Most companies have now run an AI pilot. Far fewer have anything in production.
The gap is rarely the model. It is that the agent was never wired to the systems the work actually lives in, nobody agreed what it is allowed to decide on its own, and no one owns it once the person who built it moves on. A demo that answers questions is not automation. Automation is when a task that used to occupy someone every Tuesday stops occupying them.
We build the second kind, and we have run our own portfolio this way long enough to know which parts are harder than they look.
What we do
| Automation audit We map where your work actually goes, and tell you which parts pay back when automated, which are better left alone, and which model fits each one. Written findings, ranked by return, with the cost of running each at your volume. | From US$4,000 |
| Agent build One workflow, end to end: connected to the systems the work lives in, with the limits on what it may decide, the handover to a person when it should not, and logging you can audit afterwards. | From US$12,000 |
| Sites built with AI Corporate sites produced in days rather than months, and built the way a search engine and an AI assistant can both read them. We have built several dozen this way. | From US$6,000 |
| Multilingual rollout Your existing site in any language your customers actually use, written in that language’s own script rather than transliterated, each one at its own address so it can rank, with a human pass on anything that carries legal or financial weight. | From US$4,500 |
Fixed fees, agreed before we start. No retainers.
Whichever model you use
We are not a reseller for anyone. Claude, ChatGPT, Gemini, Grok, DeepSeek, Kimi, Mistral, Llama, Qwen — or a model you host yourself, which for some work is the only answer your compliance team will accept.
That independence is the point. The right model for summarising a contract is rarely the right one for classifying ten thousand records an hour, and the price difference between them can be two orders of magnitude. A provider who only sells one will recommend that one. We cost the options at your volume and tell you which is cheapest for each job, including where a smaller open model does the work for a fraction of the price.
We also build so you can change your mind. Providers change terms, raise prices and deprecate models, and an integration wired to one vendor's exact interface becomes a rebuild when they do.
Websites, built and translated with AI
We have built several dozen sites this way — our own and clients' — and the same method applies to localisation in any language.
Machine translation alone is not enough for a corporate site, because the failures are rarely grammatical. They are a currency symbol in the wrong position, a date that reads as a different day, a right-to-left layout that breaks where Latin text is embedded, a form that rejects an address format, and a legal line that no longer says what your lawyers approved. We build the translation into the site structure rather than bolting it on: each language at its own address, correct hreflang, local formatting, and a human review pass on anything legal or financial.
Live work you can read today includes a payments site in 25 languages, a service in 14 with full right-to-left support for Arabic, Persian and Hebrew, and an education site in 21 — covering Chinese, Japanese, Korean, Thai, Hindi, Bengali, Greek, Cyrillic and most of Western Europe.
Every language in its own script, not in translation-ese. Right-to-left for Arabic, Persian, Hebrew and Urdu. Chinese in both simplified and traditional, Japanese and Korean. Thai, Khmer, Lao, Burmese and Vietnamese. Devanagari, Bengali, Tamil, Telugu, Gujarati and Sinhala. Cyrillic and Greek. Turkish, Indonesian, Malay and Tagalog. Swahili, Hausa, Amharic, Yoruba, Igbo, Zulu, Afrikaans and Somali. Portuguese, Spanish, French, German, Italian, Dutch, Polish and the Nordic languages. A visitor in Jakarta or Addis Ababa should not be able to tell your site was written anywhere else.
Where we have shipped a language already we will show you the live pages. Where we have not — most of the African set, and the scripts with their own rendering rules such as Amharic and Khmer — we say so while we are scoping it, and price the extra script work in rather than discovering it late.
What we don't promise
We will not tell you AI replaces your team. In most of the work we have seen it removes a specific recurring task, which is worth real money and is not the same claim.
We will not automate a process that is broken. Automating it makes it produce wrong answers faster, and the audit will say so before you have spent anything on building.
We do not guarantee an accuracy figure before measuring your data. Anyone who quotes you one without seeing it is guessing.
How it works
- A short call, free. You describe the work you want to stop doing by hand. We tell you whether it suits automation, and what it costs. If it does not, we say so.
- Access. Whatever systems the work runs through, under your own access rules.
- Written findings within five business days, ranked by return, with the fee fixed.
- The build, then a period running alongside your team before anything is switched over.