Loading
AI guidance usually assumes you want the most capability switched on soonest. These assume the opposite: that a wrong answer in front of a customer costs more than a slow rollout.
Four of the five are about restraint — what to keep manual, when to refuse, how to measure honestly, and the order to enable things in.
Sequenced, not alphabetised. Each one ends with something configured.
Every guide here ends up recommending the same sequence: internal first, then agent-assist with a person checking, then customer-facing on topics your articles genuinely cover.
Citations and refusals is the conceptual one — why an answer that cannot cite a source should not be sent, and why the refusal rate is a useful output rather than a failure.
Measuring what it saves is the one everybody skips, because it requires baselining before switching anything on, and by the time people want the number it is too late to collect.
Every guide is written to be finished rather than to be impressive: the screen each decision happens on, and the thing that goes wrong if you skip it.
Pulse tools →If you are about to switch anything on, read turning Pulse on safely, then measuring what it saves — the second one has to happen before the first.
If you are evaluating rather than implementing, read citations and refusals. It is the argument the rest depends on.
No. They are written to be useful against whatever you run today. Several describe decisions that apply to any pulse-shaped system, not just ours.
Each has an owner and a review date. Where something has changed, the guide is updated rather than supplemented, so there is one current version rather than a thread of corrections.
Because the limits are the useful part. A guide that recommends everything is a brochure, and the decisions worth taking care over are usually the ones with a real cost either way.
Half an hour with your own data usually saves reading three of these. The guides will still be here afterwards.