Over the past six months, nearly every AI quote I am asked to review contains the word “agent”. Same provider, same architecture as two years ago, same model wired to an exported FAQ, but the budget line has a new name and the amount has sometimes doubled.
AI agents are real, and 2026 is the first year a Moroccan SME can deploy one without an in-house research team. What does not exist is a stable definition of the word on this market. So it now covers both a genuine technical step up and plain repackaging, often at the same price.
The word that took over Moroccan quotes
You can spot the shift by putting two proposals from the same provider side by side. The 2024 one sold an “intelligent chatbot trained on your documents”. The 2026 one sells an “autonomous AI agent that handles requests end to end”. In a good half of the files that reach my desk, nothing in between has changed technically: a language model, a knowledge base, one channel (website or WhatsApp), and a handover to a human as soon as no answer is found.
This is not always bad faith. Vendors renamed their products, every conference panel talks about agentic AI, and a salesperson eventually adopts the vocabulary his prospect reacts to. The real cost lands on you: you are approving a six-figure budget without being able to name what you are buying.
What separates an agent from a conversational assistant
The useful difference is not the model. GPT, Claude, Gemini, Mistral or an open model running on your own servers: an assistant and an agent can sit on exactly the same engine. What changes is what the system is allowed to do once it has understood the request.
A conversational assistant produces text. It reads, summarises, drafts, finds information in your documents, then stops. The action stays human: someone reads the output and decides. That is precisely the scope of an internal AI assistant connected to your documents, and in many companies it already delivers most of the value.
An agent is defined by four properties I check first in any proposal:
- it breaks a goal into steps it picks itself, instead of following a dialogue tree written in advance;
- it calls tools that change the state of your systems: open a ticket, update a CRM record, issue a credit note, send an email;
- it loops: it observes the result of its action, detects failure and tries another route;
- it keeps state between steps, so it can pick up a case it started yesterday.
If the last three properties are missing, you have an assistant. The assistant versus agent question is settled there, not in the product sheet.
Why so many “AI agent” offers are a chatbot with one extra step
Two questions usually settle it. First: list for me, one by one, the actions the system performs without human approval. If the answer amounts to “it replies, and it transfers to an advisor”, there is no autonomy, there is an escalation. Second: what happens when the call to your ERP API fails? An agent retries, changes strategy, or raises an alert with usable context. A rebranded chatbot sends an apology message.
There is also a perfectly honest middle category that will nonetheless be sold to you as agentic: the deterministic workflow with a model in one box. A PDF arrives, the model extracts six fields, a script writes them into the accounting software. That is robust, usually cheaper and far easier to audit than an agent. It simply is not the same thing, and the price should show it.
Where autonomy changes the outcome, and where it adds nothing
Autonomy earns its place when the path varies case by case and the volume makes supervision expensive. Reconciling CMI settlements with online store orders when the payment labels never quite match. Processing purchase orders that arrive by email in eight different layouts, with references sometimes missing and buried in the customer history. Chasing unpaid invoices by reading the payment schedule, the last promise to pay and any open dispute before writing a word.
On the other side, autonomy adds nothing to a process that is already deterministic: if the rules fit on two pages and the inputs are structured, a classic integration costs less and breaks less often. It adds nothing at twenty cases a month, where supervision time exceeds the time saved. And it turns risky when the input data is dirty, because an agent acting on a duplicated customer master file spreads the error faster than anyone notices. Before committing, an honest AI readiness check keeps you from funding the automation of a mess.
What AI agents Morocco buyers actually pay in 2026
The ranges below are orders of magnitude seen on the Moroccan market this year. They move with the state of your systems, not with the fashion in vocabulary.
- FAQ chatbot on a website or WhatsApp: MAD 25,000 to 70,000, two to four weeks.
- Internal assistant over your documents, with no write access: MAD 80,000 to 250,000, six to twelve weeks.
- Agent with write access to two or three systems: MAD 250,000 to 600,000, three to six months, and rarely less if your APIs barely exist.
Add the running cost: model usage, hosting, supervision and fixes, in practice MAD 3,000 to 20,000 per month depending on volume. Almost none of the cost comes from the model itself. It comes from everything built around it: connectors to software with no documented API, service accounts with narrow rights, an audit log a finance controller can actually read, a test suite that replays fifty real cases before each release, and a recovery procedure for when the agent gets it wrong. A quote for an agent without those lines is a quote for an assistant, and the cost breakdown reads like any other AI project budget.
CNDP: the question Law 09-08 has not settled yet
Law 09-08 protects individuals in the processing of their personal data, and it was written long before systems started chaining their own actions. Several practical consequences follow for an agent project.
Hosting first. If the agent sends personal data to a model operated outside Morocco, you are in cross-border transfer territory, with the corresponding formalities at the CNDP. That is one of the serious arguments for a locally hosted model, even a slightly weaker one. Then purpose: a processing activity declared for customer relationship management does not automatically cover a system that takes the initiative on debt collection actions. Finally the traces, which almost everyone forgets. An agent logs its prompts, those prompts contain personal data, and those logs are themselves a processing activity with a retention period to define.
On fully automated decisions that affect a person, do not rely on a vendor’s assertion: have your specific case qualified and check the applicable formalities on the CNDP website. The workplace AI questions already arise with ChatGPT and Copilot, and an agent writing into your systems only widens the scope.
What goes wrong when an agent acts alone, and what to demand
Real incidents are unspectacular. They are duplicates: the agent retries a step after a timeout and opens two tickets, or sends the same reminder twice to an already irritated customer. They are identifier errors, where a reference read off a crooked scan points to the neighbouring order and the status changes on the wrong file. Sometimes they are expensive loops, the agent retrying thirty times a call that will never succeed.
The safeguards exist and belong in the contract: idempotent writes, a cap on actions and spend per run, write access limited to named fields, mandatory human approval above a threshold (an amount, a strategic account, a cancellation), and a log tying every action to the decision that triggered it. Three questions tell you whether any of this was thought through:
- Which actions exactly, and on which system? Ask for the named list of tools the agent can call, the service account used and its rights. A vague answer is itself an answer.
- What happens on failure? Have them describe the behaviour when an API is down, when data is missing, and when the action was already performed, then ask to see the rollback mechanism.
- Where do the model and the data run, and who handles the CNDP formalities?
Ask for a demo on your own data rather than on the in-house sample set. A genuine agent will sometimes fail in front of you and explain why. A repackaged chatbot always succeeds, on the handful of cases prepared for it.
Frequently asked questions
Can an existing WhatsApp chatbot become an AI agent?
Partly. The channel and the conversation history stay usable, but the bulk of the work sits elsewhere: connectors to your systems, permissions, testing. Budget for a rebuild rather than an upgrade, and remember that the WhatsApp Business API has its own rules on outbound messages.
Do we need an agent if the company already has ChatGPT Team?
Those are two different uses. A subscription equips individual employees, an agent automates a process that runs with nobody watching the screen. Many SMEs gain faster by first organising their subscriptions, as in this comparison of ChatGPT Plus, Team and custom builds.
Can an AI agent be hosted entirely in Morocco?
Yes, using open models on local infrastructure or a sovereign cloud. You give up some reasoning quality against the best American models and you gain on data location. For an agent handling sensitive data, the trade-off often tilts local.
How long before an agent pays off?
It depends on the volume of work it genuinely absorbs, not on the power of the model. A process running 200 cases a week usually pays back in under a year. At twenty cases, the arithmetic almost never works, however good the engineering.
Before you approve the “AI agent” line
Arrowlancer builds AI solutions for Moroccan companies: internal assistants over your documents, automation that genuinely acts on your systems, document analysis, and deployments hosted in Morocco when the data calls for it. The scope is set out on our services page. If an “AI agent” quote is sitting on your desk and you are unsure what it actually contains, get in touch: an hour of technical review costs less than a pilot abandoned after six months.