> ## Content Index > Fetch the complete content index at: https://intelligence.artificialrealitymedia.com/llms.txt > Use this file to discover other available public pages before exploring further. # What back-of-house AI is, and the five questions to ask before deploying - URL: https://intelligence.artificialrealitymedia.com/back-of-house-ai-deployment-test/ - Published: 2026-08-06T09:00:00.000Z - Updated: 2026-08-06T09:00:00.000Z - Description: Back-of-house AI puts a knowledgeable human between the output and the customer. Five questions decide whether your deployment qualifies, or should wait. - Author: Galina Fendikevich - Tags: Cross industry, #For AI Buyers, #AI tools, #Import 2026-09-03 04:53 Back-of-house AI is any deployment where the output passes through a knowledgeable human before it reaches a customer or a consequential decision. Across six operators in seven industries in Artificial Reality's first season, that placement, rather than the underlying technology, decided the outcome. The term is borrowed from hospitality: front of house is guest-facing service, back of house is the kitchen, the finance office and operations. ## Why placement beats tool choice The same technology shows up on both sides of the line, in the same companies. A major hotel chain's guest-facing concierge could not recommend a nearby restaurant when a reporter tested it — an account discussed on the show and not independently sourced by Artificial Reality — and the same company then pointed the technology at revenue management, demand forecasting and operations and recovered the cost. [ChatGPT](https://chatgpt.com/?ref=intelligence.artificialrealitymedia.com) told a yacht charter client he could safely anchor in 25 knots of wind, which the captain flatly contradicted. The same class of tool read a forty-five-page commercial lease for a restaurant founder in three minutes, found eight red flags, and saved a $1,000 attorney review that instead took ten minutes and cost nothing. The distinction is not customer-facing versus internal. It is who absorbs the error. ## The five-question deployment test Before deploying AI anywhere in your business, ask these in order. **1\. Who sees the output first?** If the answer is a customer, stop. If the answer is a member of your team who could tell a wrong answer from a right one, proceed. **2\. Would that person catch the specific errors this tool makes?** The couple with the $5 million valuation could not evaluate whether departments should be counted as assets. The captain could evaluate anchoring conditions instantly. Domain expertise in the reviewer must match the domain of the error. **3\. What does a wrong answer cost, and who pays it?** In Sofiane Ghorbel's framing, a wrong recommendation costs a relationship built over years — a cost that lands on the business, not the tool. Price the failure, not the subscription. **4\. Is the AI's confidence visible to whoever receives the output?** In all three failures the AI expressed no uncertainty. If the recipient cannot tell a reliable answer from an unreliable one, the deployment is front of house regardless of where it sits in your org chart. **5\. Is the task volume work or judgment work?** Reading forty-five pages is volume. Deciding which journalist to pitch, or what a business is worth, is judgment. Season 1's successes are overwhelmingly volume tasks with judgment retained by the human. ## What passes, and what waits Deploy now, back of house, reviewer in the loop: - Document review and summarisation before a professional signs off - Translation and drafting inside an inbox, sent by a person - Forecasting, revenue management and operations analytics - Context and knowledge organisation that feeds a human's work - Internal research where the researcher can evaluate the answer Wait, front of house, customer absorbs the error: - Customer-facing chat and concierge for anything advisory - Any tool a customer will consult independently about your domain - Automated recommendations that reach a client unreviewed - Valuations, quotes and assessments delivered straight to the person they concern - Safety-relevant answers of any kind ## Does this mean customer-facing AI never works? No. It means customer-facing AI in an advisory role, where the customer will act on the answer and cannot judge its quality, carries a risk profile most businesses have not priced. Transactional customer-facing AI such as scheduling is a different case, and several operators report it working well. ## AI tools named in this report | Tool | Named by | Verdict | Used for | | ------- | ------------------------------------------------------------- | ---------------------------------------------------- | --------------------------------------------------------- | | ChatGPT | A restaurant founder, Season 1 | Worked — back of house, attorney reviewed the output | Reading a 45-page commercial lease for red flags | | ChatGPT | Jennifer Kerum, yacht charter broker, on behalf of her client | Didn't work — front of house, no expert in between | A client asking whether it was safe to anchor in 25 knots | *Tools named by operators on the record. Inclusion is reporting, not endorsement.* ## What to do this week Run the five questions against whatever you are about to deploy, and move anything that fails question one behind a reviewer rather than cancelling it. The tool is rarely the problem. If you want the cheapest starting point, it is document review with professional sign-off: three minutes of AI work, ten minutes of review, $1,000 of billing avoided, and reproducible by almost any small business without buying anything new. Revisit the front end when you can put a competent reviewer between the model and your customer. ## Where this comes from S1E8: Why 99% of AI products fail — the full interview with Galina Fendikevich. Listen or watch: [YouTube](https://www.youtube.com/watch?v=WLzJ7U6inpw&ref=intelligence.artificialrealitymedia.com), [Spotify](https://open.spotify.com/episode/1DI6RkvekF4NF7HfpuzEqm?ref=intelligence.artificialrealitymedia.com) or [Apple Podcasts](https://podcasts.apple.com/us/podcast/why-99-of-ai-products-fail-the-gap-between/id6780699025?i=1000780319385&ref=intelligence.artificialrealitymedia.com).