At Agentforce World Tour Dubai, held at Dubai Exhibition Centre on 14 October 2026, our CEO, Mohamed Abdelsalam, sat down with Richard Doyle, Head of Middle East Ecosystems at Salesforce, to talk about what it takes to get AI agents working in the Middle East. The conversation came back to three points: agents here have to work in Arabic, people decide whether an agent succeeds, and the first release should be small. Watch it below, then read how we put each point into practice.
The short version
- Arabic is required. The region mixes Arabic, English, dialect and slang. An agent that only works in English is not built for the UAE, Saudi Arabia, Qatar, Oman or Bahrain.
- Change management makes or breaks it. People need to know how to use the agent, why it is there, and what is in it for them.
- Skip the big bang. A 12-month rollout assumes nothing changes in 12 months. Start small with your own data, find a quick win, get a prototype live in 30 to 60 days, then scale.
1. Arabic is required
Customers in the Gulf move between Arabic and English inside one conversation, often inside one sentence. Salesforce made Agentforce Service Agent and Employee Agent available in Arabic in November 2025, with more agents to follow. That makes Arabic possible. Your tests decide whether it works. Before launch, check six things:
- Real messages. Build the test set from your own chats and emails: Gulf and other dialects, Modern Standard Arabic, Arabic typed in Latin letters, and English product names inside Arabic sentences.
- The language list. An agent’s default and secondary languages are the full list of languages it will speak. Test what it says when a customer writes in a language outside that list.
- Arabic knowledge. Agentforce filters out knowledge articles whose language does not match the conversation, so Arabic answers need Arabic articles.
- Switching mid-conversation. The agent replies in the language of the customer’s latest message, while the session’s language setting stays fixed from the start. Switch language halfway through a test and check the answers and the actions.
- Custom actions and names. Actions that call an external service or a direct AI call need their language set by hand. Brand, product and place names should stay untranslated, and Salesforce says its suggested workaround may give mixed results, so test every name on your list.
- Handover. The person who takes over an Arabic conversation has to read Arabic, and the transcript has to go with it.
2. Change management makes or breaks it
An agent can pass every test and still fail if the people around it ignore it or work around it. For a first agent, a change plan that answers the how, the why and what is in it for them fits on one page:
- An owner and champions. Name a business owner, and a champion in each team who uses the agent first and answers colleagues’ questions.
- The why, written down. One paragraph on what the agent does, what it never does and who to call, shared before launch in Arabic and English. The agent charter gives you most of it.
- The benefit, role by role. Name the work the agent takes off each role’s desk, measure it from the first week, and share what you find.
- Training on real cases. Short sessions by role, in the team’s working language, using the cases they handle every week.
- A feedback loop. One place to flag a wrong answer, reviewed weekly, with fixes reported back to the people who raised them.
- Targets that fit. If a team is measured on cases closed, agree how cases the agent resolves will count, so nobody loses out by using it.
3. Skip the big bang
The products, the models and your own priorities will all move within a year, so a 12-month plan is out of date before it finishes. A first release that goes live in 30 to 60 days contains:
- One use case with one owner, chosen because its data already sits in Salesforce, such as order status questions or appointment changes.
- One team and one or two channels, in Arabic and English.
- The records and knowledge articles that use case needs, cleaned for that slice only.
- Answers first. Actions that change records come later, one at a time, after their tests pass.
- A test set in both languages, run before go-live and after every change. Agentforce Testing Center runs synthetic conversations in a sandbox for this.
- A handover to a person, a baseline measured before launch, and a date to decide whether to scale it, change it or stop it.
Leave out the integrations the use case does not need, extra channels and the long list of actions. Add them once the first release has proved itself, and scale what works.
Where to start
Our Agentforce Readiness Assessment picks your first three use cases, checks the data and permissions behind them, and gives you a costed pilot plan in ten working days. The agent charter sets out who owns each agent and how it is tested and changed. And our Agentforce team builds and runs agents in Arabic and English. Thank you to Richard Doyle for the conversation.
Frequently asked questions
Does Agentforce support Arabic?
Salesforce made Agentforce Service Agent and Employee Agent available in Arabic in November 2025, with more agents to follow. Before launch, test with your own customers’ dialects, with Arabic and English in the same message, and with Arabic knowledge articles.
How long should a first Agentforce release take?
We aim to get a prototype live in 30 to 60 days: one use case, one team, your own data, tests in Arabic and English, and a handover to a person. Scale it once it has proved itself.
What should a change plan for AI agents include?
A named owner and champions, a written note on what the agent does and never does, the benefit for each role, training on real cases, a weekly feedback loop, and targets that do not penalize people for using the agent.
