The trust ladder
Say"Holly earns its autonomy. You start it cautious, you see what it does, and you give it more only when it's earned it. It's set per person and you can always move it back."
Level names, per-person setting and defaults are confirmed. The descriptions are how the levels are designed to work; exact behaviour is to be confirmed with Dom. Rehearse each level you plan to show, and describe nothing beyond what you have seen.
| Level | Designed to | Suits | How to say it |
|---|---|---|---|
| Observe | Watch and show what Holly would do, without acting | Regulated firms, nervous users | "Holly watches and shows you what it would do." |
| Suggest | Propose ideas and drafts as cards to edit, approve or dismiss | Everyone else (the default) | "Holly proposes; you decide what goes ahead." |
| Auto-draft | Write drafts on its own for review | People approving drafts with few changes | "Holly drafts without being asked; you review." |
| Auto-publish | Publish without asking each time, within the customer's rules | Deliberate decision, person by person | "Auto-publish is the top level customers can choose. For LinkedIn we recommend a person approves every post." |
Do not demo Auto-publish until Dom confirms it is live at launch.
Two kinds of approval. "A person approves" can mean an author approving Holly's draft of their own post. "Approvals on" is the workflow: a second approver or a chain. The customer configures it, so never say "nothing ever publishes without approval". Say: "You decide what needs approval. Nothing you've set to need approval goes out without it."
Defaults (D8, confirmed 19 October 2026). Suggest for everyone else. Observe for regulated firms (law, financial services, accountancy, any firm that asks), with Suggest as the first small win. A customer may choose any level: we explain our recommendation and record their choice, never refuse. Approval workflow on for new starters and anyone posting about clients, and for every post in regulated firms.
Moving up and back. Moving up is the customer's choice, per person, reversible. Never push it. Suggest a trigger, not a date: "When someone's approving most of Holly's drafts with barely a change, that's the moment to talk about Auto-draft for them." A pre-commitment rule (for example "ten drafts in a row approved unedited, move me to Auto-draft") is a rule of thumb only, not a confirmed feature. Moving back is blameless: "If Holly has a bad week, you drop it a level."
LinkedIn and Auto-publish. Recommend a human decision on every LinkedIn post (reasons for you: platform terms on automated posting, being checked by Karl Foster; the EU AI Act's exemption for text a person has reviewed; reputation). Never explain the legal reasoning to a buyer or present it as legal advice.
Say"Auto-publish is one of the four levels, and it's your call per person. For anything going on LinkedIn under someone's name, we recommend a person approves every post. It's what we recommend to every customer."
Say (if they want Auto-publish for everyone on day one): "I'll be straight: I'd talk you out of that. Start on Suggest, look at what Holly writes for a few weeks, then decide person by person. If it's as good as you hope, you've lost a few weeks. If it isn't, you've protected your reputation."
Say (the risk-off close): "Here's what I'd recommend. Everyone starts Holly on Suggest" (Observe for regulated firms) ", with approvals on. We book your onboarding before we finish today. And we put a 30-day review in the diary now, where we look at what Holly wrote, what people approved, and decide together whether anyone moves up. Does that make sense as a next step?"
Handover to customer success. Record per person in the SPICED record: starting level, whether approval workflow is on, the agreed trigger for moving up, and the 30-day review date.