ERP software with AI: the Skyline Nexus ERP assistant

What the Skyline Nexus ERP AI assistant does

The Skyline Nexus ERP AI assistant is a chat panel on every screen that answers how-do-I questions from the ERP's own help library and returns live business figures, such as sales today or customer dues, under the user's own permissions. It matters because each answer is checked before it is shown, so users get cited help or plain search results, never an unsupported guess.

This guide explains exactly what the assistant does today, how to ask it questions, how answers are verified, which live figures it can return, how permissions limit what each user sees, and how to connect Claude Desktop to Skyline Nexus ERP through the Model Context Protocol (MCP). The last section lists what is being rolled out next. If you are comparing ERP software with AI in general, our buyer's guide on AI ERP for small business sets out the questions to ask any vendor; this page answers them for Skyline Nexus ERP.

What is live today

Six capabilities are live in Skyline Nexus ERP today. Each one is designed to be checkable: help answers carry citations, figures come from the same data as the report screens, and the only thing a user's assistant or Claude Desktop can create is a support ticket.

• An in-app help assistant on every screen that answers how-to questions from the English and Arabic help library and replies in the language you write in
• Answer verification: an answer is shown only if the help page it cites exists and was among the passages retrieved for your question
• Live business figures from a fixed whitelist of eight reports, for today, yesterday, the last 7 days, this month or the last 30 days
• Three guided walkthroughs (ZATCA onboarding, first sale setup and VAT closing) with progress tracking, and a Start my day digest
• An MCP server, the Skyline ERP Assistant, that Claude Desktop and other MCP clients connect to with a personal token issued by an administrator
• An assistant cockpit for administrators that lists unanswered questions and holds AI-drafted help pages for human review before publishing

How to ask the assistant a question

Open the assistant from any screen and type the question the way you would ask a colleague: "How do I create a credit note?", "How do I close the VAT period?" or "Where do I set the approval threshold for journals?". The assistant searches the help library by keyword and by meaning, then answers with a reference to the help page it used, so you can open the page and follow the steps on screen.

The reply follows the language you write in. Ask in Arabic and the answer comes back in Arabic; ask in English and it comes back in English. Teams that search for an ERP con IA, or a system ERP with AI, in another language can write in that language too, but the help pages the answer cites are written in English and Arabic.

If an answer does not solve the problem, rate it and escalate it to a support ticket from the same panel, so the question reaches support without being retyped. To keep usage fair, AI calls are limited per user, by default to 15 a minute.

How answers are verified before you see them

General-purpose AI assistants can produce a confident answer with a reference that does not exist. Skyline Nexus ERP checks every model answer before showing it. The help page the answer cites must exist, and it must be one of the passages that were actually retrieved for that question. If either check fails, you do not see the model's answer at all; you see the plain search results instead, and you choose the page yourself.

This design means the assistant can be wrong only in ways you can see. A cited answer points to a real page that was part of the evidence, and when the evidence is missing the assistant falls back to search rather than improvising. It does not replace reading the help page for a task you have not done before, but it removes the invented-reference problem. For more on why grounded answers matter when you choose ERP software with AI, see our guide on AI in accounting.

Live business figures from eight reports

Ask "sales today", "purchases this month" or "compare sales by branch over the last 7 days" and the assistant returns the real figure. The model never writes a database query. At most it decides which of eight whitelisted reports you asked for and for which period; the figure itself comes from the same read paths as the dashboard and report screens.

Sales, purchases and expenses can also be compared per branch and period over period. For example, a manager asks for sales by branch over the last 7 days and sees EUR 18,400 for one branch against EUR 16,900 for the previous 7 days, an increase of EUR 1,500, or 8.9 percent. The eight reports are:

• Total sales for a period
• Sales trend
• Purchases for a period
• Expenses for a period
• Customer dues
• Supplier dues
• Stock alerts
• Top products

Worked example: a figure checked before it is used

An owner preparing for a meeting with the bank asks the assistant "How much do we owe?". The assistant returns the supplier dues figure, because supplier dues is the closest of the eight reports to that question. The owner's finance manager reviews it before it goes into the bank pack and rejects it for that purpose: the bank asked for total liabilities, and supplier dues exclude VAT payable, accrued salaries and the loan. The manager runs the Balance Sheet instead and uses the assistant's figure only for the line it actually measures.

The same review habit applies to every live figure. The assistant answers the question it was asked from one of eight reports; a person decides whether that report answers the business question. Before quoting any figure outside the company, compare it with the full report screen for the same date, for example the Supplier Aging Report for supplier dues or the Customer Aging Report for customer dues.

• Question asked: How much do we owe?
• Assistant output: the supplier dues figure from the supplier dues report
• Reviewer checks: does the report measure what the bank asked for?
• Rejected for the bank pack: supplier dues are not total liabilities
• Used instead: the Balance Sheet; the supplier dues figure is only compared with its trade payables line, and any difference is explained

Permissions: the assistant sees what the user sees

Live figures are scoped to the user's business and to the locations the user is allowed to access, and each figure is gated by the same permission as its report screen. A branch manager whose user record grants access to one location sees that location's sales, not the whole company's. A cashier without permission to view the underlying report does not get the figure from the assistant either.

This is why roles matter more than the assistant itself. Set up roles and permissions under User Management, grant access locations per user, and the assistant inherits those limits automatically. For how to design roles that keep preparers and approvers apart, see our guide on roles, permissions and segregation of duties in Skyline Nexus ERP.

Guided walkthroughs and the Start my day digest

Three walkthroughs guide users through ZATCA onboarding, setting up the first sale and VAT closing, step by step, with progress tracking so a user can stop and resume. The Start my day digest gives a daily starting point. Both are deterministic: they are fixed step lists and summaries, not AI-generated text, so every user sees the same steps every time.

Use the walkthroughs for onboarding new staff and the digest as a daily routine, and use the chat assistant for the questions that arise along the way.

Connecting Claude Desktop through MCP

The Model Context Protocol is an open-source standard for connecting AI applications to external systems such as tools and data sources. Skyline Nexus ERP runs an MCP server called the Skyline ERP Assistant. Claude Desktop, or another MCP client, connects to it with a personal access token that an administrator issues to each user; tokens are OAuth bearer tokens and each is limited to 60 calls a minute. Whoever holds the token acts as that user, with that user's permissions, so treat it like a password.

The server offers four tools and nothing else can be written through it. Anthropic's help centre notes, as of September 2026, that you should connect Claude only to servers built and hosted by organisations you trust, because a malicious MCP server may include hidden instructions that try to make Claude perform unintended actions; keep that in mind before you add other servers alongside this one.

To connect, ask your administrator for a personal token and the server address, add the server to Claude Desktop or your MCP client with that address and token by following the configuration your administrator provides, and tell your administrator when you change role or leave, so your token can be reviewed. The four tools are:

• Help search: returns help passages with page citations
• Help page: fetches the full help page
• Business metric: the same eight live figures, read-only, permission-checked and limited to your own business
• Create support ticket: the only tool that writes

For administrators: tokens, the cockpit and the model

Administrators with the assistant.admin permission open the assistant cockpit. It shows the questions the assistant could not answer, known as gaps, so the help library can be extended where users actually get stuck. AI-drafted help pages wait in the cockpit for human review; nothing is published until a person approves it.

Administrators also mint personal MCP tokens per user, and should keep a simple register of who holds one and why. The assistant's model provider is set in the platform configuration with an ordered failover list, and Anthropic's Claude is the default. Keep the assistant's role in your organisation's records of processing under the GDPR, and include it in your normal user access reviews; this is general information, not legal advice.

What is being rolled out next

The next steps on the Skyline Nexus ERP roadmap extend the assistant from answering to preparing work for approval, always with a person approving before anything reaches the ledger. These are being rolled out; tell us which matter to you and we will confirm your go-live date.

Until then, the controls that approval workflows rely on are already in place: the approval threshold for journals and the Audit Trail with old and new values. Our AI accounting automation library marks every recipe as Live today or Rolling out, so you can see precisely where each task stands.

• Claude drafting journals, bills and bank matches for human approval
• Invoice and statement capture from PDFs
• AI close and audit analytics
• Ledger and trial balance reads through the MCP server
• A general-ledger API
• Bring your own Anthropic API key
• Scheduled AI automations