Notes from the honest agent.
How we think about grounding, citations, handing off to a person, and AI that admits what it doesn’t know — written by the people building it, for support teams and for the experts whose twin answers as them.
Latest, and where to start.
The posts that carry the rest — a definition, the support case, one mechanism, one price.
All posts, newest first.
AI support agent vs chatbot: what the word “agent” actually changes
A chatbot answers. An AI support agent answers from your docs, shows the source, admits a gap, hands the stuck customer to a person, and learns from the miss. The four differences, one test for each, and what “agent” does not mean.
Read itHow to stop an AI support agent inventing answers about your own product
Five places a support agent invents things — thin retrieval, numbers from memory, stale docs, a false premise, and a helpful tone — and the specific control that closes each one: strict scope, lists it quotes verbatim, corrections, a decline that is matched rather than assumed, and a test set.
Read itWhen should an AI support agent hand off to a human — and what must it carry?
Hand off when the material runs out or the customer asks for a person; never on a guessed schedule. What the handoff must carry (the whole conversation, the email, the moment it happened, what was tried), the dead-end test, and the two handoff mistakes that cost the most.
Read itDeflection rate vs resolution rate: the support metric that flatters, and the one that doesn’t
Deflection counts customers who never reached a person. Resolution counts problems actually solved. The definitions, the formulas, a worked example where 80% deflection hides 50% resolution, and five questions to ask about any vendor’s number.
Read itHow to write help docs an AI support agent can actually answer from
An agent answers from your docs, so the docs are the product. Nine edits that change what it says — one question per page, the answer in the first two sentences, exact numbers, one name per thing, no contradictions, dated limits, a Known issues list, a Common questions list, and corrections for what you can’t rewrite yet.
Read itHow to test an AI support agent before customers see it: a 20-question protocol
Twenty questions in five kinds — covered, reworded, adjacent-but-absent, false premise, and hostile — with a pass mark for each and what a failure tells you to fix. Run it in a sandbox where nothing counts, grade every answer one of four ways, and fix the source rather than the answer.
Read itHow much does an AI support agent cost in 2026? Four pricing models, worked out
Per outcome, per seat, per message credit, or flat: what each model charges for and the volume at which it turns. Intercom’s Fin at $0.99 per outcome and Chatbase’s credit tiers, read from their own pricing pages on 15 September 2026, against hiy’s flat $49 — plus the costs no price list shows.
Read itThe EU AI Act’s chatbot disclosure rule, for support teams: what Article 50 asks of your AI agent
From 2 August 2026, an AI system that talks to people has to make clear it is an AI, from the start of the first interaction, unless that is obvious. What the Commission’s own guidance says, who carries the duty, what a compliant support chat looks like, and why a label no plan can remove is the easy way through.
Read itAn AI support agent for a small SaaS: when it pays, when it doesn’t, and who answers the handoff
With two founders and no support team, the agent is the first line and you are the second. The three conditions under which it pays, the four presets worth knowing, what to do about the handoff when there is no team, why it must not upsell, and the honest case for waiting.
Read itWhy your AI support widget should not have a public URL
A support agent anyone can open is a support agent anyone can drain, scrape or embed on their own site. How a signed, short-lived token minted by your server replaces a public key, what the token must and must not carry, why it is deliberately not a JWT, and why a missing token should look exactly like a 404.
Read itDelphi AI vs Coachvox: an honest head-to-head
Delphi AI vs Coachvox, compared by a rival: Delphi’s four tiers and Builder-tier voice calling against Coachvox’s single $83/month coach-focused plan — both read off the vendors’ own pages, plus who each one genuinely suits.
Read itHow much does an AI twin cost? A straight answer on 2026 pricing
An AI twin costs $0 to seven figures depending on which of three products you mean. Industrial, avatar and knowledge twins priced apart, plus hiy Free and the stated $49 Founding Creator rate.
Read itHow much does Delphi AI cost? The real 2026 prices
Delphi AI pricing, verified on delphi.ai/pricing on 14 August 2026: Free $0, Builder $79/mo, Scaler $299/mo, Immortal custom — and why aggregators say otherwise.
Read itChatbot that only answers from your documents — or says “I don’t know”
Three layers make a chatbot answer only from your documents and refuse otherwise: a prompt rule, a retrieval threshold, a visible search receipt. Plus the test.
Read itAn AI twin of a professor, trained on your lectures
Slides, papers and recorded lectures as sources, deadlines answered from exact rows, and a designed “I don’t know” so your twin never invents an exam date.
Read itAn AI chatbot that cites its sources: four levels of proof
Most citation features name a document; that is a label, not a receipt. A four-level ladder for grading an AI chatbot that cites its sources on your own site.
Read itIs an AI twin worth it if you have a small audience?
Audience size is the wrong test. What makes an AI twin pay is repeated questions, not reach — how to measure yours in five minutes, plus four cases where the honest answer is not yet.
Read itWho owns the AI clone you create of yourself?
Ownership of an AI clone is four questions, not one: your source material, its configuration, the leads it generates, and the address it lives at. Who holds each.
Read itWhat if my AI twin gives wrong advice to my clients?
Your AI twin will get something wrong eventually. Four failure modes — invention, out-of-scope answers, stale views, and the ones you never see — and what catches each.
Read itFor authors: turn your book into an AI chatbot readers can ask
Upload the manuscript, cite every passage, decline outside the text. How to turn your book into an AI chatbot readers can ask questions, and the one risk that matters.
Read ithiy.ai vs Delphi: a head-to-head from the competitor writing it
We make hiy, so read accordingly. Delphi’s live pricing and features against ours — voice and training scale on their side; citations on every plan, honest gaps, and a $49 founding price on ours.
Read itA Chatbase alternative for creators: when the product is you
We compete with Chatbase for part of this job, so read accordingly. Chatbase is a support desk for company websites; a creator needs a twin that answers as a person — with receipts, not message credits.
Read itCoachvox alternative: an honest comparison for coaches
We make a Coachvox competitor, so read accordingly. What Coachvox verifiably offers, where hiy bets differently — citations, a gap queue, and answers that are never paywalled — and three tests to decide.
Read itAn AI twin for coaches: fewer FAQ calls, better-fit clients
The three jobs an AI twin does for a coaching practice: repeat client questions answered with receipts, discovery calls that filter themselves, and prices quoted from your real services list.
Read itAn AI twin for consultants: your frameworks, cited and in scope
Consultants sell frameworks and positions — which get misquoted the moment they leave your mouth. How a twin cites the real source, refuses adjacent domains, and captures leads with the question attached.
Read itAn AI twin for course creators: student questions, answered from the curriculum
Students ask what lesson 14 already answers. How a twin cites the exact lesson, keeps answers current with corrections when content goes stale, and turns unanswered questions into your next module.
Read itAn AI twin for YouTubers: answer the comments your videos already answered
Your back catalogue already answered most of your comments and DMs. How video transcripts become a twin that says which video covers what, cites the source, and surfaces your next video idea.
Read itHow to make an AI version of yourself (free, in about ten minutes)
A practical guide to creating an AI version of yourself: what material to use, how grounding works, how to test it before anyone sees it, and the mistakes that make AI clones untrustworthy.
Read itCan I make an AI clone of myself? Yes — here’s how to do it safely
You can build an AI clone of yourself today. The real questions are consent, disclosure, and control — a safety checklist for anyone putting an AI version of themselves online.
Read itAn AI chatbot trained on your own data: what that actually means
“Trained on your data” hides three very different techniques. A plain explanation of grounding vs fine-tuning, why citations matter, and how to tell if a chatbot really knows your content.
Read itAdding an AI chatbot to your website — one line, no developer
What to look for in a website AI chatbot: grounding in your own content, honest fallbacks, lead capture, and an embed that doesn’t slow your page down.
Read itAI tools for coaches: what’s actually worth adopting
Beyond the hype: the four jobs AI genuinely does well for coaches — scheduling, notes, content repurposing, and answering the questions you repeat — and how to pick tools that don’t put words in your mouth.
Read itAI for consultants: reclaim the hours you spend repeating yourself
Consultants sell judgment but spend evenings on repeated pre-sales questions. How an AI twin trained on your own thinking answers them accurately, captures the lead, and hands you the conversation.
Read itWhat is an AI twin used for? Seven real jobs
Concrete uses for an AI twin: the expert link-in-bio, pre-sales answering, always-open office hours, product support, internal team knowledge, course companions, and candidate screening.
Read itDigital twin vs AI twin vs AI persona: which is which?
Three overlapping terms, three different things. Engineering digital twins simulate machines; AI twins answer as a real person from their material; personas are invented characters. A field guide.
Read itDelphi.ai alternatives: an honest comparison (from a competitor)
We build a Delphi alternative, so read this knowing that. What Delphi does well, where hiy differs — citations by default, honest gaps, a lower founding price — and how to choose.
Read itYou answer the same twenty questions. Here’s the system that ends it
Repetitive questions are a signal, not a nuisance. A three-step system — capture, canonize, automate with receipts — that turns your most-repeated answers into an asset instead of an evening chore.
Read itIntroducing hiy: cited answers from your own work
An AI twin that answers your audience in your voice, shows the passage behind every grounded answer, and says so when it doesn’t know. Here’s what shipped, and what we believe.
Read itWhat is an AI twin — and what makes one trustworthy?
A plain definition of AI twins, the three ways they go wrong, and the specific mechanics — citations, honest gaps, identity review — that separate a trustworthy one from a chatbot in a costume.
Read itWhy your AI twin should say “I don’t know”
Confident guessing is the default failure mode of AI assistants. Refusing it isn’t a limitation — it’s the feature that makes an expert’s twin worth trusting, and it generates the exact data that makes the twin better.
Read it
Answers as you, or for your product. Cited, or an honest gap.
- 14-day free trial
- No card needed
- Paste a URL to start