AI is part of how PerrfectSite works. It reviews code, drafts documentation, accelerates research, and removes a large amount of repetitive effort from every project. Clients benefit through faster delivery and lower cost.
That comes with obligations. This page sets out what those are, in terms specific enough to hold someone to.
About this website
Content on this site is produced with AI assistance. Saying so is more useful than letting visitors guess.
Images are AI-generated unless credited otherwise. Written content is drafted with AI support and then edited by a person, and almost every page here undergoes human revision, correction, or rewriting before it goes live. Nothing gets published straight out of a model.
The tooling is built in-house. Custom agents handle drafting, research, and formatting according to defined instructions rather than free-form prompting, which keeps the output consistent with how PerrfectSite actually writes and works.
Figures, technical claims, and client details are verified by the project’s responsible person. A model does not get the final word on a fact about a client.
Client data does not train models
Client data is never used to train, fine-tune, or improve any AI model. Not by PerrfectSite, and not by the vendors whose tools the team uses.
Tooling is selected and configured accordingly. Business and enterprise tiers with contractual no-training terms are used instead of consumer accounts, and any setting that permits data retention for model improvement is turned off.
A tool that cannot guarantee this does not touch client work.
Sensitive data stays out
Personal data, credentials, financial records, and anything covered by a confidentiality agreement do not go into an AI tool.
Where AI genuinely helps with a task that involves such material, the material is redacted or replaced with representative dummy data first. Debugging a payment flow does not require real card numbers. Structuring a database does not require real customer records.
Secrets are handled by Infisical and passwords by 1Password. Neither ever gets pasted into a prompt. This is written into how the team works, not left to individual judgment in the moment.
When sensitive data has to be processed
Some briefs genuinely require AI-processed data that cannot leave a controlled environment. Regulated financial records, health information, and personal data, subject to strict handling obligations, all fall into that category.
Two conditions must be met before any of it starts.
Written consent from the client. The consent names what data is involved, what processing takes place, which model handles it, and how long the data is retained. Agreement in a call does not count, and a general project scope does not count either. Consent for this is separate, specific, and in writing.
Local models on internal hardware. Processing runs on models hosted on PerrfectSite hardware under direct control. No third-party API and no external inference provider. Nothing leaves the machine the data was loaded onto, and the model runs without network access for the duration of the work.
Data is removed once the task is complete, along with derived artefacts such as embeddings, caches, logs, and intermediate files. Written confirmation of deletion follows on request.
Where local processing cannot produce an acceptable result, that gets said plainly rather than solved by quietly reaching for a hosted model.
A person is accountable for everything delivered
AI output is a draft. Every line of code, every document, and every recommendation that reaches a client has been reviewed by the person whose name is on the work.
Nobody at PerrfectSite ships something they do not understand. Where AI produces a solution the reviewer cannot explain, the solution gets rewritten rather than shipped and hoped over.
Architectural decisions, security judgements, scoping, and pricing stay with people. Those decisions carry consequences that a model does not bear.
Tools first, automation second
The engineering position is straightforward. Effort goes into building tools that perform a task reliably, rather than asking a model to perform that task from scratch every time.
A model asked to do a job produces a slightly different result on every run. A tool built once, calling a model only where judgement is genuinely required, produces the same result every time and can be tested like any other software. Deterministic code handles what deterministic code is good at. AI covers the parts that need language or pattern recognition.
The consequences are practical.
- Output is reproducible, so a review checks the tool rather than re-checking every result by hand.
- Failures are debuggable, because the thing that broke is a named function rather than a prompt that behaved differently on a Tuesday.
- Cost and energy per task drop, since a script does not need to reason its way through an already solved problem.
- The tool becomes an asset that keeps working, instead of a prompt somebody has to remember and retype.
Automation without a tool around it is a demo. The tool is what makes it work on the hundredth run.
Restraint is deliberate
Every AI query consumes energy and water. Running a large model to rename a variable is waste, and the industry does a great deal of it.
The team uses AI where it saves meaningful time and skips it where it does not. Smaller models handle smaller tasks. A well-scoped single request beats twenty rounds of vague prompting, which is the same discipline that makes any project cheaper.
This connects directly to how PerrfectSite approaches development generally. Good decisions made early reduce total work. That principle applies to AI usage as much as to architecture.
People are trained, not replaced
Using these tools well is a skill, and an untrained person with AI produces confident nonsense faster than they used to.
Training covers what the tools do reliably, where they fail quietly, how to verify output, and how to recognise a fabricated answer. Developers are expected to understand the code they accept. Writers are expected to check the facts they publish.
Headcount has not been reduced because of AI. The work capacity per person went up, so projects move faster and cost less.
Transparency
Ask what role AI played in your project and you will get a straight answer.
Clients with policies restricting AI involvement should say so at the scoping stage. Those constraints get documented and followed, and where they change the timeline or price, that is said upfront rather than discovered later.