Why this case matters
- AI the system is accountable for, not “the model's mood”. Two models check each other, every call has a known cost, and if a provider fails, a backup model takes over automatically. The client knows the cost of every order; the data and the process belong to the client, not to the AI vendor.
- Sensitive data, taken seriously. Documents with special categories of personal data never leave the client's server in plain form. The architecture was designed from GDPR and ISO 17100 requirements up, not retrofitted to them after the fact.
- A deep dive into the industry. We learned how the agency works down to the details, from industry discount grids and the four-eyes principle to government requirements for electronic signatures. That's how we approach any industry: process first, code second.
- A custom system vs. subscriptions. Its own pricing rules, client data on its own hardware, independence from the AI vendor, and no per-seat fees that grow with every new hire.
- Speed without cutting corners. The client saw the first working screens within a few weeks, and the core went live in 2–3 months. Every release runs 211 automated tests first, so updates don't break the agency's work.
The same architecture for any regulated data
Swap GDPR for your own data protection rules, and the problem will feel familiar to anyone who handles medical records, contracts, court documents or HR data. No personal data goes to the cloud, only masks, so there's no cross-border transfer of personal data and no processing of it by a third-party AI model. The setup was built for GDPR, where fines reach 4% of annual turnover. This approach opens up powerful large language models (LLMs) to fields where they used to be off the table: healthcare, law, finance, HR. This is exactly the kind of setup we design as part of our AI integration service, and the platform itself follows the playbook of custom SaaS development.
FAQ
How long does it take to build a platform for a translation agency?
The core (stage pipeline, roles, translator workspace) takes 2–3 months to go live. After that, the platform grows in one-week iterations with live releases: the client sees progress in working screens, not in reports.
Does the data go to OpenAI or Anthropic?
Only anonymized text: names, numbers and addresses are replaced with masks on the client's server before any call to the cloud. The mapping table is encrypted and never sent out. So there's no cross-border transfer of personal data and no processing of it by a third-party AI model, which takes care of the main data protection concern. There's also a mode with no AI at all and a “cloud stores nothing” switch.
How much does a platform like this cost?
The price depends on the number of seats, how deep the AI layer goes and the data protection requirements, so there's no honest one-size-fits-all price list. We work out a preliminary range for free on the first call: you walk us through your process and leave with an architecture sketch and a ballpark budget.
Who owns the code, and who supports the system after launch?
The source code, data and documentation belong to the client: under the contract we hand over the entire repository. Support is optional: our team can keep developing the system in iterations or hand it over to your developers.
Why not an off-the-shelf SaaS like memoQ, Phrase or Smartcat?
If your process is standard, a ready-made service is faster and cheaper, and we'll tell you so on the first call. A custom platform wins when data ownership, non-standard pricing rules, public-sector requirements and AI under your own control all matter. Here, all four did.
Need a system like this for your process?
We design and build industry platforms with an AI layer end to end, from process analysis to launch and support. Leave a request below: on a 30-minute call we'll show you a live Polyglot demo and sketch the architecture and a budget range for your task for free. If an off-the-shelf SaaS covers your process, we'll tell you so plainly.
