The three things people mean by “AI”
| Chatbot | Automation | AI agent | |
|---|---|---|---|
| What it does | Answers questions | Runs a fixed sequence | Decides and acts toward a goal |
| Who decides the steps | Not applicable | You, at build time | The agent, at run time |
| Can it use your systems | Usually not | Yes, along a fixed path | Yes, choosing which to use |
| Handles the unexpected | No | No — it stops or errors | Often, within its guardrails |
| Cost to build | Low | Moderate | Highest |
| Right when | Questions repeat | The process never varies | The right step varies |
A concrete example
Take an enquiry arriving at a construction supplier in Kigali.
- A chatbot answers “what are your opening hours?” and stops there.
- An automation takes every enquiry, writes it to a spreadsheet, and sends the same acknowledgement email. Reliable, and completely blind to what the enquiry said.
- An agent reads the enquiry, works out that it is a bulk cement request from a contractor rather than a retail customer, checks current stock, routes it to the trade sales person instead of the general inbox, drafts a reply quoting trade pricing, and flags it for a human to approve before sending.
The agent version is more useful and considerably more work to build correctly. That is the whole trade-off.
What an agent needs to actually work
This is where most agent projects fail, and it has little to do with the model:
- Real access to real systems. An agent that cannot read your stock levels or write to your CRM is a demo. Wiring it into the tools you actually run is the bulk of the work.
- Clean, reachable data. If your customer records live in three places and disagree, the agent will make confident decisions on bad information.
- Guardrails. Clear limits on what it may do without a human. Anything touching money, contracts or customer-facing promises should be approved by a person until it has earned trust.
- Visible failure. When it cannot complete a task it must say so loudly. An agent that fails silently is worse than no agent, because you stop checking.
- A way to measure it. If you cannot tell whether it is doing better than the manual process, you cannot justify keeping it.
Where agents genuinely help
- Enquiry triage — sorting incoming messages by intent, urgency and value, and routing each to the right person.
- Research and summarising — pulling together background on a supplier, a tender or a prospect.
- Follow-up judgement — deciding who is worth chasing again, when, and with what message.
- Reporting with commentary — not just assembling the weekly numbers but noting what changed and why it might matter.
- First-pass document handling — reading invoices or delivery notes and extracting the fields that matter.
Where they do not
Being straight about this saves money. An agent is the wrong tool when the process never varies — that is what automation is for, and it will be cheaper and more reliable. It is also the wrong tool when the cost of a mistake is very high and there is no human check, when the underlying data is too messy to trust, or when the real problem is that nobody has agreed what the process should be. Automating an undefined process just produces confusion faster.
How to start without over-committing
- Automate the fixed parts first. They are cheaper, and they clean up the data an agent would later depend on.
- Pick one decision that genuinely varies and is currently made by a person spending real time on it.
- Keep a human in the loop at first. Let the agent draft and a person approve. You will learn quickly where its judgement is weak.
- Measure against the manual baseline — time saved, response time, missed follow-ups.
- Widen its autonomy only where it has earned it.
If you are weighing up what this costs, our guide on AI automation costs in Rwanda covers how this kind of work is priced.