AI agents by industry
Same loop, different tools and different stakes. Four patterns below are chosen because the inputs are unstructured (voice notes, photos, mixed-language text) and the action is real. Each is a starting design, not a product; the cost bands are illustrative run costs from the price basis.
Mining: shift and incident reporting by WhatsApp voice note
Problem. Supervisors at dispersed sites report by phone call or paper; the data reaches the office late and incomplete. Agent. Supervisors send a voice note ("night shift, 42 tonnes, pump 3 down, one near-miss at the ramp"). The agent transcribes, fills the shift form with structured fields, asks one clarifying question if a required field is missing, files the report and raises an alert for any incident keyword. Tools: transcribe, form fill, ERP write (staging), alert. Autonomy: assist — the supervisor confirms the form. Risk: medium — safety-critical context, so the agent never gives safety instructions and never controls equipment; SCADA access, if any, is read-only. Cost: audio tokens make this dearer than text — US$50–350/month at 6,000 messages. Governance for mining operators (security, OT networks, data classification) is covered from the enterprise angle at enterpriseai.co.zw.
Agriculture: agro-dealer ordering and stock enquiries
Problem. Smallholders and commercial farms ask an agro-dealer for prices and availability of seed, fertiliser and chemicals in seasonal bursts, often in Shona or Ndebele, often by voice note. Agent. Reads the request, answers from the price list tool (with date), checks stock at the nearest branch, takes an order with a Paynow EcoCash prompt for deposit, and sends a pick-up reference. Tools: catalogue, stock by branch, order create (gated), Paynow, template. Autonomy: act with confirmation. Risk: medium — the agent gives no agronomic advice ("how much should I apply?") and hands those to an agronomist; product prices and stock come only from tools. Cost: comparable to a sales agent, US$20–80/month at a few hundred orders, plus utility templates for pick-up notices.
Hospitality: lodge and guesthouse bookings with deposits
Problem. Enquiries arrive on WhatsApp at all hours; answering slowly loses the booking to the next lodge. Agent. Confirms dates, party size and room type; quotes from the rate tool in the currency the guest asks for (USD or ZiG at the rate the tool returns); holds the room; requests a deposit via a Paynow link or EcoCash prompt; confirms with a PDF and directions on payment. Tools: availability, rates, hold/confirm booking, Paynow, template, handoff. Autonomy: act with confirmation. Risk: low-to-medium — money is involved but through a payment provider, and holds expire automatically. Cost: booking-agent band, US$25–60/month at 1,000 enquiries; payment-provider fees are separate.
Logistics: dispatch updates and proof of delivery
Problem. Customers ask "where is my delivery?", drivers send proof-of-delivery photos to a group chat, and the office reconciles by hand. Agent. Answers ETA questions from the fleet-tracking tool; accepts a POD photo from the driver, extracts consignment number and signature presence, files it against the delivery, and notifies the customer with a utility template. Tools: tracking read, consignment lookup, POD file (write), template. Autonomy: act for notifications, assist for exceptions. Risk: medium — customer data and location data; the agent must verify the asking number belongs to the consignment before answering. Cost: images cost more than text; US$50–250/month at a few thousand consignments.
The common thread
In all four, the model reads unstructured input from a phone and turns it into a structured action in an existing system. None of them needs a new app for the worker or the customer — WhatsApp is already there, which is why the channel constraints page is required reading before any of these is scoped.