Insights · 6 min read
Why business agents need local business data, not translated prompts
Translating the words is easy. Knowing how each country does business is the job.
Most AI training effort so far has gone into code. But the work companies want to hand to agents is mostly operations: paying suppliers, closing the books, answering customers, chasing invoices and keeping deals moving. That work runs differently in every country.
Same task, three markets
Take one of the most ordinary jobs in any company: handling a supplier invoice. The steps look the same everywhere. The rules underneath do not.
| Country | How the invoice arrives | What the agent has to know |
|---|---|---|
| Saudi Arabia | Electronic, under ZATCA rules | Since 4 December 2021, invoices must be generated and stored through compliant electronic solutions. From 1 January 2023, taxpayers integrate their e-invoicing with ZATCA in waves. |
| United Arab Emirates | Electronic, through accredited service providers | Under Ministerial Decisions 243 and 244 of 2025, B2B and B2G e-invoicing becomes mandatory in phases, starting 1 January 2027 for businesses with revenue of AED 50 million or more. |
| Germany | Structured e-invoice | Since January 2025, businesses must be able to receive e-invoices that follow EN 16931. Paper invoices are phased out from 2027 for companies with prior-year turnover over €800,000, and from 2028 for everyone else. |
An agent that learned accounts payable from English prompts translated into Arabic will write fluent Arabic and still pay the wrong amount, request the wrong document or skip a step the tax authority requires.
Why translation isn’t enough
- Documents: every market has its own invoice formats, required fields and tax codes. A translated prompt carries none of them.
- Rules: what makes an invoice valid, deductible or payable is set country by country, and it changes.
- Register: a Gulf customer reply, an Egyptian supplier email and a German payment reminder each follow their own conventions.
- Cross-border work: real work crosses languages. An instruction in Arabic, an invoice from Germany, a reply in German.
Scenarios with outcomes, not questions with answers
A question and answer pair tests whether a model knows a fact. A business scenario tests whether an agent can do the job. We build each one around a final state that can be checked: the right amount scheduled for payment, the right correction requested, the right message sent in the right language.
Our samples page includes a fully worked, clearly labeled synthetic example: a Riyadh distributor, a German supplier, three documents that don’t agree, and a grader that checks what the agent actually did.
How we make them
Our Guild writes every scenario in Arabic, formal and dialect, and the English, French and German the region trades in, by professionals who do that work in the region. Each one is built on the documents and rules of its market, reviewed by a second specialist and graded automatically. We deliver in batches, on the schedule a lab sets.
Questions we get
- Does Bearish AI only make coding data?
- No. Our engineering archive from Bearish OS is our founding dataset. The Guild makes business scenarios across finance, accounting, procurement, logistics, sales, customer service and administration.
- Which languages do you cover?
- Arabic, formal and dialect, and the English, French and German the region trades in. Every scenario is written natively by a professional in that market, never translated.
See the format for yourself.