Everything up to here happened in a chat window. You uploaded something, you asked, you read the answer, you acted on it. A person was always in the loop, copying the result to where it needed to go.
Stage 3 is where that changes. The AI stops advising and starts acting - inside your systems, without a person shuttling things back and forth. This is the stage people find hardest to picture, so this article shows it rather than describes it. It is also our proof-of-capability piece: we build and run these, we do not just write about them.
Shape one: invoices that land in the ERP by themselves#
Picture the accounts process at most businesses. A supplier invoice arrives - PDF, email, or a photo on WhatsApp. Someone opens it, reads the figures, and types them into the accounting system or ERP, line by line. Every day. It is slow, it is dull, and tired data entry makes mistakes.
Here is the stage 3 version. The invoice arrives. The AI reads it - supplier, invoice number, line items, tax, total - and the data appears in ERPNext as a draft purchase invoice. No one pressed a button. A person opens the system later to a queue of drafts already filled in, checks them, and approves.
Notice what changed. The human went from typing to approving. The boring, error-prone work is gone; the judgement stays. That is the shape of a good stage 3 system - it removes the labour, not the oversight.
Shape two: a WhatsApp agent answering from live stock#
A customer messages the shop on WhatsApp at 9pm: "Are you open Sunday, and is the 50-inch TV in stock?"
At stage 1 or 2, an AI could draft a reply, but a person would still have to send it - and would have to know the live stock level. At stage 3, the agent answers on its own, and the stock part is the key: it does not guess or quote a stale number. It checks the live inventory, sees the 50-inch TV is in stock, and replies with the truth, at 9pm, with no staff member awake.
The enquiry is answered and logged. If the customer then wants to reserve it, the agent can do that too - within limits the business set.
The real work is the guardrails#
Here is the part that separates a system you can trust from a demo that embarrasses you. Stage 3 is defined less by how clever the AI is and more by how carefully its limits are drawn. A mistake here is visible to customers and touches real records, so the engineering is mostly about restraint.
- Clear limits. The agent can answer stock and hours; it cannot invent a discount or promise a delivery date it has no basis for.
- A confident hand-off. The moment it is unsure - an unusual request, a complaint, anything outside its defined job - it stops and passes the customer to a human, cleanly, without pretending.
- Nothing final without a check where it counts. The invoice lands as a draft for approval, not a posted transaction. The AI does the work; a person still owns the decision that matters.
At stage 3, the skill is not making the AI do more. It is knowing exactly where it should stop and ask for a human.
Why you cannot skip to here#
It is tempting to want this first - it is the stuff that looks impressive. But stage 3 rests on the rungs below it, and it fails without them:
- Without stage 0, there is no live stock or clean system for it to act in.
- Without stage 1, the team cannot brief or judge what the AI is doing.
- Without stage 2, the AI cannot reliably ground its answers in your real data
- and a WhatsApp agent that answers from a guess is worse than none.
A business that tries to buy stage 3 on top of a stage 0 foundation gets an expensive failure. A business that climbed the rungs gets a system that quietly does the work.
Where this leaves you#
If shapes one and two are what you want, the honest question is not "how do I get an AI agent?" It is "am I standing on the three rungs this needs?" That is exactly what next week's article - the six-point readiness check - is for.
For the full climb, see the five stages of AI adoption. This is the work we do: we build and run systems like these, and we will tell you plainly whether your business is ready for one yet.
