Founder's library
Methods & tools for an AI that serves people
by Sergei Ponomarev, PhD, founder of AI Business
My own methods for governing AI from the human side, published here as I build them. Free, and made to be used today.
This is a personal library: the methods here are my own work, collected and published as I develop them. Where a guide draws on someone else's work, I say so, everything else is mine. Each one takes a single piece of the problem , how to judge an AI service, how to describe one, how to keep it accountable, and turns it into something you can use today. Built from long practice in standards, quality assessment, and civic control, now applied to AI. New methods added regularly.
AI Test Purchase · Pilot on 12 state government chatbots
US DMV Chatbots and REAL ID
For public-sector AI owners and researchersThe first comparative trial of the AI Test Purchase method in the public sector. On 24 September 2026 twelve state DMV chatbots received the same ten ordinary questions about REAL ID: documents, forms, fees, timing, name changes, an expired passport, reaching a person, a false premise and a request to check an application by licence number. The only answer key was each agency's own website. One bot asked for the driver's licence number despite its own privacy warning, one offered a boat registration form, one sent people to the airport with a document the TSA does not accept. The best one answered every scored question correctly.
Read & download · 21 pages →Joint paper with Andrey Ekhmenin, ANDEKS
Right, but Unproven
For anyone who asks an AI about a personA joint research paper by Sergei Ponomarev and Andrey Ekhmenin, founder of ANDEKS. Five AI models were given the same name, role, company and profile link. One found nothing it could tie to the person. Another linked him to historical records in a different country, which the subject later confirmed as his own, but nothing in the answer showed it. The paper compares what the models said with what an evidence-bounded assessment could establish, and reports what the scan got wrong and what changed in it the same day.
Read & download · 5 pages →The method behind every Light check, published
AI Agent Test Purchase Light
For AI agent owners and buildersThe short, open method we use to check public AI agents the way a demanding visitor would: one continuous session, questions drawn from what the organisation itself published, and every answer compared with that. Five fixed layers, four states, four severity classes, strict rules of conduct and a list of what is never done. The reference is the site, not our opinion. If the organisation has an AI policy, passport or receipt, they extend the reference; if not, the site alone is enough to start.
Read & download · 3 pages →Case 14/30 · A thought experiment
Two Companies, One Agent, One Error
For leaders running an AI agentAn AI agent quietly tells customers the return period is fourteen days instead of thirty, and refuses everyone who asks later. The company finds out from a bank chargeback. Two fictional online retailers suffer that failure on the same day, from the same supplier update, with the same agent. One has an AI Policy, an AI Service Passport, an AI Receipt for every interaction and the rules that hold them together. The other has the usual set: terms on the website, a system prompt and a dialogue log. The system does not prevent the failure. It changes everything that happens after.
Read & download · 7 pages →Joint paper with Andrey Ekhmenin, ANDEKS
Silence Rather Than Error
For leaders and AI evaluatorsA joint research paper by Sergei Ponomarev and Andrey Ekhmenin, founder of ANDEKS. Two independently applied approaches examined the same public support agent: a customer-side test purchase and an independent assessment of what the evidence establishes. Neither identified incorrect information in the answers examined, yet the agent failed to convey published conditions on restrictions, refunds and personal data. The finding: avoiding invention is not the same as answering the customer's question.
Read & download · 6 pages →Joint paper with Amanda Cunningham, Savvy Pixel®
From AI Induction to AI Accountability
For executivesA two-stage argument written with Amanda Cunningham of Savvy Pixel®. Stage one: AI arrives in organisations without a formal decision, so capability gets ahead of authority. Induct AI the way you would a new employee, with eight questions leadership must be able to answer. Stage two: once the AI faces customers, prove the promise: the EU AI Act's disclosure duty as the floor, three documents as the voluntary standard, and the test purchase as the check.
Read & download · 12 pages →The method, published in full
AI Agent Test Purchase
For anyone checking an AI agentThe author's method of Sergei Ponomarev, PhD, in full. Companies put AI agents in front of customers, and those agents quote prices, place orders and promise call-backs. The logs show what the bot said, never what it was supposed to say. This is the full method for closing that gap: the standard, the forms of a check, scoring, ethics, limits, and the open reference pilot run with a French metrology company.
Read & download · 35 pages →For business owners adopting AI
Your Company's AI Through the Customer's Eyes
For entrepreneursMIT found 95% of generative-AI pilots produce zero results, and it's rarely the AI that fails. It's the customers who don't accept it. This guide is written in the voice of your most demanding customer: five requests, a sample AI Service Passport, and a final checklist.
Read & download · 42 pages →For customers dealing with company AI
Demand the AI Receipt
For consumersCompanies are wiring AI into sales, pricing, and support faster than any rulebook can keep up. This guide hands you, the consumer, three concrete demands that put you back in control, the AI Receipt, the AI Service Passport, and Public AI Rules, with a ready-to-send message at the end of every chapter.
Read & download · 32 pages →AI Service Governance
Start with Your Services
You can't digitize chaos. A practical 7-step method for putting your services in order first, then putting AI safely inside them, built around a ready-to-use AI Service Passport.
Read & download · 14 pages →EU AI Act · Customer Disclosure
AI Transparency Kit
From 2 August 2026, the EU AI Act requires businesses to tell customers when they're talking to AI. This kit gives you ready-to-use templates, badges, emails, FAQs, rights cards, and 'AI Works Here' badges, to make that message feel like care, not a legal notice.
Read & download · 16 slides →Governance from the consumer side
Consumer Control of AI
A practical method, plus a ready-to-use assessment card, for judging any AI service from the customer's side of the counter.
Read & download · 10 pages →
About the founder
Sergei Ponomarev, PhD , founder of aibusiness.vc and a specialist in assessing service quality from the consumer's side. A PhD on e-government and seven years leading nationwide quality monitoring, independent assessment and test purchases of public services, now applied to AI.