Six months of building B2B AI - the surprises that didn't make the optimistic Twitter version
Build-in-public threads usually report what's working. Worth occasionally reading the version that reports what isn't.
Some surprises from the last six months of building B2B AI tooling for enterprise buyers:
→ Sales cycles are 2-3x longer than the optimistic forecast — even when the buyer is motivated.
→ 'No-code' is a selling point for marketing — and a friction point in deployment when engineering teams need to extend the configuration.
→ 'Model-agnostic' is genuinely hard to explain to non-technical buyers. Most don't have a strong feeling about which LLM is under the hood.
→ Community-building as a marketing channel is the highest-ROI investment that still feels the slowest to compound.
→ The most useful customer conversations happen 90 days after deployment, not before.
What's the honest, non-optimistic version of your last six months?
Useful threads come from people who are willing to say the thing that doesn't fit on a victory lap.
Replies
The 90-days-after-deployment point feels especially underrated. Pre-sale feedback often tells you what buyers wish were true; post-deployment feedback tells you what actually survived messy workflows, politics, and handoffs.
My non-optimistic version: “AI quality” is rarely the blocker people think it is. The blocker is usually whether the customer has a clean enough source of truth for the AI to keep being useful after the demo.
@jim_jeffers The "clean enough source of truth" point is painfully accurate. We've had clients blame the AI for giving bad answers when the real problem was their internal knowledge base hadn't been updated in 8 months. The AI was doing exactly what it was told, just with garbage in. Nobody wants to hear "your documentation is the bottleneck" during a post-deployment review, but that's been the case way more often than actual model quality issues for us.