Case Study — Catching AI Drift Before Your Customers Do

Client: ValorHub (self-owned)

What it protects: ~35 social media posts per campaign across LinkedIn, Twitter, Telegram, and Instagram.

Timeline: Built in weeks 3–5, after the first drift incident.

The pattern:

Content safety net AI-generated posts flow through three sequential checks: hard rules, gold-standard comparison, and escalation threshold. Good posts publish. Drift triggers an alert. Hard fails are rejected. AI drafts 35/campaign Hard rules banned words format · length Gold check score vs baseline Threshold PASS / WARN / FAIL Published on-brand Rejected hard fail Alert drift detected On-brand post Drift — flagged for review Hard fail — blocked automatically
Every AI-drafted post passes through three checks in order. Green flows through to publish. Gold routes to an alert. Red never ships.

The eight-week problem

You paid someone to set up an AI to help write your marketing posts. Two weeks in, it sounds like you. Comments are positive. Followers grow. You breathe out.

Eight weeks in, a customer emails: "Did you rebrand? The last few posts don't sound like you anymore."

You go read the last month. They're right. The AI drifted. It uses exclamation marks again (you told it not to). It's stopped mentioning your product by name. Every third post opens with a variant of "In today's fast-moving landscape…"

You changed nothing. This is what every AI writing setup does when nobody scores what comes out — it slowly pulls toward its own average. Nobody catches it because nobody re-reads last week against last month. And by the time a customer catches it, you have a month of off-brand content already in front of the audience you wanted to impress.

The fix is not "write a better prompt." The fix is a safety net that scores every post before it ships.

The safety net (three checks, in order)

1. Deal-breakers. Banned words. Wrong length. Wrong sign-off. Founder's name misspelled. A computer either sees the pattern or it doesn't — instant fail. The post is blocked; I get told what broke.

2. "Does this sound like us?" For each post type (LinkedIn, Instagram) I keep 5–6 past posts I personally approved as "this is what we sound like when we're good" — the gold library. Every new post gets scored against it. If today's post scores clearly worse, the safety net flags drift even when no rule was violated.

3. Severity. Small dip: a note in the log. Big dip: full stop. Post doesn't ship until I look at it.

What the net actually caught

"Maksym" instead of "Max." My legal name is Maksym. I go by Max publicly. The AI kept slipping "Maksym" into posts because that's how my LinkedIn profile is listed. The deal-breaker check caught it three times in the first month. Without the check, at least one post would have gone out signed with the wrong name.

Voice drift with no rule broken. A post that would've scored 8.6/10 against the gold library came in at 8.0. No rule violated, nothing looked wrong on read-through. The gold-comparison check flagged it. I read both versions side by side, saw the drift, updated the prompt. The next post came back to 8.6. Without the net, that post would have shipped, and the next twenty would have drifted further from there.

"Unlockable" flagged as if it were "unlock." Early version of the deal-breaker check used a naive substring match — "unlock" was on the banned-word list, and it started failing legitimate posts that contained "unlockable." Tightened the check to whole-word matching. This is the class of bug you specifically want the safety net to not have — better to catch and fix its own bugs than to let them silently trash your content.

What working with me looks like

Drift audit — €500 refundable (~3 business days). For a team asking "has our AI actually been drifting, and how bad?" Send me your last 30 days of AI-assisted posts. I score them against a rubric you get to keep and send a 2-page written report: what has drifted, ranked by severity, and three things you can do this week. If drift is inside acceptable range, I say so and refund the €500.

Build — €5,500–€12,000 (3–5 weeks). One content channel wired end-to-end. Safety net running against your live pipeline, a "sounds like us" library curated with you, a runbook so anyone on your team can update the rules. One off-brand post seen by 5,000 followers dents trust it takes months of good posts to rebuild. On a 4-post-per-week pipeline that's already drifting, the build pays back inside 3–4 months.

Contact: info@valorhub.eu · linkedin.com/in/maksymdonets · Munich (CET)