CLOSE
megamenu-tech
CLOSE
service-image

Company

CLOSE
CLOSE
CLOSE
Embassy Case Study

Org GPT

Embassy Development Limited

A conversational surface for the whole enterprise. Reckonsys is engineering Org GPT as Embassy's system of intelligence: one permission-aware interface where employees ask questions of the company's knowledge estate and complete real workflows, on every device they already use.

One prompt, end to end

Employee asks: “What's my remaining leave balance, and what's the SharePoint policy on carry-forward?”

1
Client attaches identity
Web, desktop, PWA or Teams sends the prompt with a signed Entra ID access token over TLS.
2
Gateway hydrates policy context
Token validated; roles, agent entitlements and tenant resolved. A trace id is issued that follows the request to the end.
3
Router classifies intent
A router model decides whether one skill handles this or several. Here it fans out: Leave Agent and SharePoint retrieval.
4
Leave Agent calls HRMS
Employee id always resolves from the token, never from user input. A user cannot ask about someone else by naming them.
5
Retrieval, pre-filtered by ACL
A delegated Graph token scopes the search to sites the user can already read. Top candidates are re-ranked.
6
Answer model composes
System prompt, retrieved chunks, tool results and conversation summary within a budgeted context window. The response streams back.
7
Everything is written down
Citations, tool calls, token counts, latency and per-provider cost persist to the audit store. A follow-up question is suggested in parallel.
Client
Embassy Development LimitedAn Embassy Group company
Industry
Real Estate & Property Development
Built with
Claude API One of several providers behind a model-abstraction layer
Services offered
AI DevelopmentCustom Software DevelopmentRAG Model DevelopmentAI Agents Development
01

The brief

Embassy runs on systems of record: SharePoint for documents, HRMS for people, SAP for finance, plus payroll, travel and expense platforms. Each holds what an employee needs. None of them answer a question.

Org GPT sits alongside those systems and exposes them through a single conversation. An employee asks in plain language; the platform retrieves what they are entitled to see, calls the right business system, and completes the transaction. It runs on the web, on a Windows desktop client, on native iOS, as an installable PWA and inside Microsoft Teams, from one backend and one design system.

Reckonsys is engineering the platform end to end: architecture, build, hardening and release across all five surfaces.

02

Design tenets

  • Six commitments set before any code, agreed with Embassy and applied as constraints on every decision that followed.
Identity-first
Every response is scoped to what the requesting user is entitled to see. No query bypasses SharePoint ACLs, SAP authorisations or HRMS PII boundaries.
Multi-model
Providers are abstracted, so workloads route to Claude, GPT or a self-hosted model without application-level rewrites.
Composable agents
Each business process is an isolated agent with its own toolset, system prompt and eval harness, orchestrated by a router.
Surface-agnostic
One backend serves web, Windows, native iOS, an installable PWA and a Teams tab and bot.
Auditable by construction
Every prompt, tool call, retrieved document, citation and response is logged to an immutable audit store.
Quiet authority
In keeping with Embassy's brand, the interface is considered and reliable. Never gimmicky.
03

Architecture

  • A modular, cloud-native design organised into five horizontal planes, with identity, observability and security running vertically through all of them.
Plane 01Experience
Web app, Windows desktop client, native iOS, installable PWA and a Microsoft Teams tab and bot. All share one design system, one auth SDK and a typed API contract.
Plane 02API & orchestration
A gateway exposing versioned REST and streaming endpoints. Hosts the agent router, the tool-call engine and the streaming response pipeline.
Plane 03LLM & retrieval
A provider-abstracted client with unified token accounting, cost controls and fall-over. Retrieval runs over a vector index pre-filtered by per-user ACLs.
Plane 04Enterprise integration
Connectors to SharePoint, SAP, HRMS, payroll and travel systems, all through a reusable framework with circuit breakers and typed schemas.
Plane 05Data & governance
Append-only audit store, per-user memory store, policy store for RBAC and feature flags, and the admin console backend.
Cross-cutting
  • Identity and authorisation - Microsoft Entra ID as identity provider, delegated Graph tokens for retrieval, and a policy engine gating agents and features.
  • Observability and audit - distributed traces, structured logs, per-prompt audit records, cost dashboards and PII-redacted conversation replay.
  • Security - managed secrets, encryption at rest and in transit, isolated egress, and a hardened rendering pipeline for rich in-chat responses.
04

The agent catalogue

  • Rather than one model attempting everything, each Embassy process becomes an isolated agent with its own tools, prompt and evaluation set. A router dispatches to the right one, or to several at once.
Leave
Apply for leave and approve requests from the conversation.
HRMS
Holiday calendar
Query the organisation's holiday calendar directly.
HRMS
Travel
Raise travel requests and move them through approval.
Travel system
Expenses
Submit and approve travel expenses, with attachments and a state machine behind the chat.
Finance
Reimbursements
Batch multiple invoices into one submission, parsed and validated before it reaches finance.
Finance · OCR
Payslips
Retrieve payslips under strict PII gating, with every access audited.
Payroll · SAP
SAP password reset
Trigger a reset through a secure, verified flow instead of a helpdesk ticket.
SAP
Knowledge retrieval
Grounded answers from SharePoint, scoped to the sites each user can already read.
SharePoint
Anthropic partnership

Claude in a multi-model platform

Org GPT does not bet the platform on a single model, and that was a deliberate architectural position rather than a hedge. Embassy needs to route different workloads to different models, and to change that routing as models improve, without rewriting the application. So the LLM plane sits behind a provider abstraction with unified token accounting, cost controls and fall-over. Claude is a first-class provider within it.

Where Claude earns its place is the work that carries consequences. Grounded answers over Embassy's document estate have to stay inside retrieved material and cite it, because a confident but unsourced answer about a leave policy or an approval threshold will be believed and acted on. Agent routing has to classify intent reliably enough that a payroll question never lands in the travel agent. Both are instruction-following problems under constraint rather than raw generation problems, and both are where Claude's consistency does the most work.

The provider abstraction also makes that claim testable rather than asserted. A dedicated evaluation harness re-runs a curated prompt set on every model change and blocks a merge on regression, so provider selection per workload is settled by measurement.

Model choice becomes a routing decision, not a rebuild.
Claude API Provider abstraction layer LLM eval harness Agent router RAG · ACL pre-filtered Per-provider cost accounting
05

How it ships

The programme is structured so Embassy makes its second investment decision on working software rather than a slide. Two commercially independent phases, each sized and accepted on its own.

Delivery runs as dual-track agile: a discovery and design track one sprint ahead of build, and a build track shipping behind feature flags with a demo every two weeks. Trunk-based development, mandatory review, infrastructure as code for every environment, and an LLM eval pipeline that blocks merges on regression.

  • Phase 1: MVP in users' hands
    The Windows and native iOS clients, the permission-aware SharePoint backbone, the RBAC governance module, the multi-provider switch, and the first slice of business agents: holiday calendar, travel expenses, multi-invoice reimbursements, payslips and SAP password reset. At the end of it, Org GPT is live, governed, and transacting against real Embassy systems.
  • Phase 2: Depth and reach
    PWA and Microsoft Teams surfaces, multi-model response comparison, task management, long-context memory, agent routing depth, rich in-chat UI responses, curated golden databases, structured-database retrieval over an entity ontology, and the remaining leave and travel agents. Triggered only once Phase 1 is accepted.
06

What it's for

The outcomes the programme is being measured against.

01
Fewer routine tickets reaching HR, IT helpdesk and finance shared services.
02
Faster onboarding, with new employees transacting with enterprise systems in natural language from day one.
03
A unified, governed surface for executives to query portfolio, tenant, lease, project and HR data without waiting on ad-hoc reports.
04
A defensible, audit-ready position for the CISO and Data Protection Officer on the safe use of generative AI inside Embassy.
Key takeaways
71
Projects executed
60+
Team members
28+
Products built for clients
28+
Projects funded
Contact us

Let's build the surface your whole company can ask.

Get in touch

Modal_img.max-3000x1500

Discover Next-Generation AI Solutions for Your Business!

Let's collaborate to turn your business challenges into AI-powered success stories.

Get Started