The default answer to a pipeline problem is "hire more SDRs".

It's usually the most expensive answer in the room.

The other option is to stop spraying and start prioritising. Work out which accounts have a reason to buy this week, and reach those. That takes a system, and somebody has to build it.

So I built it, then I sold with it.

Every play runs the same loop:

signal → enrich → score → route → act → measure → kill or freeze

The discipline is those last two steps. Most teams build signals and never close the measurement loop, so nothing gets killed and nothing gets templated. You end up with a graveyard of clever ideas nobody can repeat.

£10k MRR closed, in five months. Another £25k+ MRR sitting live in pipeline behind it.

Three prior years, with a person on the ground, had closed nothing.

No budget. No CRM. No telephony. I stood up the operating layer myself, then ran deployment, onboarding, expansion and the partner channel. Seven clients signed and live, from first touch. Six agencies signed as strategic partners.

I also built the enablement the team runs on: a six stage onboarding path, eleven training modules covering 77 tracked competencies, and an objection knowledge base mined from ten recorded customer calls, every answer attributed to its source. The method is public as Blueprint 22 in gtm-stack. The content stays with the company, as it should.

Selling something nobody has bought before means building the case from scratch. What the category is. Why a senior buyer should care. What the number looks like on their side of the table.

That's the job. The engineering just makes it repeatable.

I needed the CRM and email platform for 490 retail brands.

The obvious build was AI agents plus a scraping API. 53k tokens per brand. A ten brand pilot burned 532k proving it worked.

Then I looked at what the agents were actually doing: fetching a page and matching strings against vendor names.

A regex does not need a language model. 😅

Rebuilt on SPF and DKIM records plus a homepage regex, it returned the same answers. 195 platforms confirmed, about 20 minutes, effectively zero tokens.

Most enrichment problems are pattern matching problems wearing an AI costume.

→ crm-scan · signal-scan

10,000+ sends. 16 campaigns. Ten Gmail-only domains.

The campaigns I didn't build were bouncing at 7 to 15% on the same infrastructure.

The difference wasn't a tool. It was verifying every lead before it ever touched a sequence, and refusing to load anything I couldn't confirm.

The engine books two meetings a week. Microsoft-stack targets I worked by hand, and cold called from my own phone, because a rule that says "this list is harder to land" is not a reason to skip the accounts.

46% of a 1,015 domain list sat behind Cloudflare and returned 403.

I could have bought a rendering service. Instead I used the signals that survive the wall.

Shopify answers on /products.json even when the homepage refuses. Store locators live in sitemap.xml, served to crawlers by design.

Recovered physical presence data for 1,483 domains. Cost: nothing.

A 180 row pass on pattern guessed emails came back 75% NXDOMAIN.

The domains did not exist. I had invented them.

The fix was not a better guessing algorithm. It was a hard rule: no guessed data ever reaches a system of record. Unknown stays blank and flagged.

Every rule I operate by has a scar like this behind it. I'd rather show you the scars than a case study with the failures edited out.

gtm-stack Audit a business, map the signals it can actually reach, rank plays by impact × ease, run one, then freeze the winner into a one command skill. 22 portable blueprints. No client data in it, by design. Python · PowerShell

signal-scan Prospect intelligence from public web signals. Loyalty platform, CRM/ESP, live offers, voucher code circulation, referral and student platforms, CDP, real app adoption. Scoring is config driven, so it retargets to any ICP without touching the code. Measured at 93 to 94% success and 1 to 3% field noise on a real 501 domain run. Python

crm-scan Detect a brand's email platform, loyalty stack and mail host from DNS and one page fetch. No keys, no vendor, no LLM. PowerShell

baskit A wishlist app that scores every saved item on price history, cool off and budget. Documented decision engine, 141 unit cases, a Playwright suite gating every deploy, self serve GDPR deletion. Built end to end by directing AI, with no engineering background. I still can't read most of the code. 🤷🏽‍♂️ Live at baskit.suhmantics.com TypeScript · Next.js · Prisma · Postgres

Python · PowerShell · TypeScript · Next.js · Prisma · Postgres · DNS / SPF forensics · Companies House API · Google Sheets API · Clay · n8n · Instantly · Klaviyo · Shopify · Vitest · Playwright

I don't think the interesting question is whether AI replaces salespeople.

It's what a salesperson can build now that they couldn't build two years ago, with no CS degree and no permission from anyone.

Turns out: quite a lot. 🚀

If you're building a GTM motion and the answer keeps coming back as "more headcount", I'd enjoy that conversation.

LinkedIn · London, UK

_{Every figure here comes from a run log or a signed contract, not an estimate. Where something is untested, it says so.}