Daniel Brunsdon

The gap

AI was creating a two-speed company. Engineering was seeing 30x productivity gains through Claude Code — but every other department was stuck at roughly 30% improvement. Marketing, BD, ops, HR — all using AI, all underwhelmed.

The assumption was that non-technical work was harder to automate. That wasn't it. The real bottleneck was legibility: the company's knowledge was scattered across a dozen SaaS platforms that AI agents couldn't read, couldn't search, and couldn't act on. Notion wikis, Framer pages, password vaults, HR platforms — all walled gardens with their own auth, their own structure, their own search. Agents could write code because codebases are organized. Everything else was a mess.

The diagnosis

I mapped the actual workflows — not the org chart, not the tool stack, the workflows. What surfaced: most non-engineering work followed a pattern of gather context → make a decision → produce an output. The gather step was where 70% of the time went, and it was where AI hallucinated most frequently. Agents were guessing because they couldn't see.

The intervention point wasn't better prompts or fancier models. It was the information architecture underneath.

The redesign

Built a structured internal knowledge system — indexed markdown with YAML frontmatter covering roadmap, OKRs, active tasks, CRM data, brand guidelines, product specs, and operational procedures. Everything an agent needed to do its job, in a format it could actually parse.

Then built a unified credential layer across 25 services — AWS, Cloudflare, Twitter, databases, email, analytics, DNS providers. One interface for agents to authenticate and act, with human approval gates on anything destructive.

This wasn't a cost play. It was a legibility play. The SaaS tools got replaced not because they were expensive, but because they were opaque to AI:

• Framer → markdown site templates agents could generate and deploy
• Notion → structured docs with machine-readable frontmatter
• Typefully → direct API scheduling through the credential layer
• 1Password → unified secrets management agents could query
• HR platform → structured employee data in the knowledge base

Five tools eliminated. ~$5,700/yr in direct savings. But the savings were a side effect — the real unlock was that agents could now see everything.

The results

The before/after was immediate:

Three months of work in three days. A GTM push that would have taken a quarter — branded materials, campaign copy, partner outreach sequences, landing pages — was generated, reviewed, and shipped in a long weekend. Not by cutting corners. By removing the context-gathering bottleneck entirely.

Hallucinations went from frequent to near-zero. When agents can reference the actual brand guidelines, the actual product specs, the actual CRM data — they stop making things up. Context management, not model selection, was the quality lever.

Non-engineering productivity went from +30% to multiplicative. The structured docs system turned every department into something closer to a codebase — organized, searchable, version-controlled. The same Plan-Execute-Verify loop that made engineering fast now worked everywhere.

25 credentials, one interface. Sensitive operations — DNS changes across cloud providers, production database migrations, infrastructure audits — handled through a single authenticated layer with human approval on anything irreversible.

We wrote about the technical approach in detail on the company blog.

The takeaway

Most companies are bolting AI onto existing workflows. The unlock isn't there. It's in making your company legible to AI — restructuring how knowledge is stored, accessed, and acted on so that agents can do the gather step as well as a tenured employee.

You don't automate roles. You make your organization readable. The automation follows.