When an agent is the wrong tool

An implementation company that publishes a reference site has an obvious incentive to say yes to every agent request. This article is the list of times we say no, and what we suggest instead. Each situation comes from a real pattern in scoping conversations; the examples are composites, not clients.

The frame

An agent is a probabilistic decision-maker — a model choosing actions — wrapped in deterministic controls. It is the right tool when the inputs cannot be structured in advance and the job requires an action. Move either condition and something simpler wins. The diagram above is the whole framework: bottom-left is a menu, top-left is automation, bottom-right is a chat with retrieval, and only the top-right is an agent.

Seven situations

1. The inputs fit a menu

“Customers ask for their balance, our opening hours and how to pay.” Every one of those is a menu item. A WhatsApp flow with reply buttons answers them deterministically, at zero per-message cost inside the service window, with no hallucination and no golden set. Zimbabweans are fluent in menus — USSD trained a nation — and a good menu is faster than typing. Build the menu; add an agent later, behind the one branch that needs free text.

2. The action is fixed and the trigger is structured

“When a Paynow payment succeeds, send the receipt and update the CRM.” That is a workflow: trigger → condition → action, with structured data at every step. n8n, Zapier, Make or the CRM’s own automation does it for the price of a subscription and produces a log that is the audit trail. An agent here pays tokens to guess what a rule already knows and adds a probabilistic step to a deterministic job.

3. The decision has significant legal effect on a person

Credit approvals, account suspensions, hiring screens, insurance claims. Zimbabwe’s data protection regime requires consent or legal authorisation for automated decisions affecting individuals’ rights. You can build an agent that prepares the decision — gathers documents, checks completeness, drafts a recommendation with evidence — but a person decides. If the client’s goal is to remove the person, the agent is the wrong tool and the law is the reason.

4. The source documents are not ready

An internal Q&A agent mirrors your document hygiene. Two versions of the leave policy, a price list in someone’s inbox, a procedure that lives in a WhatsApp group: the agent will find the wrong one and quote it confidently. Six weeks of editorial work — one owner per document, version dates, a superseded folder — is the project. The agent comes after, and is then a small job.

5. Your customers are not where the agent would live

An agent on WhatsApp assumes the customer has a smartphone, data and the app. Internet penetration reached 84.55% in Q4 2025 on POTRAZ’s count, but MISA’s August 2026 analysis notes that many Zimbabweans spend more than 10% of monthly income to stay online, against a UN affordability benchmark of 2%, with typical monthly spend around US$10. For a rural customer base, a low-income segment or an older one, a voice line or a human on a phone reaches more people than an agent on WhatsApp — and voice has its own problem, next.

6. The economics of the channel do not work

A voice agent through an international carrier pays about US$0.86 per minute to a Zimbabwean mobile before the model says a word. For reminders, that is a US$1.30 call doing the job of a US$0.004 utility template. Unless the client can terminate on a local trunk or genuinely needs voice (no smartphones, after-hours reception), the agent is the wrong tool for the channel, not the job. Move the job to WhatsApp or keep the person.

7. Nobody can write the policy down

“Just make it handle whatever comes up.” An agent needs a scope, a refusal list and an escalation rule that can be tested. If the client cannot write five sentences describing what the agent must never do, there is no golden set, no evaluation and no way to know whether a prompt change helped. The right next step is a two-hour workshop that produces the policy — then, and only then, an agent.

A table to keep

You are asked for an agent because…	Build instead	Why
Customers ask the same five questions	WhatsApp menu / rule-based flow	Deterministic, free in-window, no hallucination
Data must move between systems on an event	Workflow automation	Structured in, structured out, log is the audit trail
Someone wants to remove a human from a consequential decision	Assist-only agent + human decision	Automated-decision rule; liability
Staff cannot find the right document	Document clean-up, then retrieval	Agent quality equals document quality
Customers are offline, rural, or price-sensitive on data	Voice line, USSD, human	Reach beats sophistication
The channel is international voice	WhatsApp agent or local telephony	Line cost dominates the model cost 10:1
The scope cannot be written down	Policy workshop	No policy, no evaluation

When we do say yes

The inputs are messy and human — a voice note from a mine supervisor, a sentence from a patient, a supplier’s PDF — and the outcome is an action in a system: a filed report, a held slot, a proposed reconciliation. The tools are narrow, every write has a confirmation gate, and the client can say what “wrong” looks like. Those are the top-right cases; the agent types pages describe nine of them, and the comparison database prices them.

The honest version of “we build AI agents” is “we build the smallest thing that does the job, and sometimes that is an agent”. This article is the other half of that sentence.