A demo of the fictional company NovaWorks. It reproduces a real business workflow and uses no real customer data.
One task, two ways
Open a service request for a contract and draft its first invoice
AS-IS
A person moves across three systems
0:00 / 1:45
Start
The task
A NovaWorks operator handles the activation request and first invoice draft for Hanbit Tech contract CTR-1001 by hand. The counters on the right add up the inputs and systems.
Form inputs
0 / 12
Systems visited
CRMServiceERP
Steps
Select a step to jump to that scene.
→ The same contract details were typed into forms 12 times across 3 systems.
TO-BE
One request, one approval
“Prepare activation for Hanbit Tech contract CTR-1001 and draft the first month's invoice.”
0:00 / 1:19
Start
The task
This time an AI agent handles the same task. The person makes one request and one approval.
Person· AI
Request0
Read the request
Pick a Skill
Look up CRM contract
Check Service status
Check ERP status
Draft the plan
AI preparation · 6 steps
Approval0
Service tenant
Service activation request
ERP billing account
ERP first invoice draft
AI execution · 4 steps
Verify ERP and Service
Steps
Select a step to jump to that scene.
→ The work was requested once and the plan approved once.
AS-IS · connected by a person
12 inputs · 3 systems
TO-BE · connected by an AI agent
1 request · 1 approval
Both ways end in the same result: one activation request and one first invoice draft.
Videos are paced for explanation. Durations are not productivity figures. During recording, one request-interpretation step used a fixed response instead of the live model.
Know-how lives in Skills. Execution stays in your systems.
Language→Skill→CLI→API
AI Agent
Interprets the request and picks the right Skill and customer.
Skills
Define the order, required data, approval rules and completion criteria.
CLI
A separate process per system returns results as JSON.
OAuth API
Each system validates user permissions and inputs itself.
Execution service
Stores plans and approvals, tracks stages, prevents duplicates and resumes.
Why you can hand it to AI
Confirm before running
No change API can be called until a person approves the plan.
Resume where it stopped
If a system fails, finished stages are kept and the failed stage runs again.
Tracked to the result
Completion is verified against each system's real data.
Exceptions follow the rules too
Customers with the same name
BehaviorShows the candidates and asks which one
Unsigned contract
BehaviorBlocks any change
Read-only permission
BehaviorReads work; changes are blocked by the API
ERP outage
BehaviorResumes from the failed stage, keeping finished ones
Repeated request
BehaviorReturns the existing run without creating duplicates
Situation
Behavior
Customers with the same name
Shows the candidates and asks which one
Unsigned contract
Blocks any change
Read-only permission
Reads work; changes are blocked by the API
ERP outage
Resumes from the failed stage, keeping finished ones
Repeated request
Returns the existing run without creating duplicates
The demo stops at filing the activation request and drafting the invoice. It does not activate services, issue invoices or send email, and it is not connected to enterprise SSO.
How adoption starts
Connect
Pick one repetitive task and connect your systems' APIs and CLIs.
Define
Define the procedure as a Skill and verify permissions, approvals and completion rules.
Expand
Grow from stable flows and integrate with your real SSO and approval systems.
The same idea in a real project
In the automotive parts case, Ops CLI — the system built for the VIBE OPS methodology — made AI a participant in development work.