AI agents for business: what they can actually do.

First, agent or workflow?

The word “agent” is used for everything from a chatbot to software that runs a company. A useful line was drawn by Anthropic in Building effective AI agents (December 2024): workflows are “systems where LLMs and tools are orchestrated through predefined code paths”, while agents are “systems where LLMs dynamically direct their own processes and tool usage”.

In practice, most of what earns its place in a business is a workflow with a model inside, and a true agent only where the steps cannot be predicted. The same article advises finding “the simplest solution possible, and only increasing complexity when needed” — which matches what we see in production.

Six jobs agents do today in our projects

Each of these runs in a real product or client platform. What they have in common is at the end.

1. Take a restaurant booking on WhatsApp. For a restaurant group, a WhatsApp assistant books tables and manages a waitlist in the same system the hosts use. What makes it work: the agent can only act through booking tools we wrote, so it cannot invent a table that does not exist. (AI for restaurants)
2. Answer a voice note, by voice. For gyms, members send voice notes — including in Kiswahili — and the agent answers by voice, books, alerts the owner and hands over to staff. What makes it work: a dry-run mode for testing, a daily usage cap and an off switch.
3. Turn a PDF menu into a structured menu. In Vimus, a restaurant uploads a PDF, a photo or a link and gets dishes, sections and prices, ready for video. What makes it work: the owner reviews everything before it is published — extraction is a draft, not a decision.
4. Draft a reply a professional approves. In BrokerPilot, incoming client messages are classified and a reply is drafted; the broker edits and sends. What makes it work: the model never sends on its own.
5. Answer at 3 a.m., and know when to stop. Mama Bot answers new mothers' questions on WhatsApp and, on any warning sign, tells them to call the emergency number for their country. What makes it work: a short list of red flags written by people, not left to the model's judgement.
6. Run a team's operations. In Takat OS, our own assistants turn chat and WhatsApp messages into tasks, decisions and follow-ups in the right project. What makes it work: the assistant's reply is rewritten from what the system actually recorded.

What it takes to run one

A demo agent needs a prompt. A production agent needs everything around it:

• Tools, not free rein. The agent acts through functions you wrote and tested — create a booking, update a record — each checking its inputs.
• Guards. A usage cap per day, a dry-run mode, a switch to turn it off, and rules for when a person must decide.
• A human handover that works at any moment, not only when the script runs out.
• Logs of every action, so you can answer “what did the agent do yesterday?” with facts.
• Your data in your database, with backups that have been restored at least once.

Where agents still fail

They report things that did not happen. A model will happily say “done” after a failed call. The rule we apply, first of all to our own assistants: an agent may only report what a system confirmed, and its reply is checked against the real results before anyone reads it.

Small errors compound. The Anthropic article warns that agents mean “higher costs, and the potential for compounding errors”. Short chains of well-tested steps beat long autonomous runs for most business tasks.

The real world is not the prompt. Our first meeting notetaker was built well and still missed meetings: it waited in lobbies nobody was watching. The fix was not a better model but a message to the organiser the moment it waited. Most agent failures we see are of this kind — a missing step in the process, not a weak model.

How to start

Pick one conversation that costs your team time — bookings, renewals, quotes, questions after hours. Decide what the agent may do alone, what it prepares and what it never touches. See it working on a preview before any customer does. Then measure, and only then pick the next one. That is how we build our AI Pilots, and it is why they go live in two to six weeks.