11 AI agents that run your back office

The work that never gets done

Every small manufacturer carries a backlog of jobs that are important but never urgent enough to beat today's fire. Following up the quotation sent three weeks ago. Chasing the invoice that went 20 days overdue. Checking whether the components for next week's panels are actually on order. Reading the GST position before the filing date rather than on it.

None of these are hard. They are simply more work than the two or three people running the office can reach, and they are the work whose absence quietly costs the most: the lost repeat order, the cash sitting in someone else's bank account, the panel that stalls for one part.

That is the gap the agents are aimed at. Not replacing the estimator or the accountant, but doing the patient, repetitive chasing and checking that nobody has time for.

What each agent does

All eleven are included on every paid plan. Each can be turned off per company if it is not useful to you.

• Quotation drafting agent — Takes a description, photo or drawing and produces a complete draft quotation with a costed bill of material behind it.
• Smart procurement agent — Sweeps daily for shortages against open sales orders and low stock, picks the cheapest capable vendor from your price lists, and drafts requisitions for approval.
• Collections agent — Chases overdue invoices with escalating reminders and a payment link attached, from a gentle nudge to a final notice.
• CRM follow-up agent — Keeps cold leads warm with escalating, individually written emails instead of one forgotten follow-up.
• Production planner agent — Ranks open jobs into a suggested build order and flags what is late-risk or blocked on material. Advisory only.
• Daily briefing agent — A morning summary for the owner: what came in, what shipped, what is overdue, and what looks anomalous.
• GST compliance agent — Computes GSTR-1 and GSTR-3B, validates GSTINs, place of supply and HSN codes, and flags input tax credit drift.
• Invoice and delivery reader — Reads supplier invoices and delivery notes, matches them to the purchase order, flags over-billing and drafts a goods receipt note.
• WhatsApp sales agent — Answers enquiries and captures leads on WhatsApp within the 24-hour reply window, handing over to a human when it should.
• Ask-your-data analytics — Answers questions about your own numbers — sales trends, top customers, receivables ageing — with a chart rather than a paragraph.
• Security sentinel — Watches for suspicious access patterns, maintains an IP blocklist and records incidents for review.

Where the line sits between suggestion and action

An agent that quietly emails your customers or raises purchase orders on its own is a liability, however clever it is. The rule in Zaptiz is that anything with an external consequence produces a draft in a waiting-for-approval state: procurement proposals, requisitions, quotations, collection emails.

Agents that only read and report — the daily briefing, the GST health check, the production planner, the analytics — run freely, because there is nothing to approve and no way for them to do harm.

Outbound collections email is off by default. You turn it on when you have read a few drafts and trust the tone.

Auditable and metered

Every AI run is recorded with what it did, when, for which company, and how many credits it consumed. Credits are computed from the real work — tokens, model, whether an image was involved — rather than a flat per-request charge, so reading a drawing costs more than answering a question and the bill reflects reality.

There is a per-user rate limit as well, so a runaway loop or an enthusiastic new user cannot burn a month of credits in an afternoon. Super administrators can see cross-company AI logs including the originating IP.