Tawasol
Better conversations, measured: our first AI support benchmark
How does Wasel handle real support questions? Explore our first Egyptian Arabic and English benchmark: conversation quality, response times and estimated costs in EGP.
A customer asks about a course, follows up on its price, then wants to speak to your team. Good AI support needs to follow that whole conversation.
We tested Wasel, Tawasol’s AI support assistant, on everyday questions in Egyptian Arabic and English. Here is what worked, where the answers could be clearer, and what the test tells us about speed and cost.
A first look at the results
20 scenarios. Three runs each. 60 conversations and 108 replies. The test covered a training academy, an online store and a home-services business, using fictional business information and real AI replies on 28 October 2026.
- 47Pass
- 10Pass with improvement
- 3Needs correction
| Business | Pass | Improve | Correct |
|---|---|---|---|
| Training academy | 30 | 6 | 0 |
| Online store | 6 | 4 | 2 |
| Home services | 11 | 0 | 1 |
Pass means the assistant completed the required task correctly. Pass with improvement means the result was acceptable in context, but the reply could be clearer or more helpful. Needs correction means an important answer was missing or incorrect.
The business groups contain different numbers of conversations. Each bar shows the share within its group; the numbers show the actual counts. These are results for this small test set, not an overall accuracy score.
Building smart, affordable ACX
Our goal is smart, affordable Agentic Customer Experience (ACX) for Egyptian and Arabic-speaking teams. That means helping customers with your business information, following the conversation and bringing in your team when needed.
Sharing measured results is part of that goal. You should be able to see both useful answers and areas that still need work before trying AI support for your business.
Following the conversation
One academy conversation started with a beginner asking which database course would help them at work. Wasel recommended SQL Foundations and explained why it suited a beginner.
The customer then asked:
طيب اختياري ده مدته كام وسعره كام؟
Wasel replied:
كورس SQL Foundations مدته 20 ساعة تدريبية، وسعره 1800 جنيه.
The assistant kept track of the selected course and gave its correct duration and price from the supplied information. These course details belong to the fictional test business.
All six requests for team help were also recorded for review in the test. That checks the assistant’s handoff decision; it does not measure how quickly a person follows up.
Where the answers can improve
A useful outcome can still leave room for a better reply. We separate those improvements from mistakes that leave the customer’s task unresolved.
10 conversations with improvement notes; 3 needing correction. Each appears once below.
- Explain that contact details are optionalConversations: 6Pass with improvement
- Make refund follow-ups more helpfulConversations: 4Pass with improvement
- Use the available return conditionsConversations: 2Needs correction
- Answer the service question directlyConversations: 1Needs correction
For example, after an optional invitation to share contact details, a customer asked whether a phone number was required. Wasel confirmed that the request was awaiting review. A direct “No, sharing your number is optional” would have been more helpful.
The three conversations needing correction had clearer gaps: two replies treated available return conditions as unknown, and one asked an unnecessary clarification after the customer had already named the service. Those answers need to use the available information more reliably.
How quickly did Wasel reply?
- Median
- 10.68 seconds
- 95th percentile
- 15.66 seconds
| Reply time | Replies |
|---|---|
| Under 5 s | 5 |
| 5–<10 s | 40 |
| 10–<15 s | 55 |
| 15 s or more | 8 |
The median reply took 10.68 seconds. The 95th-percentile reply time was 15.66 seconds: at least 95% of the measured replies were at or below that time.
These timings cover the test assistant’s processing. They exclude delivery through a live messaging channel and are not a response-time guarantee.
Estimated AI credits in EGP
The 108 replies used an estimated 19.47 EGP in test credits, averaging about 0.18 EGP per attempted reply. The estimate includes the conversations that needed improvement or correction.
This is a limited test estimate. It excludes preparing the knowledge base, live knowledge search, hosting and channel delivery or external fees. It does not establish how many production replies 1,000 EGP will cover. Check our pricing page for plan details and your workspace usage for actual consumption.
How to read this benchmark
We used the same business information for each repeat and checked facts, conversation context, task completion and requests for team help. Answers were assessed by their meaning, not by matching a fixed phrase.
This is a small first-party test with simulated customer scenarios and automated assessment. Independent human review of the Arabic results is still pending. It is not an independent audit, a competitor comparison or a test of live message delivery. This edition covers Egyptian Arabic and English; it does not establish performance across every Arabic dialect.
Common questions
Will my business get the same results?
Results depend on your knowledge, policies and setup. Start with your common customer questions, then add follow-ups and requests that need your team.
Does this measure customer resolution?
It measures whether the assistant completed the task in each test conversation. It does not measure live customer satisfaction, human follow-up or production resolution rates.
How can I try Wasel with my own information?
Use our knowledge and testing guide to prepare your information and test realistic conversations. Explore knowledge-grounded replies, or start with Tawasol.