The work that does not need a person.

Lead qualification, support triage, reconciliation, scheduling. When an AI agent is the right answer, how the economics work, and where agents fail in ways that humans do not.

There is a category of work in every business that you keep meaning to systematise. Qualifying inbound leads. Answering the same five client questions. Reconciling two systems that should agree but never quite do. Scheduling. Following up. Triaging support tickets. The work is straightforward. It happens at a rate that justifies a person. The problem is that you cannot find a person who will do it for long.

This is the natural habitat of an AI agent. An agent is a small program with three things: a trigger, a set of tools it can use, and a loop that decides which tool to call next. Tools are usually APIs your business already has — your CRM, your inbox, your calendar, your billing system. The loop is provided by the model.

The most common agent we build is a lead qualifier. The trigger is a new contact submission. The agent reads the message, looks up the company on the public web, checks if the company is in your target segment, drafts a personalised response, and either books a meeting on the calendar or routes the lead to a human for the borderline cases. End to end, under two minutes. Cost: less than a rand per lead. The number of qualified meetings goes up. The time the sales team spends qualifying goes down.

The second is a support triage agent. Incoming tickets get read, categorised, and routed. The simple ones get answered directly using your documentation. The complex ones get tagged and queued for a human with the most likely answer pre-drafted. Customers get faster responses. Agents handle the work that needs them.

The third is a reconciliation agent. Two systems hold overlapping data — your billing system and your CRM, your inventory system and your fulfilment system. They drift. The agent runs nightly, compares the two, flags exceptions for a human, and quietly fixes the trivial ones. The number of meetings titled “data quality” drops.

The economics are unusual. An agent costs the same to run on a Saturday at 3am as on a Tuesday at 11am. It does not take leave. It does not need an onboarding period. It does not resign. The capital expenditure is the build (typically two to six weeks); the operating cost is fractions of a rand per task.

Agents fail in ways that humans do not. The model makes a confident error on a subtle case. A tool returns unexpected data and the agent misroutes a ticket. These failures are real and have to be designed around. The two answers are evals and human-in-the-loop. Evals catch the model in a test environment before deployment. Human-in-the-loop catches it in production by routing low-confidence cases to a person.

The wrong reason to build an agent is to replace a team. The right reason is to redirect a team. The agent handles the volume. The team handles the cases the agent cannot. The team also keeps the eval harness sharp, because they are the ones who notice when the agent starts drifting.

Three signs your business is ready for an agent: there is a workflow that runs more than ten times a day, it is mostly the same shape each time, and the cost of getting it wrong on a single case is bearable. If all three are true, this is where to start.

→ Find out which workflow in your business should not need a person. Book a Discovery.