Agentic AI
An AI agent does not answer questions. It moves the work forward.
That is the whole difference between a chatbot and an agent: an agent acts. It reads from your systems, writes to your systems, chases, opens a task and hands over, in admissions as in payments, academic life or reporting. At GEERD one agent already runs in production, Enro, in admissions. The others we build with you, process by process, wired to what you already have. This page says how, and where an agent stops.
Last updated · September 2026
Definition
An agent is not a better-written chatbot.
Both of them reply. Only one leaves the institution in a better state than before.
| A chatbot | An agent | |
|---|---|---|
| What it knows | A script written six months ago, and whatever the model believes about the world. | What your systems hold on this record, right now: the payment schedule, the timetable, the programme sheet, the missing document. |
| What it does | It replies, then the conversation falls flat. | It replies, writes on the record, opens a task, sends the document, chases, and leaves the team a line. |
| When it does not know | It replies anyway, or sends you to a form. | It stops, hands over, and tells the team what it was missing. |
| What it leaves behind | A history nobody reads back. | An entry in your system, and a trace of every action. |
| Who keeps control | The supplier who wrote the script. | Your team: it reads before sending, then decides, process by process, what may go out on its own. |
Process by process
The posts an agent can hold in your institution.
An agent is defined by a process, a source and a limit, not by a channel, and not by a product. These are the posts we most often open with an institution. The first already runs; the others are written at the diagnostic, before a line of code, and only start if their answers can be checked against a source you hold.
Admissions · in production
Answering a family at 10.40 pm, in their language, with the campaign’s facts; recording interest in a programme; opening the call-back task; leaving the team a line. That is Enro, BrightStep’s agent, in production today.
Payments
Chasing an instalment by name, on the right date, with the amount from the schedule (never a calculation of the agent’s own) and stopping at the first no. Wired to your accounts or your ERP, not to a copy.
Academic life
Timetables, absences, certificates, terms: the same questions, a hundred times, to teams with other work. The perimeter gets written first: what an agent may say about a pupil’s or a student’s record, to whom, and from which source.
Incomplete files
Every year a share of applications and re-enrolments stops on one missing document nobody had time to chase. Chasing the right document, in the family’s language, is work an agent would do without tiring, and must stop at the first refusal.
Reporting and returns
The tables leadership and the ministry expect call for no intelligence: constancy and a single source. An agent fetches them from the same place, at the same hour, every month, and does not touch them up. That is where the checking matters more than the writing.
Graduates
An alumni network dies of silence. Getting back in touch, checking a professional situation, inviting to an event: regular work, low stakes, high volume. That is exactly an agent’s profile, and the easiest to bound.
Your systems
Wired to what you already have.
An agent does not replace your systems: it works inside them. Your student-records software, your accounts, your ERP, your messaging, even your spreadsheets: those are its sources and its targets. Our platforms are among them when you run them; they are not a condition.
Connecting systems that were never designed to talk to each other is part of our work: the executive-programme platform of Africa Business School is connected to the HR system of Groupe OCP, its first industrial client. An agent is wired the same way: through existing interfaces where they exist, through a connector written for you where they do not.
What an agent reads and what it writes is listed before go-live, system by system, field by field. What is not on the list, it does not touch.
In this order
How we build an agent, whatever the process.
Always in this order. The reverse (plugging a model into your data and watching what it makes of it) is what produces the stories you have read.
01Observe the process as it is really done
Not as it is described: as it is done, with the exceptions, the manual chasing, the spreadsheet beside the software. That is where the work an agent can take lives, and the cases where it will have to stop.
02Write the perimeter
What the agent is allowed to do, what it is not allowed to do, where it stops and whom it warns. Written, read by you, and signed before a line of code exists.
03Wire the source
One fact, one source: the amount comes from the schedule, the date from the calendar, the missing document from the file. The agent receives, for one action and that action alone, what that action calls for, not your database, not the internet, not the other records.
04Check every action before it leaves
Nothing goes out as it stands. Do the facts cited exist in the source, does the action stay inside the perimeter, is the reply in the person’s language, has the agent been talked into echoing its instructions. Two of those checks are themselves a model, handed the facts and asked two things only: is this supported by these facts, and is it proper. Those two can retract an action. That is how Enro works today, and the template for every agent we build.
05Start in Assist
For the first weeks the agent prepares and your team sends: nothing goes out without a person having read it. Their corrections say what is missing from the source and what the institution cannot bear to hear in its own voice. Moving to autonomy is a decision, taken perimeter by perimeter, with a spending cap the agent does not cross.
06Leave a trace
For every action: what the agent received, what it did, the checks that ran, the one that blocked, and why. That is what makes an AI’s work something you can argue with in a meeting rather than take on trust.
Built at your institution, not delivered from afar
These six stages happen on site. A GEERD engineer sits with the team concerned, watches the process while it is done, writes the perimeter with them, wires the agent to your systems and stays until it runs in production. In Morocco, and wherever we are called, in French as in English.
The hard lines
What an agent will never do, whatever its post.
These refusals are not preferences you tick. They are written into the policy of every agent we build and verified on every action; Enro enforces them today on every reply. An action that crosses one is not carried out.
- It decides nothing about a person (no admission, no grade, no exclusion, no scholarship, no discount), and never implies a decision is secured.
- It judges no one: it records what was said and done, never what it thinks of someone.
- It does not negotiate and invents no figure: an amount comes from the schedule, a rate from a published grid, a date from the calendar, never from itself.
- It applies no pressure: no manufactured urgency, no asking again for what has already been accepted or already declined.
- It does not leave its perimeter: the payments agent does not talk about admissions, the academic agent does not talk about money. Not visas, not law, not health.
- It does not pass itself off as a person. Asked “who am I talking to?”, it answers that it is an AI assistant working with the team.
- It invents no fact about your institution. What is not in its source does not become a reply: it becomes a hand-over to a person.
- It writes nowhere it has not been allowed to: the list of systems and fields it may touch is closed before go-live.
Perimeter written before the code · checks on every action · action retracted on failure · trace kept
In production
Enro, the agent that already runs.
Enro is BrightStep’s admissions agent. It works on WhatsApp, on the campaigns you give it, in the mode you choose. A family writes in the evening, on a Sunday, the day before a holiday: fees, dates, documents, stages; the reply goes out in the language of the message, with the campaign’s facts and nothing else.
In the same movement it records interest in a programme, leaves the team a factual line, opens a task when it has promised a call-back, sends the brochure. Every draft goes through the checks described above; if one blocks, the reply is retracted and the conversation goes to your team. Every lead gets a score from 0 to 100: in the morning the team opens a sorted list, not an inbox.
It starts in Assist, without exception: it drafts, your team sends. Moving to Autopilot is decided campaign by campaign, and a monthly credit cap stops it rather than letting it carry on with nobody having decided. The demonstration shows it on a real campaign, with the checks beside every draft.
When nothing exists yet
What we do when the agent you need is nowhere.
GEERD is not only the publisher of three platforms. Part of our work is software written for a single institution, because its need had no product: a training platform for an executive programme, a written-examination environment for an engineering school. An agent for a process that is yours alone is built the same way, with the same method.
And there is work we turn down. Anything that puts a decision about a person (ranking applicants, a grade, an exclusion, a scholarship) in a model’s hands. A machine can prepare a decision, gather what is needed to take it, and make a file readable in two minutes. It does not take it, and nobody in your institution should ever be able to say “the system decided”.
Data
What the model sees of your institution.
An agent does not receive “your data”. It receives, for one action and that action alone, what that action calls for: the source of the process in question, the last exchanges of this conversation, and the record of the person it is answering. Not your database, not the other records, not another institution’s.
What goes in and what comes out stays in your organisation and can be read back action by action, including what was blocked. And if your DPO has questions about the processing, we would rather answer them in writing before go-live than after the first incident.
For your DPO
Hosting and the platforms’ guarantees are set out on the Trust and security page; processors, retention periods, transfers and the status of the CNDP filing are in the privacy policy. The data processing agreement is the document we send before a go-live, and the one we hold ourselves to.
If you are not sure you need an agent
Most institutions we meet do not have an AI problem. They have information that does not circulate, and an agent plugged into that would simply go faster with wrong data. That is why we start with a diagnostic rather than a demonstration: first we look at where the information comes from, process by process, then we decide whether a machine has any business there.
On the same subject
Read next.
- What an AI agent may tell a familyThe real question is not the quality of the reply, it is what the machine may commit your school to.→
- Seven questions to ask before letting an AI talk to your applicantsNone is about the model. All are about what is written, what stops, and what can be read back.→
- Grow your admissionsFollow every applicant from first enquiry to enrolment.→
- Track paymentsGive the right teams clear financial information.→
- Simplify academic lifeCourses, attendance, grades, transcripts and academic operations.→
- Run your institutionGive leadership a reliable view without multiplying internal requests.→
- BrightStepAdmissions→
Which process should come first?
A free diagnostic looks first at where your information comes from, process by process, which is what decides where an agent would be any use, and where it would only go faster with wrong data. And if you want to watch one work, the demonstration shows Enro on a real campaign, in Assist, with the checks beside every draft.