AI agents that take on your internal operations — bids, onboarding, licence management, alert triage — and, if you run a desk, resolve customer requests end-to-end. Every action verified, governed, and written back to your systems as an audit trail.
Built in production for a UK Microsoft-focused MSP, replacing a legacy on-premises PBX with Amazon Connect and AI triage.
The Model
Every request follows the same governed path — whether it's handled by an agent or a human, customer-facing or internal
Every request starts with a verified identity — SMS one-time passcode on voice, single sign-on on chat, staff identity for internal work. Verification is decoupled from execution, so every downstream action knows exactly who asked.
Governed tools execute the request with scoped, per-tenant credentials — no shared admin accounts, no credentials in prompts. Destructive actions are approval-gated.
Anything the AI shouldn't handle is warm-transferred to your people with the verified identity, a summary, and the full tool history — no 'can you repeat all that?'
Every action — who requested it, what was done, in which tenant — is written back to your system of record. The automation leaves the same paper trail your engineers do.
Internal Operations
No ticket-heavy desk? This is where the bigger prize usually is — the agent platform runs your internal operations standalone, no contact centre required
Agents draft SOWs and proposals grounded in your own CRM, past proposals, and pricing history. Senior staff review instead of authoring from scratch — bid for more, faster.
Reads each client's Microsoft 365 licence estate to surface unused and under-used seats for reclaim — a recurring saving hiding in plain sight across your whole client base.
Internal joiner workflows run under staff identity — a separate trust boundary from client data — so new hires are productive on day one.
Event-driven agents watch the queue, attempt resolution, run status checks, and post updates back to tickets automatically — no caller required.
NOC and RMM alert tickets arrive summarised with suggested fix steps from your runbooks, so engineers act instead of triaging.
Multi-tenant by design: bringing a new client under management is tenant consent and mapping — configuration, not an engineering project.
Customer Service Desk
For MSPs running first-line support: the high-volume, repetitive contacts — resolved end-to-end, not just deflected to a chatbot
Verified caller → reset executed via Microsoft Graph → confirmed on the call → ticket logged. The highest-volume L1 contacts, handled end-to-end.
"I can't access X" gets a live diagnosis: the agent gathers user, group, and licence facts in parallel, reasons over them, and asks clarifying questions when the request is vague.
Mailbox and shared-mailbox permissions, distribution list and security group membership, and joiner-mover-leaver flows like new-starter onboarding.
Account disablement and access revocation, gated behind a verified request from an authorised approver before anything destructive runs.
Locates and restores missing files across OneDrive and SharePoint, escalating to a human when standard search and restore don't find it.
Callers get live status from your PSA without holding for an engineer, and everything else routes to the right queue with full context attached.
Under the Hood
Five layers, front to back — the same tool-use loop that powers Claude Code, running inside your own AWS on Amazon Bedrock AgentCore, not a third-party black box
Amazon Connect replaces the legacy PBX as the single intake point. Voice runs on Amazon Nova Sonic speech-to-speech; chat and email converge on the same routing. Optional — internal-only deployments skip it entirely.
Microsoft Entra is the identity fabric across client tenants; your PSA is the system of record that maps identity to client, authorisation level, and audit trail.
Claude-based agents built on the Claude Agent SDK run on Amazon Bedrock AgentCore — inside your own AWS account, grounded on your knowledge base and live PSA state.
Every action flows through MCP tools on AgentCore Gateway — single-purpose functions with authorisation re-checked at execution time. Integrations plug in behind this boundary; Microsoft Graph, PSA, RMM, and documentation tools slot in without touching the agents.
Every automated action lands on the PSA ticket, with end-to-end tracing. Deflection rate, handle time, and first-contact resolution are measured from day one.
Model usage is tiered — fast models classify and route, stronger models reason over the hard cases — so cost scales with the work actually done, not with call volume.
Outcomes
Password resets, licence audits, proposal drafts, alert triage — the high-volume, low-judgement work across your desk and back office. Every task an agent completes returns capacity to work that actually needs a person.
Whatever reaches a human — a routed call or a drafted proposal — arrives with verified identity, a summary, and the work already attempted. Nobody starts from zero.
One platform serves your whole client base. Onboarding a new client is configuration — tenant consent and mapping — not a rebuild.
The entire stack runs in your own AWS account. Models don't train on your data, and workloads can be pinned to your region for residency.
Delivery is phased and scope-gated — identity foundation first, then live automation where it pays back fastest, whether that's self-service on Microsoft 365 or your internal workflows. No big-bang cutover.
Tell us where the repetitive work piles up — your desk, your back office, or both — and we'll show you what the agents would handle. Starting with a working demo, not a slide deck.