In short: a secure AI assistant can answer staff questions from your ERP, CRM, SQL databases and SharePoint in seconds, in Arabic or English, without moving your data out of your environment. A focused pilot takes about eight weeks, with a clear go / no-go decision at the end.
The scenario
This blueprint describes how FOXMINDS would plan and build an AI knowledge assistant for a typical 200-person company in Lebanon or the Gulf, such as a distributor, a services firm or a group with several branches. It is a worked example of our method, not a client case.
- An ERP for orders, stock and invoicing, and a CRM for customers and opportunities
- SQL Server databases behind reporting and in-house applications
- Policies, procedures, contracts and proposals in SharePoint and shared folders
- Microsoft 365 and Teams for email and daily collaboration
- Staff who work in both Arabic and English
The problem is familiar: the answers exist, but they are spread across four systems and a spreadsheet. A manager who needs a number waits a day for someone to pull it together, and staff spend time searching for documents instead of using them.
What the assistant answers
The assistant lives inside Microsoft Teams. Staff type a question in plain language, in Arabic or English, and get an answer with links to the sources it used. These are the kinds of questions a pilot is designed around:
| Team | Example question | Where the answer comes from |
|---|---|---|
| Sales | Which customers bought less this quarter than last, and by how much? | CRM + ERP |
| Sales | Draft a proposal for this client using our three most similar past offers. | SharePoint + CRM |
| Operations | Which orders are late this week, and why? | ERP + SQL |
| Operations | What does our procedure say about returning damaged goods? | SharePoint |
| Finance | Which invoices over $10,000 are more than 60 days overdue? | ERP |
| HR | ما هي سياسة الإجازات المرضية لدينا؟ (What is our sick-leave policy?) | SharePoint |
| Management | Summarize yesterday’s sales, stock alerts and open complaints in five lines. | ERP + CRM + SQL |
Each answer shows its sources, so users can check the original record or document in one click. If the assistant does not find a reliable answer, it says so instead of guessing.
How it is built
Nothing is replaced. Your systems keep running as they do today, and the assistant sits on top of them in four layers:
- Secure, read-only connectors to the approved data sources: ERP, CRM, SQL databases and SharePoint libraries. We start with the minimum data each use case needs.
- A searchable index of the approved content, refreshed on a schedule. It keeps each record’s permissions, so the assistant knows who is allowed to see what.
- The AI model, running in your own cloud tenant (or a dedicated environment we manage for you, in an in-country region where available). It reads the relevant records and documents and writes the answer.
- The chat interface in Microsoft Teams, where staff already work, with sources cited under every answer.
Arabic and English
Staff can ask in Arabic, English or a mix of both, and the assistant answers in the language of the question. It can search an English procedure to answer an Arabic question, and the reverse. During the pilot we test both languages with real questions from your teams, including local terms and product names, and tune the results before rollout.
Security and governance
For most companies, security is the deciding question. These controls are part of the design from day one, not added later:
- Your data stays in your environment. Content is processed in your tenant or a dedicated environment, and is not used to train public AI models.
- Permissions mirror your systems. If an employee cannot open a file or a record today, the assistant will not show it to them.
- Read-only first. The pilot only reads data. Any future action, such as creating a ticket or sending a reminder, needs a person to approve it.
- Full audit trail. Every question, answer and source is logged and can be reviewed by IT and compliance.
- Sensitive data stays out. Salaries, personal files and other restricted areas are excluded unless you decide otherwise.
- Local law. Controls are designed around Lebanon’s Law No. 81/2018 and the UAE Personal Data Protection Law, working alongside your legal advisers.
Fix permissions first. An AI assistant is only as safe as the permissions underneath it. If a confidential folder is shared with “everyone” today, the assistant will respect that setting. That is why discovery includes a permissions review before any data is connected.
The 8-week pilot plan
| When | Phase | What happens |
|---|---|---|
| Weeks 1–2 | Discovery | Workshops with each department, pick 2–3 use cases, map the data sources and who owns them, agree success measures and baseline |
| Weeks 3–6 | Build the pilot | Read-only connections to 2–3 sources, build the searchable index, set permissions, connect to Microsoft Teams, test in Arabic and English |
| Weeks 7–8 | Test and measure | 10–15 pilot users ask real questions; we track accuracy, time saved and adoption, then fix what users report |
| End of week 8 | Go / no-go | A short report with results against the baseline, and a costed plan to add users, sources and use cases, or to stop |
The pilot has a fixed scope and a clear decision at the end. You scale only once the results are proven against your own baseline.
What drives the cost
There is no one-size price. The cost depends on a few factors, which we review with you in discovery before giving you a clear estimate:
- Number of users and how often they ask questions
- Number of data sources and how much content is indexed
- Where it runs: your existing cloud tenant, an in-country region or a dedicated environment
- Languages and use cases included in the pilot
Ask us for an estimate for your company
What to measure
We agree the measures before the pilot starts, measure today’s baseline, and report against it at the end. Typical targets for a pilot like this one:
- Time to answer: from hours or a day to under a minute for common questions
- Time on recurring reports: a large share of manual preparation removed
- Answer quality: the share of answers pilot users rate as correct and useful
- Adoption: how many pilot users ask at least one question per week
These are targets, not past results. Your baseline and targets are set together in discovery.
Common mistakes to avoid
- Connecting everything at once. Start with 2–3 sources and a few clear questions; add more once the first ones work well.
- Ignoring permissions. Over-shared folders become over-shared answers. Review them first.
- No data owner. Each source needs someone who can say what is correct and keep it up to date.
- Testing in one language only. If your teams work in Arabic and English, test both from the first week.
- Skipping training. A 30-minute session on how to ask good questions makes a visible difference to adoption.
Readiness checklist
If you can answer “yes” to most of these, you are ready for a pilot:
- We can name 2–3 questions that take too long to answer today
- We know which systems hold the answers, and who owns each one
- Our users sign in with Microsoft 365 accounts
- Our SharePoint and file permissions are reasonably up to date
- We know which data must stay out (for example HR or salaries)
- A manager will sponsor the pilot and give feedback every week
- 10–15 staff are willing to test it for two weeks