About the AI Adoption Agent

1. What this is

The AI Adoption Agent is a working prototype that automates one specific piece of B2B sales intelligence: finding companies, estimating their AI readiness from public signals, and producing a draft first-touch email calibrated to that readiness.

It is built as an application artifact — a proof of concept demonstrating autonomous agent architecture in a context directly relevant to GO MO Group's existing work with AI-led B2B marketing.

2. Pipeline

1. 1
   Search — Serper.dev queries Google for companies matching the ICP description. Social media, news aggregators, and marketplaces are filtered out.
2. 2
   Scrape — Homepage and About page content is fetched via cURL. Text is extracted (scripts, nav, footer stripped). Truncated to 5,000 characters per page.
3. 3
   Classify — Claude Haiku assigns a 1–5 AI maturity score with rationale and explicit signals. It never invents signals — if the text doesn't support a higher score, it doesn't get one.
4. 4
   Outreach — Claude Sonnet drafts a segment-calibrated email. Score 1–2 gets an education angle. Score 3 gets a pilot angle. Score 4 gets an optimization angle. Score 5 gets a skip note.

3. AI Maturity Rubric

4. Why this was built

GO MO Group has published detailed work on AI-led B2B marketing — including a deployed GenAI Content Supply Chain for a global manufacturer and a productised Generative Engine Optimisation (GEO) service. The ASBX process and Pull Marketing framework show a systematic approach to AI-augmented sales.

This prototype demonstrates the specific capability gap that comes before that framework kicks in: automatically identifying which prospects are ready for what conversation. The classification at the top of the funnel changes which content personalisation strategy is appropriate — education for some, optimisation for others.

It is built to live under gomogroup.com, not as a standalone product. This is an input into an existing system, not a replacement for it.

5. Known Limitations

• ▪Public signals only. Classifications are based on what companies publish on their website. A company doing significant internal AI work without publicising it will score low. This is a known bias.
• ▪Phase 1 sample. The Heatmap covers 27 hand-seeded Västsverige companies, not a statistically representative sample of the region.
• ▪Text only. The scraper reads HTML text. It does not process video, PDFs, LinkedIn, press releases, or job postings — all of which would improve signal quality.
• ▪One classification pass. Scores are not averaged across multiple runs. A single scrape on a day when a page is cached or down could affect the result.
• ▪Cross-domain duplicates not deduped. The same company can appear multiple times in one run if its presence spans multiple domains (e.g., main site, investor portal, third-party employer page). Deduplication is by domain, not entity.
• ▪Thin-content fallback. When a scraped page returns very little usable text, the classifier still assigns a score 1 ("Dormant") rather than flagging insufficient data. A future patch will gate classification on minimum scraped content.
• ▪No send functionality. Outreach drafts are copy-to-clipboard only. This is intentional — review and personalisation by a human is required before any contact.
• ▪Shared hosting constraints. The live agent is limited to ~90 seconds and ~8 companies per run due to Oderland's execution limits.

6. Tech stack

7. Contact

Built by Andreas Persson.
Questions about this prototype: andreas@perssonofsweden.com