How We Made Centre 8’s Marketing Outreach Smarter, with a Targeting Model That Never Touches Admissions

Published 6 July 2026 · Last updated 7 July 2026 · Client engagement · Centre 8 Education

A note on scope, up front: the model described here is used purely for marketing, to help decide who to reach out to about a course. It does not decide, rank, or influence who is admitted to any programme, or which level they are placed in. It runs in shadow mode and does not touch anyone’s queue automatically. Every outreach and enrolment decision is made by a person.

The situation: outreach was a guessing game

Centre 8 Education runs courses, certifications, and seminars, and a real part of the job is reaching the right people about the right course at the right time. The catch was that “the right people” was mostly a guess. The team worked through contact lists broadly, spending the same effort on someone unlikely to be interested as on someone who was an obvious fit, and marketing time spread evenly across everyone is just marketing time spread thin.

There was already a hand-tuned scoring rule trying to help. We replaced it with a proper model. The hard part, it turned out, wasn’t the maths, it was refusing to lie to the model.

The judgement call that shaped everything: what counts as a “no”

The CRM held 515 expired outreach suggestions, contacts the system had once flagged, where the window had lapsed. The obvious move is to treat those as negatives: these people didn’t respond, so label them “not interested” and feed them in. We threw every one of them out instead.

An expired suggestion doesn’t mean “no.” It means an operator ran out of time that week. That’s a scheduling failure, not a signal about the contact, and if you feed it to a model as a negative, you don’t teach it who’s uninterested, you teach it to reproduce your own operational chaos. So the label became ground-truth outcomes only: a small set of genuine positives against a large pool of person-and-course combinations, a roughly one-in-fifty imbalance. Honest and hard to learn from beats plentiful and misleading, every time you have to choose between them.

What we built: a calibrated course-interest model

The model scores existing contacts by how likely they are to be interested in a given course, so the marketing team knows who to prioritise for outreach. It’s a targeting tool, the same category as any lead-scoring system a sales team already uses, applied to education marketing, and pointed only at prioritisation, never at anything that decides access to a course.

We validated it carefully, cross-checking in a way that never lets one course’s data leak into another’s evaluation, and it landed at a genuinely useful level of accuracy, a large lift over contacting people at random. The strongest single signal was how recently a contact had come in: fresh contacts engage, stale ones don’t. Which is precisely what the operator had been saying by hand for months, the model agreed with the human, it didn’t overrule them.

Then the honest part, which matters more than the accuracy. The raw score would read something like “0.92,” and it was tempting to call that a 92% chance. It wasn’t, it was a ranking artefact from balancing the classes. We added a calibration step, and the average predicted probability fell from a flattering 38% to a truthful ~1.8%, which is the real base rate. Same ranking, honest numbers. A score that looks impressive but lies about what it means is worse than no score, because someone will eventually trust it.

What changed: sharper outreach, same human judgement

The marketing team stopped spreading effort evenly and started spending it where interest was actually highest, better response rates, less wasted effort, and no change whatsoever to who is ultimately accepted onto anything.

Crucially, the model runs in shadow mode. It logs its predictions and emails a live accuracy report every Monday, but it does not touch the operator’s queue. If it picked who to contact, it would be grading its own homework, so the decision to act on it stays human, and the switch to ever change that stays a human decision too. It retrains itself monthly and keeps reporting on its own accuracy in the open.

The takeaway

There’s a responsible way to use predictive models in education and an irresponsible one, and the line between them is about what the output is allowed to touch. Pointed at marketing prioritisation, kept in shadow mode, honest about its own confidence, and agreeing with rather than overriding the human, a targeting model saves time and sharpens outreach. Pointed at gating access to courses, it becomes something regulators treat very differently, and rightly so. We built this one to sit firmly on the marketing side of that line, and to stay there by design.

If you want to target your marketing more intelligently while keeping the decisions that matter in human hands, that’s exactly the kind of system worth scoping carefully. Book a free discovery call and we’ll talk through what’s appropriate, and what isn’t, for your business.

Related service: Lead Qualification