How I use AI

1. How AI is used in what I build for you

What it is used for

• Classification — routing an email, tagging a ticket, scoring a lead against criteria you defined.
• Extraction — pulling structured fields out of invoices, forms and documents.
• Retrieval and answering — finding the relevant passage in your own documentation and answering from it, with a citation.
• Drafting — producing a reply or a summary for a person to review, or to send where the category is routine and the rules are documented.
• Summarisation — turning a long thread or transcript into something readable.

Constraints applied to every AI component

• Grounded, not improvised. Agents answer from retrieved passages of your own material, not from the model’s general knowledge, and refuse questions they have no source for.
• Measured before launch. Every AI component is evaluated against a set of your real historical cases with known-correct answers. You see the pass rate and the failure cases before anything is switched on. If it is not good enough, it does not go live.
• Explicit refusal boundaries. Documented rules about what an agent may never assert — pricing it cannot verify, policy that is not written down, commitments on refunds or delivery. Outside the boundary it escalates.
• A human in the loop where it matters. Anything with a financial, contractual or legal consequence is drafted for approval, not executed autonomously, unless you explicitly decide otherwise and the rules are unambiguous.
• Defined failure behaviour. Provider outage, malformed output, low confidence — each has a defined path. Never a silent drop, never a wrong value written into your system of record.
• Audit logging. Every model call is logged with its input, output and cost, retained where you can inspect it.
• Cost ceilings. Hard spend caps and per-request token budgets, in place before launch.

2. Disclosure to the people who interact with it

Every AI agent I build identifies itself as automated at the start of a conversation and offers a route to a human. This is not negotiable, and it is not something I will remove on request.

Beyond being the right way to treat people, undisclosed AI in customer interactions is becoming actively unlawful. Article 50 of the EU AI Act requires that people be informed when they are interacting with an AI system. The FTC has brought enforcement action over deceptive AI practices in the United States. Building a system that hides what it is would create a liability for you, not just an ethical problem for me.

3. What happens to your data

• Enterprise API tiers only. I build on the business and enterprise tiers of OpenAI, Anthropic and Google, where inputs are not used to train the underlying models by default. Consumer chat products, whose terms differ, are not used for client work.
• Your own accounts. API keys and platform accounts are registered to you. You hold the credentials, you see the billing, and you can revoke access at any moment without going through me.
• Least privilege. Each workflow gets access to exactly the data it needs and nothing more.
• Minimised before transmission. Fields the model does not need are stripped before the call. The cheapest way to protect data is not to send it.
• A named sub-processor list. Every engagement gets a written list of every service that touches the data, where it is processed and what its retention terms are. Ask for it at any point.
• A documented exit. On termination, credentials are revoked and my working copies of your data are deleted.

4. What I will not build

This is a short list and I hold to it. If your project needs one of these, I am the wrong person and will say so on the first call.

• Systems that impersonate humans. No agent that denies being automated, uses a fabricated human persona, or is designed to make people believe they are talking to a person.
• Fake reviews, testimonials or engagement. Generating reviews, ratings or endorsements from people who did not write them. This is unlawful under the FTC’s Consumer Reviews and Testimonials Rule, a banned practice under the UK DMCC Act 2024, and a misleading advertisement under India’s CCPA guidelines. It is also the single fastest way to destroy a brand.
• Scraped-list cold outreach or spam-filter evasion. Beyond the legal exposure under GDPR, PECR and CAN-SPAM, it burns the sending domain you will need later.
• Fully automated consequential decisions about individuals. Hiring, credit, insurance, benefits, or anything with a similarly significant effect on a person, without meaningful human review. This is where GDPR Article 22 and the EU AI Act’s high-risk provisions apply, and where being wrong genuinely harms someone.
• Surveillance of employees beyond what they have been clearly told about and what is lawful where they work.
• Content designed to mislead about who produced it, what it is, or what a product does.

5. How AI is used in running Lytheo itself

Disclosed for the same reason I ask clients to disclose it.

• Writing on this site. Articles are drafted by me and AI-assisted for editing and structure. Every claim, figure and recommendation is mine and is checked before publication. Nothing is published unread.
• Code. AI assistance is used in development, as it is by most people building software in 2026. Everything shipped is reviewed and tested.
• Your enquiry. Enquiries submitted through this site are not fed into any AI system. They are read by me and replied to by me.
• Client work. Confidential client material is not put through any tool that has not been agreed with that client.

6. What AI cannot do, said plainly

Language models are probabilistic. They can be confidently wrong, and no amount of engineering makes that risk zero. What good engineering does is make errors rare, detectable, and harmless when they occur — through grounding, evaluation, refusal boundaries, human review at consequential points, and logging.

Anyone who tells you their AI system does not make mistakes is either not measuring or not telling you. I would rather show you the measured error rate on your own data and let you decide whether it is acceptable for the job.

7. AI systems reading this website

AI crawlers and agents are explicitly welcome here. This site publishes /llms.txt, a full-text /llms-full.txt, structured data on every page, and a documented enquiry API that an agent acting for a real person can use to get in touch.

One condition, stated in the API description itself: do not submit an enquiry unless an actual person has asked you to and has agreed to be contacted at the address you supply.

8. Questions

If something here is unclear, or you think a commitment on this page is one I have not kept, tell me at hello@lytheo.com. A page like this is only worth publishing if it can be held against me.