About the Playbook
Read enough teardowns and the same hard problems surface again and again, solved in recognizably similar ways. The Applied AI Playbook lifts those problems out of any single company and treats each one on its own: what makes it hard, the patterns that recur, the tools and popular choices teams reach for, a reference architecture, and the best practices — each tied back to the teardowns it’s drawn from.
This is the cross-company comparative map the teardowns point toward. It’s synthesis, not a per-company reconstruction: the unit here is the problem and its solution space, evidenced by linking to the teardowns where each move shows up — not a per-claim confidence tier (that discipline lives in the teardowns themselves). For the vocabulary underneath these patterns, see AI Engineering Terms.
The map
Section titled “The map”Each row is a problem common to applied-AI products and the page that works it through. Pages link to the teardowns that supply the evidence.
| Challenge | The problem in one line | Seen across |
|---|---|---|
| Graduating an agent from assistant to actor | Crossing from “suggest” to “act” without losing user trust | Antimetal, Prophet Security, Pallet, Basis, Confido, Amperos |
| Testing output that isn’t reproducible | No fixed expected output, so a normal test suite can’t gate changes | Glean, Rilla, Traba, Momentic, Basis |
| Keeping inference cheap & fast | Frontier-model calls on every step blow up cost and latency | Basis, Glean, Momentic, Traba |
| Reaching systems with no clean API | The systems of record are legacy portals built for humans, not machines | Pallet, Amperos, Confido, Momentic |
| Retrieval at multi-tenant scale | Grounding every tenant in its own knowledge without per-customer code | Pallet, Glean, Rilla |
| Surviving long, multi-day workflows | Multi-party flows must survive crashes, retries, and partial failure | Pylon, Gradient Labs, Harvey |
| Beating context degradation | One agent’s context grows until quality quietly falls off | Traba, Antimetal, Basis |
| Encoding dense domain rules | Regulatory and underwriting logic must be exact, testable, and auditable | Pylon, Basis, Comp AI |
| Own vs. rent the model | Deciding where to spend a training budget vs. renting a frontier LLM | Rilla, Basis, Harvey, Pallet, Glean |