AI agents
Agents that read, decide and act inside your tools, with a person on the approval gate.
For work that needs judgement, not a fixed script: reading a supplier document, matching a payment to a quote, drafting a reply from three systems. We build the agent, the tools it may use, the limits it may not cross, and the log that proves what it did.
The Zimbabwean situation
What usually goes wrong
Most of what Zimbabwean businesses ask for is a flow, and we say so. The agent cases are the ones where a person currently reads something, checks two systems, and decides: a supplier invoice that needs matching to a purchase order, a WhatsApp voice note describing a repair, a customer asking whether a part will fit their vehicle, a payment notification that must be tied to one of forty open quotes.
The risks are specific. An agent that quotes from memory will invent a price in USD when your list is in ZiG. An agent with write access to the CRM and no gate will "helpfully" close deals. Neither is acceptable, and both are design problems, not model problems.
What we build
What gets built
- One job, defined. Inputs, allowed tools, required output, and the rules it must obey ("credit only for account customers", "never quote outside the price list").
- Tools, not access. The agent gets narrow functions (find customer, read stock line, draft quote), never a database login.
- Approval gate. Proposed actions that touch money, credit or a customer promise go to a named person on WhatsApp or email as approve/reject.
- Audit log. Every action stored with the evidence it used, so you can answer "why did it do that?".
- Evaluation set. Your past examples become the test; the agent goes live when it passes them, and is re-tested monthly.
How it works
The flow, end to end
What it needs from you
- One job, described in a paragraph, with what "done" looks like.
- Read access to the systems involved (API keys, or exports if there is no API).
- The rules, written down, including what must never happen.
- 20–50 past examples with the correct outcome.
- A named approver with a WhatsApp number and a backup.
Timeline shape
| Phase | Typical span | What happens |
|---|---|---|
| Scoping | 2 weeks | Pick one job with clear inputs and a measurable output. Write the rules the agent must obey. Collect 20–50 past examples. |
| Build | 3–6 weeks | Tools (read/write connectors), prompts, approval gate, audit log, evaluation harness run against the examples. |
| Supervised pilot | 2–4 weeks | Every action reviewed by a named person; accuracy tracked weekly; scope widened only when it holds. |
| Run | ongoing | Monitoring, drift checks, monthly re-evaluation, new tools added one at a time. |
Spans are typical shapes, not quotes. The process page explains what moves them.
What it costs
What drives the price
Agents cost more than flows because of the evaluation and the gate, not because of the model. The cheapest agent reads one system and writes to none; the most expensive reads several, writes to two, and must be right every time.
We do not publish package prices because the drivers below move the number by multiples. The pricing page lists every driver with sourced platform costs.
| Driver | Effect on cost |
|---|---|
| Number of systems the agent reads or writes | Each connector is built and tested; systems without an API cost the most. |
| Approval complexity | One approver on WhatsApp is simple; multi-step approvals with limits per role are a bigger build. |
| Evaluation depth | More past examples and stricter accuracy targets mean more test cycles before go-live. |
| Model usage | Charged per token by the provider; long documents and high volume move this. Estimated at scoping from your volumes. |
| Hosting and data location | Standard cloud is cheapest; private or regional hosting for sensitive data costs more. |
Questions we get asked
What is the difference between an agent and the WhatsApp automation?
A flow follows a fixed path you designed. An agent is given tools (read the CRM, read the stock sheet, draft a quote) and a goal, and decides the steps itself within limits you set. Use a flow when the path is known; use an agent when the work needs reading and judgement. aiagent.co.zw explains the categories in depth.
Will it invent prices or promise things we do not offer?
Not if it is built properly. Quotes may only be assembled from your price source; if the item is not there, the agent asks a person. Every proposed action is logged with the evidence it used, and anything touching money or credit waits for approval.
Which AI models do you use?
Whichever fits the task and the data rules. Hosted frontier models for language, smaller or private models where data must stay under your control. The choice is made at scoping and written down.
What happens when it gets something wrong?
The approval gate catches the expensive mistakes before they happen; the audit log shows what it did and why; a rollback path exists for every write it can make. During the pilot a person reviews every action.
Can the agent work when our systems are offline?
It queues. If your accounting system or CRM is unreachable, actions are stored and retried, and a person is told if the backlog grows.
Next step
Scope it before you book it
Fill in the scoping worksheet (five minutes, no sign-up). It produces a summary you can paste into WhatsApp or email, so the first call starts with the actual work.