Investor relations site and AI assistant

Investors ask in Arabic and in English, by text and by voice. For them a guessed answer is worse than no answer.

I treat a wrong answer here as worse than none, so I built the assistant around a fact sheet and a question bank, not around the model. Nearly every caller is Arabic, so I started there. I test changes on a separate test assistant at no cost, replay real calls to see what went wrong, and copy a change to the live one only after reading the bank's results.

• The public investor relations site.
• An access-controlled dashboard behind it.
• An assistant that answers investors in Arabic and English, by text and by voice.

• A financial line got the figure of a similar-sounding line.

  Retrieved excerpts sometimes handed back the number for a neighbour, such as an amount due to the company for one due from it. A prompt rule didn't stop it. I moved the lines callers ask for into an always-loaded fact sheet, each with its Arabic name: 147 of 148 statement answers came back right.

• Spoken answers stopped mid-phrase.

  The model's hidden thinking counted against the reply length cap, so 8 of 44 turns in two real calls hit the limit. I turned thinking off, raised the cap, and set the clock to Riyadh time. I also rewrote the answers to be short and to lead with the latest period.

• The same date read differently on different phones.

  A plain Arabic locale gives Hijri dates on most phones and Gregorian in my tests, so the assistant was told different dates for the same moment. A caller asked for Gregorian. I pinned the calendar in code and added a rule. It only showed up when I checked on real devices.

• The site is public, and I run it.
• The assistant answers from an approved knowledge bank, never from guesswork.

• Enterprise CRM and Sales Hub

  The group's CRM portal. Staff in 60+ branches across seven countries work in it every day.