# Business automation
> VI: https://5ac.vn/wiki/business-automation/ · EN: https://5ac.vn/en/wiki/business-automation/ · lang: en · spec_as_of 2026-10-19 (spec no formal spec) · cluster C7
Repetitive business workflows handed to AI agents running as pipelines with owners, statuses, checkpoints and handoffs — instead of disconnected prompts nobody owns.
---
## Definition
Business automation is the practice of organizing repeatable work into stateful pipelines where AI agents carry each step while people keep approval rights on the high-risk ones. Unlike disconnected prompting, every task has an owner, a definition of done, and an accountable checkpoint.
## Mechanism
Disconnected prompting breaks down because operations are chains of dependent actions: a lead enters the system, an agent qualifies it, an owner reviews it, the CRM updates, marketing triggers a nurture sequence, and reporting flows back to the CEO. Without a shared status board, the team sees scattered outputs instead of workflow movement. Kanban forces the four questions every serious automation system answers: who owns the task, what stage it is in, what "done" means, and who is accountable when something breaks.
| Ad-hoc approach | Stateful pipeline |
| --- | --- |
| Prompt an agent one task at a time | Create a card with an owner and completion criteria |
| Results scattered across chats | Status lives on one shared board |
| Hard to see where work is stuck | Bottlenecks and approval queues become visible |
| Review happens at random | Review becomes a required checkpoint |
| Cycle time unmeasurable | Time in process, backlog and completion rate are measured |
The minimal board for an SMB has five columns: Intake (new requests arrive), Agent Running (the agent drafts or analyzes), Review (a person checks quality and risk), Approved / Scheduled (cleared and queued for the next step), Done / Logged (finished and recorded for measurement). Background agents run on a schedule — recurring reports, reminders, alert triage — under one rule: automate where frequency is high and decision authority is low, and leave a log of what was read, flagged and sent on every run.
Three numbers separate a system that compounds from one that burns team time: cycle time (card created to done), review pass rate (work that clears review first try), and blocked rate (cards stuck on missing data, missing owner or missing permission). On the safety side, the principles are least privilege per agent, agents split by role rather than one do-everything assistant, approval checkpoints near money and customer data, and a mandatory audit log.
Flow: Intake → Agent Running → Review → Approved / Scheduled → Done / Logged
## Application context
For Vietnamese SMBs, the best starting point is a revenue-adjacent pipeline that is not yet sensitive: receive leads, qualify intent, draft the first reply, update the CRM, create follow-up reminders. After two to four weeks of measuring first-response time, leads dropped, and work sitting in review, expand into content, reporting or post-sale care. The larger prize is the operating layer: as many agents run in parallel, the competitive edge shifts from the cleverest prompt to the ability to organize stateful work — which is exactly where a multi-agent Kanban board beats a shelf of separate chatbots. Contracts, pricing and internal data should travel through infrastructure the business controls.
## Sources fetched for this entry
- — C6 pillar: multi-agent pipelines, 5-column board, three metrics — accessed 19 Oct 2026
- — background agents, high-frequency/low-authority rule, 30-day path — accessed 19 Oct 2026
- — six safe-automation principles and the access matrix — accessed 19 Oct 2026
- — English pillar on business automation pipelines — accessed 19 Oct 2026
- — English background-agents piece — accessed 19 Oct 2026
- — English safe-automation playbook — accessed 19 Oct 2026