Chatbot, RPA, or agents: which AI fits your work
Three vendors pitch you in the same month. One sells a chatbot trained on your documents. One sells robotic process automation that clicks buttons faster than your admin team. One sells something called agents. All 3 say AI. All 3 quote wildly different prices. None of them explains which one your work actually needs.
The word AI now covers at least 4 different tools. They are as different as a calculator, a phone, and an employee. Matching the tool to the shape of your task removes most of the risk before a single dirham is spent.
Deterministic automation and RPA
Scripts and screen robots that repeat exact steps: open this system, copy these fields, paste them there. Superb when the task never varies. Brittle the moment it does: a renamed column or a reworded email breaks the run, silently. RPA suits high-volume, zero-variation clerical steps, invoice entry, data migration between systems, and very little of the document work professional firms actually sell.
Chat assistants
ChatGPT and its siblings. Genuinely useful, and the honest assessment is that your team should be using one for drafting fragments, summarising, and unblocking. But an assistant answers one question at a time, remembers nothing about your workflow, and leaves a person doing the work: gathering the inputs, prompting, checking, assembling, formatting. It makes your people somewhat faster. It does not remove the production hours; it reshuffles them.
Agent systems
Software workers assigned to roles, wired into a workflow: one gathers sources, one drafts to your template, one checks numbers and citations, one routes the result to a person for sign-off. The difference from a chat assistant is that nobody sits in the middle: the brief goes in, the checked draft comes out. The difference from RPA is that variation does not break it: a strangely worded brief is read, not pattern-matched. The cost of those properties is engineering: an agent system must be built around your specific workflow, integrated with your tools, and given a review gate. It is a system, not a subscription.
Mixed architectures
Real deployments are rarely pure. The strong pattern is agents for the fluid middle, deterministic steps for the parts that must never vary, rate lookups, template filling, sending, and a human at the end. Distrust any pitch that is 100% one ingredient.
| RPA | CHAT ASSISTANT | AGENT SYSTEM | |
|---|---|---|---|
| ZERO-VARIATION CLERICAL | |||
| ONE-OFF QUESTIONS | |||
| MULTI-STEP DOCUMENT WORK | |||
| PHONE-NATIVE WORK |
The questions that expose a mismatch
Ask any vendor 3 things. Does the task vary run to run? If yes, RPA will break; you need something that reads. Who does the work between input and output? If the answer is your staff plus their tool, you are buying an assistant, not capacity. What happens when the system is unsure? If the answer is not a specific mechanism, flags, a review queue, a person who signs off, the pitch is a demo, not an operation.
Price is a clue, not a verdict
Chat assistants cost tens of dirhams per seat per month because they leave the work with you. RPA licences run mid 5 figures a year and pay off only at clerical volume. Agent systems are built, so they price like a small construction project: a fixed build, then an operating arrangement. Each price is fair for what it is. The waste is paying any of them to do a job shaped for one of the others.
DoxaMind builds the third kind: systems of AI agents for judgment-heavy document work, research, quotes, proposals, reporting, mixed with deterministic steps where rigidity is a feature, and always ending at your team's sign-off. If you are weighing the 3 pitches on your desk, join the waitlist and we will tell you honestly which shape your work is.