What exactly is Daxonet AI Frontier, is it just another chatbot on top of AutoCount?
No. A chatbot answers. Frontier operates. Frontier is an AI operating layer with seven named AI Departments (Sales, Service, Stock, Finance, HR, Marketing, Ops), each running as a closed feedback loop, observe an outcome, log it back into the Frontier Brain, tune the next response. The Brain is a structured 3-layer queryable model of your business (global identity, business data like SKUs and customers, department SOPs), not a vector store of documents. Every output is checked against Guardrails you write, the rules become your organisational IP. AutoCount or D365 sits underneath as the system of record. A WhatsApp wrapper cannot do any of that and never will.
Who is Frontier for, a small AutoCount business, or a large D365 enterprise?
Three customer shapes, one platform. (3) Existing-ERP extenders (already on SAP, Oracle, Odoo, SQL Account) get Frontier connected via APIs, same Brain, same Departments, no rip-and-replace. The Brain, Departments and Guardrails are the same underneath; only the integration shape changes.
What are the seven AI Departments and what do they actually do?
Sales-AI handles after-hours WhatsApp enquiries, drafts quotations, runs customer re-engagement, flags churn risk. Service-AI does smart booking with technician and weather context, callback prevention, route sheets. Stock-AI watches stockout risk, suggests inter-branch transfers, flags slow-movers. Finance-AI produces the daily cashflow brief, runs smart dunning, routes exceptions, surfaces e-Invoice status. HR-AI triages leave / OT requests, screens CVs, generates onboarding packs. Marketing-AI drafts festival campaigns, refreshes marketplace listings, drafts testimonials. Ops-AI ships a 07:00 daily brief, detects anomalies, preps and recaps meetings, builds the weekly board pack. Most SMEs start with three; the rest are licensed as needed.
What is the Frontier Brain, and why does it matter that it isn't 'just a vector database'?
The Brain is a three-layer structured representation of your business: Global (identity, fiscal year, ownership, locations, brand voice, languages used, BM / English / 中文), Business (SKU master, customer master, supplier list, pricing logic, payment terms, AutoCount or D365 live data, the last 12 months of orders), and Department (SOPs, escalation rules, harness criteria, approval thresholds, one editable markdown file per department). A vector database can retrieve passages of text. A structured Brain can be reasoned over, AI can apply your actual pricing logic to a quotation, not guess at one. Daxonet builds and maintains the Brain alongside your team; it becomes your operating IP.
How do Frontier Guardrails work, and how are they different from 'system prompts' in ChatGPT?
Guardrails are customer-defined harness rules that every AI output is checked against before it reaches you or a customer. Concrete examples from real Frontier instances: 'Never quote outside the AutoCount price for this customer group,' 'Never extend credit beyond the customer's credit limit,' 'Never send dunning to relationship-hold customers,' 'Always reply in the customer's language (BM / EN / 中文).' Every check is logged in the audit trail, rule, override, change. Unlike a system prompt, a Guardrail is structured, versioned, auditable, and assigned to specific departments. The longer you run Frontier, the more Guardrails you accumulate, and that library of rules is your switching cost.
Why multiple large language models, why not just pick one?
Because token cost is a competitive lever, not a fixed cost. Frontier's router picks the cheapest model that clears the harness quality bar for each call. A stock lookup or a Malay reply runs on Claude Haiku or MiniMax, cheap, fast, accurate enough. A multi-paragraph quotation with reasoning runs on Claude Sonnet, quality bar high, cost moderate. Mandarin customer ops route to MiniMax for native CJK quality. Long-document OCR and invoice extraction route to Gemini for the best price-to-quality on long context. Regulated tenants with sovereignty requirements run Qwen on-prem at zero per-token cost. Same customer experience, 30–60% lower run-rate than a single-model competitor.
How fast can a Malaysian SME on AutoCount actually go live?
Four weeks, on the productised path. Week 1 (Learn), boss plus one admin onboarded, Frontier installed, five demo scenarios run against your own AutoCount data. Week 2 (Wire), AutoCount database connector live, WhatsApp Business connected, Business Layer questionnaire completed (pricing logic, customer tiers, SOPs). Week 3 (Automate), first three AI departments live in shadow mode, approval rate above 80% by Day 4, first skill switches to autopilot. Week 4 (Scale), KPI review against your baseline, two more departments added, the Frontier Operator's Manual handed over, quarterly review cadence begins. The four-week timeline assumes your AutoCount data is reasonably clean; we audit and tell you honestly before Week 2 starts.
What about data residency and PDPA? Is anything trained on our data?
All customer data is hosted in the Azure Malaysia region. Cross-border data transfer happens only on an explicit, logged customer request, never silently. Frontier is PDPA-compliant by default, with a Data Processing Agreement template, customer-controlled retention windows, and right-to-erasure built into the admin console. Your data is never used to train the underlying base models, that is a contractual guarantee. Encryption is AES-256 at rest, TLS 1.3 in transit, with customer-managed keys on the Enterprise tier. For regulated industries (financial services, healthcare, government supply chain) we deploy Qwen or Llama on-prem so your data never leaves your tenant at all.
How much does Frontier cost, and is it grant-eligible?
Pricing is scoped to your size, modules and integrations. Daxonet quotes fixed-price after a short scoping call so there are no surprises. Most clients reach payback within the same project window.
What is the first thing you actually deliver? I have heard 'AI consulting' before, what's different?
Three concrete artefacts before any subscription starts. (1) An Impact Score map, which of your operations have the highest AI return on the lowest implementation risk, ranked and dollar-sized. (2) A Brain audit, a written assessment of whether your AutoCount or D365 data is clean enough to wire up, with a separate quote for any clean-up work that is needed. (3) A 30-day proof-of-value pilot on one department, with KPIs measured against your existing baseline. Only after the pilot proves out do you graduate into the subscription. We do not collect recurring revenue on customers who have not progressed past their pilot, that is built into our pricing model.
