AI agents that draft, people who decide: practical, safe AI in business software

AI in business software has moved quickly from chat windows that answer questions to “agents” that take actions: draft a purchase order, chase an overdue invoice, prepare a goods receipt. That is genuinely useful. It is also where things can go wrong, because an ERP holds records that have legal, financial and customer consequences. A wrong answer in a chat window is an inconvenience. A wrong purchase order sent to a supplier is a problem.

This article sets out a practical position: let AI agents do the preparation, and keep people responsible for the decision. It also covers the safeguards a business owner should check before trusting any agent with real data.

What an agent is, in plain terms

An AI agent in an ERP is a piece of software that looks at your records, works out what needs doing and prepares the next step. Some typical examples:

• Noticing that a sales order is short of material and preparing purchase orders from the cheapest vendor.
• Reading a supplier’s invoice, matching it to the purchase order and preparing a draft goods receipt.
• Finding customers with overdue invoices and drafting reminder messages with a payment link.
• Spotting leads that have gone quiet and drafting follow-ups.
• Summarising the business each morning for the owner.

Each of these used to take a person time to notice and time to prepare. The value of the agent is in both the noticing and the preparing.

Three modes of agent

It helps to separate agents by what they produce, because the risk is different for each.

Mode	What it produces	Risk if wrong
Report	A summary, a check or a forecast	Low: a person reads it and decides
Advisory	A ranked list or recommendation	Low to medium: it shapes a decision
Drafts for approval	A record ready to be confirmed	Medium: only if confirmed without review

Notice what is missing from that table: agents that create, send or post records on their own. For most businesses, most of the time, that mode is not worth the risk.

Why drafts beat full automation

Business records have consequences

A confirmed purchase order commits money. A goods receipt moves stock and cost. An invoice creates a tax liability. These are the records your auditor, your banker and your customers rely on. A person should own each one.

AI is good at preparation and weak at context

AI can read a messy supplier invoice faster than any clerk. What it cannot know is that you agreed a different rate with that supplier on the phone yesterday, or that the customer you are about to chase paid by cheque this morning. A person reviewing a draft brings that context in seconds.

Drafts keep the speed

The time saved by an agent comes mostly from the preparation, not from skipping the click to confirm. A purchase order that is nearly ready, with differences flagged, takes moments to approve. Keeping the human step costs very little and protects a great deal.

Safeguards worth insisting on

Before you switch on any AI agent, in any product, check for these.

1. Drafts, with differences highlighted

The agent should show what it is unsure about. A draft goods receipt should flag quantity or rate differences from the purchase order. A draft BOM should mark lines that need review. A draft that looks equally confident everywhere is hiding something.

2. The AI sees only what the user can see

If a site engineer asks the assistant about project cost and is not allowed to see financial figures, the assistant should not reveal them. The AI should act as the signed-in user, inside their company, limited by their role.

3. Numbers from records, words from the AI

Figures in an AI summary should come from the system’s own records. The AI’s job is the explanation. A good design checks the AI’s text against the numbers it was given and removes anything it cannot support.

4. Auto-send off by default

Some agents, such as payment reminders, can be allowed to send on their own. That should be a deliberate choice by an administrator, not the default.

5. Metered, logged and rate-limited

AI calls cost money. Usage should be measured, limited per user and logged, so that cost is predictable and you can see who used what.

6. Protection against misuse

Prompts to an AI can be crafted to make it misbehave. A business system should screen AI requests and give administrators a way to act on suspicious activity.

Getting started without the risk

1. Start with report agents. A daily briefing or a GST check carries no risk and builds trust in the data.
2. Add one drafting agent in the area where preparation takes the most time, often purchase or collections.
3. Review every draft for the first few weeks. You will quickly learn where the agent is reliable and where it needs a closer look.
4. Clean up the masters it depends on. Agents working from duplicate products or outdated vendor prices will draft duplicate and outdated records.
5. Only then consider auto-send, and only for low-risk messages.

Where Zaptiz fits

Zaptiz is built on the principle in the title. Across both products, agents prepare quotations, purchase orders, goods receipts and bills as drafts, and a person reviews and confirms before anything is created, sent or posted. Automatic sending, where it exists, is off by default and must be switched on by an admin. The assistant answers as the signed-in user, limited by role and, in Zaptiz Construction, by project access. AI usage is metered in credits with per-user rate limits, and every call is logged. In Zaptiz Electrical a Security Sentinel screens AI prompts and proposes account or IP actions.

Zaptiz Electrical agents include Quotation Drafting, CRM Follow-up, Smart Procurement, Invoice and Delivery Reader, Collections, Production Planner, GST Agent and Daily Briefing. Zaptiz Construction agents include the Daily Site Briefing, BOQ Generator, Procurement Agent, Invoice and Challan Reader, RA Bill Drafting and Cost Overrun Watch, with more marked as coming soon. The GST agents prepare and check figures; they never file returns.

See the full list and how each agent works on the AI page, or read the guide Using AI agents safely in Zaptiz.