Case Study · Industrial Distribution

Our first workflow: lead generation at MP Cutting Tools.

MP Cutting Tools is an industrial cutting tool distributor in Montréal, run by Alessandro Mattei. MP finds tools, parts, and equipment for industrial companies, which means its week runs on research, outreach, and quotes. Exactly the kind of work an AI workflow should handle. Meridian built one, broke it, rebuilt it, and it now runs inside MP's tools every morning, with Alessandro reviewing what goes out.

Full disclosure up front: MP was our first client and the business kept running while the workflow learned. Everything below, including what broke, is written from the build log.

The starting line

Before the agent, lead generation at MP was entirely manual:

  • Research industrial companies one at a time — sector, size, what they likely buy
  • Dig for the right contact — purchasing manager, maintenance lead, owner
  • Write every outreach email from scratch
  • Track follow-ups by hand

It's not hard work. It's relentless work — the kind that gets skipped the moment a live deal needs attention, which is exactly when the pipeline quietly dies.

What we built

One agent, plugged into the tools MP already used. It handles the pipeline end to end:

01

Research

Reads a target list and builds a short profile per company: what they do, what they likely source, why MP is relevant to them.

02

Contacts

Finds the right person and role. Flags low-confidence matches for a human instead of guessing.

03

Drafting

Writes outreach in MP's actual voice — trained on real sent emails, referencing what makes each target relevant.

04

The human gate

Nothing sends itself. Alessandro reviews every draft: one click to send, edit, or kill. The workflow does the hours of digging; the judgment call stays human.

What changed
Research, done
Company profiles and contacts are waiting before the day starts, instead of being the day.
Drafts, not blanks
Every outreach starts from a draft in MP's voice. The human edits and sends, or kills it.
Weeks, not months
From first sketch to the full pipeline running, tracked in the build log.

We keep the measurements in the build log rather than on this page. A multiplier with no baseline behind it is just marketing, and this page exists to be checked, not to impress. Ask us and we'll walk you through the log.

What broke

The first version was worse than doing it by hand. Two failures were worth the scars.

The drafts sounded like a template. Our first instinct was to write the agent more rules — tone guidelines, do's and don'ts, a style guide. It got more generic with every rule. What fixed the voice was the opposite move: fewer instructions, more real sent emails in its context. The agent learned MP's voice by reading MP's actual mail, not by being told about it. Show, don't instruct.

It guessed at contacts it shouldn't have. Early on it would confidently name a "purchasing manager" it had half-inferred, and a wrong name in a cold email is worse than no email. The fix was a confidence rule: below a threshold, the agent stops and flags the contact for a human instead of guessing. Being willing to say "I'm not sure" turned out to be the feature that made the rest trustworthy.

Some things we left alone on purpose. The agent doesn't decide who's worth pursuing, and it doesn't send. Those stayed human because they're judgment, not digging — and judgment was never the bottleneck.

What this proved

The build for MP is the same build we now sell: workflow audit → business knowledge written down → AI workflow in your existing tools → human review on everything that leaves the building. We didn't design that process in a pitch deck. We designed it on a real pipeline, with a client's revenue on the line.

Your week probably has a version of this.

Research, outreach, quoting — the pipeline work nobody has time for. The first conversation is free, and we'll tell you honestly if AI isn't the right answer for it.

Send a message → See how we'd scope it
Free, reply within a business day. If your workflow doesn't need AI, we'll say so.