DEV Community: Mike Wei

The fix for bad AI isn't a bigger brain. It's better structure.

Mike Wei — Thu, 09 Jul 2026 23:30:25 +0000

Most AI support bots top out at 30-50% resolution and get stuck around 3.5/5.0 CSAT.

Customers describe them with one word: robotic.

We kept hitting the same wall, and the reason turned out to be structural. A single generic model tries to force every refund, every connectivity issue, every warranty claim, every shipping question through one path. But real business problems aren't linear. They're full of ambiguity, overlapping systems, and incomplete information. Rigid decision trees and one big model both break the moment real complexity shows up.

So we stopped trying to build one model that knows everything.

Instead: divide and conquer. A sub-agent is a specialist focused on one domain — refunds, connectivity, shipping, warranty, finance, product damage. On each execution, a super agent plans the work, activates the right specialists, gathers facts from them, and makes the final call. Multiple sub-agents can contribute to the same request and cross-check each other — which is exactly why the output is more reliable than any single-agent system.

The results across Q1 2026 deployments:

83% resolution (vs. the 30-50% ceiling)
4.8/5.0 CSAT (vs. ~3.5)

And because each sub-agent is measurable — traffic, resolution, NPS, CSAT at the specialist level — you don't just learn whether the AI works. You learn about your own products, customers, and processes.

The counterintuitive part: running multiple agents per request should cost more than a single model. After continuous engine optimization, it's often more cost-effective than single-model alternatives — while doing deeper reasoning.

The lesson we keep relearning: reliability in AI doesn't come from a bigger model. It comes from better structure.

How the architecture works: https://aissist.io/technology/multi-agent-platform

The "Moore's Law for AI" everyone promised you? It didn't happen.

Mike Wei — Thu, 09 Jul 2026 23:24:39 +0000

Everyone building AI agents is about to learn the same expensive lesson: capability and economics are two different problems.

I've spent 15+ years building large-scale ML systems, and here's the uncomfortable math behind agentic AI in 2026:

→ Token prices haven't fallen in a straight line. The cheapest models have flattened out, and flagship pricing turned upward again this year (GPT-5.4 launched at 2x GPT-5's input price).

→ Reasoning tokens are billed as output tokens — deeper thinking raises cost even when the visible answer stays short.

→ One agentic request expands into planning, retrieval, verification, and tool calls. More useful work, way more tokens.

Betting your unit economics on "models will get cheaper" is not a strategy. Power constraints alone make that assumption shaky — AI demand is now colliding with the electricity grid, not just GPU supply.

What actually works is treating token efficiency as an architecture problem:

System design is the strongest lever. A system tightly coupled to the problem spends less because it already knows the workflow, the boundaries, and the next step. Generic agents pay a "rediscovery tax" on every single run.

Task optimization comes second. Routing, classification, extraction — these don't need frontier reasoning. The first big savings usually come from stopping the model from doing work it never needed to do.

Custom models come last, and only at scale. Trade a recurring inference bill for a predictable training cost — but only when the task is stable and you have real evaluation discipline.

This is why we treat cost as a core design constraint at Aissist, not a cleanup task for later. AI that can't scale economically fails the business case no matter how smart it is.

Full analysis: https://aissist.io/technology/token-efficiency