There are more AI tools than any team could evaluate in a year, and a new “must-have” lands in your inbox every week. Most of them are built to win a demo, not to survive your Tuesday. The stack we actually recommend is shorter than you’d expect — and that’s the point.
Here’s the rule we hold every tool to: does it remove work, or just add another dashboard to check? If it adds a tab someone has to remember to open, it’s already losing.
Start with one model, not ten
You don’t need a shelf of AI products. You need one capable frontier model wired into the places work already happens. We default to Claude for most reasoning, writing, and analysis, and we resist bolting on a second or third until there’s a concrete job the first one can’t do. One model, used well, beats five used shallowly.
Put it where the work already is
The fastest adoption we ever see comes from meeting people inside the tools they already live in — the inbox, the doc, the CRM, the help desk — not a shiny new app you have to train everyone to open. A good integration is invisible. The win shows up in the workflow people already run, not in a separate destination.
The layers that actually matter
When we map a stack with a client, it comes down to a handful of layers — and most teams only need the first two to start:
- Model — one frontier LLM as the workhorse.
- Where it runs — embedded in existing tools and workflows, not a new surface.
- Knowledge — connect your own data only when a generic model genuinely can’t answer; retrieval is powerful and frequently over-applied.
- Orchestration — light automation to chain steps once a manual workflow has proven it’s worth automating.
- Guardrails — logging, review, and a human in the loop wherever output touches a customer or a dollar.
The best AI stack is the smallest one that ships the outcome — every extra tool is a tax you pay forever.
What we tell teams to wait on
Plenty of impressive technology isn’t ready to be load-bearing in your business yet. We routinely tell clients to wait on fully autonomous “agent swarms,” rip-and-replace platform migrations, and anything whose entire pitch is that it’s new. None of it belongs in your first ninety days. Sequence it deliberately — the roadmap should say as much about what to hold as what to buy.
Small on purpose
A stack you can explain on one page is a stack your team will actually use. Start with one model, put it where the work happens, attach a metric, and add the next layer only when the current one is paying for itself. Everything else is noise dressed up as progress.