Talk to your ERP. Get answers and actions.
A conversational AI assistant embedded inside Microsoft Dynamics 365 Business Central. Read live BC data, run analytics, create records — all in plain English. Every write previews first; nothing posts without your “yes”. Honours your BC permissions, by design.
- Native chat panel — any BC page
- Preview-and-Confirm on every write
- Provider-neutral: Claude · GPT · Gemini
- Page-context AI on filtered lists
Built for the Microsoft stack
The Business Central assistant that knows where the ‘commit’ line is.
Most AI in ERP today is a chat widget that paraphrases your data and hopes you don't notice. Ours reads live BC data, runs the analytics, executes the action — and stops before the write so a human always gets the final word.
Preview-and-Confirm by default
Every write is a two-step pattern: AI shows the preview card, user replies 'yes' to commit. No autopilot, no surprises. Audit log captures the prompt, the preview, the confirmation and the result — for every single write.
Honours your BC permissions
The assistant operates within the signed-in user's BC permission set. Can only read what they can read; can only write to tables they can write. Cannot bypass RBAC. Your existing security model still works — the assistant composes with it.
Provider-neutral, vendor-hidden
Choose Claude, GPT or Gemini at install time — switch later with a Setup change. End users see only the 'SourceForge AI' brand. Provider errors, references, and product names are sanitised before reaching the user.
Talk to Business Central. Get the answer — and the action.
From any BC page, open the chat panel. Ask in plain English. The assistant queries your live BC data, runs the analytics, and — when you ask for a change — previews exactly what it will write and waits for your “yes”.
Every write previews first. The assistant cannot bypass your BC permission sets — it can only read and write what you can.
Every page, every record — at the speed of a sentence.
Replace multi-step navigation with a single English sentence. Reads compose freely — “top 10 customers, only the ones over their credit limit, sorted by oldest invoice” works.
The assistant creates records. You still hit ‘commit’.
Sales orders, invoices, credit memos, purchase orders, items, contacts, customer master data — all creatable from a single sentence. Every write is a two-step pattern: the AI builds a preview card with everything it's about to do, and waits for your explicit confirmation.
- 1You: “Create a sales order for Mednovation, 5 ATHENS desks at ₹12,000 each.”
- 2AI shows preview: customer, items, taxes, totals — and waits.
- 3You: “yes” (or describe changes). Only then is the record committed.
Six forecasting models. One prompt away.
No BI tool to set up, no SQL to write, no dashboard to design. The assistant runs these directly against your live BC ledger entries.
Sales forecasting
Linear-regression forecast with quarterly seasonality. Returns historical actuals + forward-looking projection with confidence ranges — ready to paste into a board deck.
Stockout forecasting
Predicts when each item runs out using consumption rate, on-hand, open POs and committed sales orders. Categorised CRITICAL / URGENT / WARNING / MONITOR.
Customer churn forecast
Compares recent vs earlier-half buying activity for every active customer. Flags decline; sorts worst drop first; ranks by revenue at risk.
Cash position forecast
Weekly cash projection using open AR and AP bucketed by due date. Returns inflows, outflows and cumulative net for the chosen horizon.
Sales anomaly detection
Z-score analysis on daily sales vs rolling baseline. Surfaces unusual spikes (opportunities) and drops (issues) for the chosen window.
Pivot tables (any dimension)
2-D pivot over Customer Ledger Entries — rows × columns × measure. Outputs a clean markdown table with row + column totals.
Eight surfaces where AI meets your team.
The chat panel is the headline; the supporting surfaces are where habit forms. Morning Briefing, Smart Alerts, Page-Context AI, Knowledge Library, My Insights — designed to fit the rhythm of a working day.
Native AL. Native BC permissions. No middleware.
A production AL extension that respects BC's security model, runs entirely inside your tenant, and never bulk-syncs your data to anyone. Calls to your chosen AI provider are outbound from BC, scoped to the current prompt.
No third-party middleware to licence, monitor, audit, or budget for. The provider's API key sits in BC Isolated Storage and is never logged or exported.
- BC Online (SaaS) + On-Premises v25+
- Platform 25.0+ · Runtime 14.0+
- Compatible with W1 + localisations
- Native chat panel from any BC page
- Page-context AI (filtered list grounding)
- Honours standard BC permission sets
- Anthropic Claude (recommended)
- OpenAI GPT
- Google Gemini
- Tier 2+ paid plans for production
- Vendor-neutral end-user branding
- Provider switch is a setup change
- Two-step Preview-and-Confirm on writes
- Honours user's BC permission set
- Error sanitisation (vendor names hidden)
- Audit log: who / what / when / before / after
- API key in Isolated Storage
- [NonDebuggable] key getter
- Usage dashboard (tokens · queries · cost)
- System Prompt Editor
- Prompt Templates (1-click standardisation)
- Knowledge Library (curated answers)
- Briefing schedule (auto-emails)
- Feedback review for tuning
Multi-language by default
The assistant understands and responds in any language the underlying model supports — English, Hindi, Bengali, Tamil, German, French, Arabic, Japanese, Spanish and 100+ more. Users can switch mid-conversation; a single thread can read a German invoice and answer in Hindi.
Vendor-neutral user experience
End users only ever see the “SourceForge AI” brand. Error messages, branding and product references from your chosen AI provider (Claude / OpenAI / Gemini) are all sanitised before reaching the user. Switching providers is invisible to your team.
Anywhere BC runs, this fits.
The capabilities are vertical-agnostic — what changes is the prompt cookbook. Manufacturing teams ask about BOMs and POs; retail asks about top items and stockouts; finance asks about cash position and DSO.
Five phases. 6–10 weeks. No surprises.
Most of the elapsed time is pilot adoption and prompt cookbook development — not technical configuration. The extension installs in 2–3 days.
A 2–4 week pilot with 3–5 power users builds the prompt cookbook the rest of your team will live by.
Discovery
Confirm prerequisites, agree pilot user list, choose AI provider, set API consumption budget.
Sandbox install
Deploy to BC sandbox, configure API key (Isolated Storage), assign permission sets, smoke-test.
Pilot
3–5 power users live daily for 2–4 weeks. Feedback loop, prompt cookbook v1.
Production rollout
Deploy to production, broader user enablement, training sessions, dashboard configuration.
Hypercare
Two weeks of daily check-ins, fast feedback loop, prompt refinement, then SLA-based support.
Ongoing
Quarterly reviews, system prompt tuning, knowledge library updates, new feature rollouts.
Audit-grade by construction.
API key in Isolated Storage
`DataScope::Module`, AES at rest, `[NonDebuggable]` getter. Never visible to other apps; never written to tables, exports or logs.
Audit log on every write
User, timestamp, prompt, preview, confirmation, before/after values. Immutable; survives BC upgrades.
Error sanitisation
Raw provider errors never reach users — vendor names, endpoints and stack traces are filtered. Unified branding regardless of provider.
Two BC stories — placeholders till the real ones land.
Real customer names and metrics available on a discovery call.
Inside-sales team spending 60% of the day navigating BC menus for customer balances, top-items, open orders, posted invoices. New hires took 6+ weeks to learn the menu structure.
Deployed SourceForge AI as a chat panel from every BC page. Pilot ran for 3 weeks; built a 30-prompt cookbook with the team. Production rollout in week 5.
- Page navigations per user-day: 220 → 78
- New-hire ramp time: 6 weeks → 11 days
- Top prompts: customer briefing, open orders, top items
Production team needed daily stockout risk, but BI reports lagged a day and couldn't be filtered ad-hoc. Buyers reordered late twice a month, causing line stops.
Enabled the stockout forecast capability with morning briefing automation for the buying team. Smart alerts configured for CRITICAL items.
- Line stops from stockout: 2/month → 0
- Buyer prep time: 90 min → 10 min daily
- Auto-briefing reaches 12 buyers at 8am
Common questions
Microsoft Dynamics 365 Business Central, Platform 25.0 or higher (Runtime 14.0+). Works on Business Central Online (SaaS) and On-Premises v25+. Compatible with the W1 base app and most Microsoft-supported localisations including India GST, UAE VAT, KSA, UK, Germany.
30-minute demo on a BC sandbox configured to your data shape.
Send us your BC version, user count, and the 3 most common questions your team types into a BC menu today. We'll demo a sandbox where SourceForge AI answers those three in plain English — plus the Preview-and-Confirm pattern on a real write.
- 30-minute live demo on a sandbox tenant
- Three prompts modelled on your real questions
- Preview-and-Confirm pattern walk-through
- Written quote within 48 hours