Response to Wiseway Logistics — AI Assistants Brief
Three assistants in the brief. A foundation that quietly extends to the next ten — customer service, AP, AR, reporting, compliance — without rebuilding. Built on open standards so you can swap the model, the harness, or the host when the landscape shifts. And it will shift.
01 — Approach
The AI tooling market is moving faster than any enterprise procurement cycle. A platform locked to one vendor's bundled product today is a rebuild in 2027. Our recommendation is a small set of open, swappable layers glued together by open standards — so when a better model, a faster harness, or a cheaper host appears, you change one component, not the whole system.
The agent harness (Claude Agent SDK) is open source. Skills live in your GitHub. Every integration is a Model Context Protocol (MCP) server — the emerging open standard for tool calls. Replace any layer without rewriting the others.
Skills, prompts, evals, rate-card logic, TMS adapters — all in a Wiseway GitHub repo, code-reviewed, branch-protected. By the end of the engagement your team merges PRs, not raises support tickets.
Harness and MCP servers run in Azure Australia East alongside your existing Microsoft 365 footprint. Documents never leave your tenant. Only model inference traffic egresses, and that endpoint is configurable.
Knowledge, Quoting and Transport assistants share the same harness, the same identity, the same admin console, the same logs. Adding AP, AR, customer service or reporting later means a new skill — not a new product.
02 — The stack
The brief asks for a platform that can be extended without rebuilding. We get there by keeping the layers thin and the interfaces open. Click any component in the diagram below to see what it does and what it can be swapped for.
Anthropic offers a managed-agents capability where the harness itself runs on Anthropic's infrastructure. It's elegant but, as of today, that infrastructure is US-based — which sits awkwardly with your "keep our data in Australia" preference and means your SharePoint and TMS traffic would cross the Pacific on every call.
Our recommendation is to self-host the Claude Agent SDK in Azure Australia East. The SDK is open source and lightweight — a container you run on App Service, Container Apps, or AKS. Anthropic will support the engagement, but the host is yours. If their managed offering lands in Sydney later, migrating is a redeploy, not a rebuild — because the SDK interface is the same.
03 — Architecture
This is the picture for all three assistants — and the ones you'll add later. The harness on the left orchestrates skills; skills call MCP servers; MCP servers wrap your real systems. Nothing here is invented for Wiseway — every piece is an open standard or your existing system.
04 — Initial scope
The brief is pragmatic: pilot small, prove value, expand. We agree. Knowledge Assistant comes first because it has the lowest integration risk (read-only, SharePoint only) and the broadest staff exposure — the fastest way to build internal trust in the platform. Sales Quoting follows once the rate-card model is settled.
05 — Before we can give you a timeline
The brief is clear on outcomes but light on the systems underneath — which is fair, because that's our job to discover. We can't put numbers against the first two use cases until we've answered the questions below. This is a two-week discovery, not a six-month one.
Blocks: Phase 1 & 2 estimate
Blocks: Knowledge Assistant scope
Blocks: Go-live security review
Blocks: Channel build effort
Blocks: Deployment plan
Blocks: Adoption & ownership transition
Two weeks. Three workshops (systems, security, users). Output is a fixed-price Statement of Work for Phase 1 (Knowledge), a tight estimate band for Phase 2 (Quoting), and a written architecture decision record you keep — whether you proceed with us or not.
06 — Next steps
We'd like to spend a fortnight inside Wiseway's systems and with Wiseway's people, then come back with a fixed-price Phase 1 and an estimate band on Phase 2. No deck-ware. No surprises after sign-off.