready. Recording both shows whether work needs correction or is waiting on a dependency.

Test Repeatability

Compare quality, time, rework and cost before and after the change. Track manual intervention and exceptions alongside generation speed.

Start with the Use Case

Set the scope, owner, baseline and stop conditions before the pilot. Continue observing if evidence is unclear, and roll back if risk grows.

Connect the Steps

Separate observed causes from hypotheses in the review. Write tested changes back into SOPs, responsibilities, playbooks and skills, with version ownership and rollback.

Keep a Record and Review

Agree on interfaces, delivery, acceptance and dispute handling with external partners. Each organization can keep its own process while sharing a standard at the handover.

Compare on the Same Basis

Possible deliverables include a process map, a human–AI responsibility matrix, acceptance criteria or a pilot review. Confirm deliverables and timing after discussing scope.

Who Should Be Involved

V2.0 is an interim research framework. Test implementation conditions and business impact on real work, then use the evidence to decide whether to expand.

Use four levels to describe results: business outcomes, processes, functional responsibilities and individual actions. Then verify that the work at each level contributes to the overall goal.

Name the operating goal and the people who use the result. Choose one workflow with observable outputs and manageable risk.

Trace a real deliverable back through the workflow. Record its inputs, owner, acceptance criteria and downstream recipient.

How to Map the Workflow

Use four levels to describe results: business outcomes, processes, functional responsibilities and individual actions. Then verify that the work at each level contributes to the overall goal.

Use the same process definition for role responsibilities, playbooks and performance criteria. Describe who handles exceptions, what authority they have and how decisions are reviewed.

Check two things at handover: whether the upstream output passes, and whether the next step is ready. Recording both shows whether work needs correction or is waiting on a dependency.

Start with the Use Case

Compare quality, time, rework and cost before and after the change. Track manual intervention and exceptions alongside generation speed.

An AI skill brings together the knowledge, tools and rules for a task. The human playbook and the skill share acceptance criteria and specify when a person must take over.

Make the Method Reusable

Ask another colleague to repeat the task using the playbook or skill. Note gaps that still need verbal explanation. Expand only when the process can be repeated reliably.

Define Authority

Set rules for data scope, access and retention before using sensitive information. Escalate to an authorised person when access, inputs or risk limits are unclear.

Delivery, Authority and Review

Agree on interfaces, delivery, acceptance and dispute handling with external partners. Each organization can keep its own process while sharing a standard at the handover.

Bring one live workflow to the discussion. Describe the desired outcome, the bottleneck, the roles involved and the samples available for comparison.

Possible deliverables include a process map, a human–AI responsibility matrix, acceptance criteria or a pilot review. Confirm deliverables and timing after discussing scope.

Methods and Further Reading

V2.0 is an interim research framework. Test implementation conditions and business impact on real work, then use the evidence to decide whether to expand.

Frequently Asked Questions

How Do We Choose the First Workflow?

Name the operating goal and the people who use the result. Choose one workflow with observable outputs and manageable risk.

Do We Need New AI Tools First?

Start with tools you already have after checking data permissions and acceptance standards. Evaluate purchases only if there is a genuine capability or integration gap.

How Should Sensitive Data Be Handled?

Set rules for data scope, access and retention before using sensitive information. Escalate to an authorised person when access, inputs or risk limits are unclear.

How Do We Know the Pilot Helped?

Compare quality, time, rework and cost before and after the change. Track manual intervention and exceptions alongside generation speed.

When Is It Ready for a Wider Rollout?

Ask another colleague to repeat the task using the playbook or skill. Note gaps that still need verbal explanation. Expand only when the process can be repeated reliably.

Bring a Problem to Discuss

Use four levels to describe results: business outcomes, processes, functional responsibilities and individual actions. Then verify that the work at each level contributes to the overall goal.

Discuss a Pilot

Amber Liu

Name the operating goal and the people who use the result. Choose one workflow with observable outputs and manageable risk. About Amber Liu