Workflow agents
Agents that carry a piece of work from intake to completion across your ERP, CRM and ticketing systems.
More on workflow agentsAgentic AI for enterprises
AI agents that run inside your business. Engineers who sit inside it too.
Unmatrix AI builds agentic systems that do real work across your ERP, CRM and ticketing stack, and sends forward deployed engineers to your office to ship them into production.
First notice of loss · Insurance
Simulated run
Simulated example of an Unmatrix agent handling first notice of loss for insurance: it reads the case, checks your systems, sends the risky decision to a person, and writes the result back with a full audit trail.
Show me my industry
Deployed inside
- Northwind Insurance
- Halvard Energy
- Meridian Health
- Corvid Logistics
- Atlas Manufacturing
- Ferris Capital
What we build
Agents for the work that keeps your operations running
Not chat windows. Systems that take a case, act across your tools, and hand the risky decisions to a person.
Knowledge and retrieval agents
Grounded answers over policies, contracts and engineering documents, with citations and your existing permissions.
More on knowledge and retrieval agentsOperations copilots
Assistants that sit beside support, operations and field teams to draft, triage and escalate in real time.
More on operations copilotsAgent platform and guardrails
The shared layer every agent runs on: evals, permissions, audit logs, review queues and model routing.
More on agent platform and guardrailsHow we build them
Three architectures, chosen by how well the work is known
Autonomous agents for open-ended investigation, LangGraph workflows for known processes, RAG for questions whose answers live in your documents.
- How it works
Autonomous agents (ReAct)
Agents that reason, call a tool, read the result and decide the next step in a loop, inside a harness that enforces tool allowlists, step limits and human review.
- How it works
LangGraph workflow agents
Graph-based workflows where the model decides only at chosen nodes, with checkpointed state and human-in-the-loop interrupts. Example: a resume screening agent.
- How it works
RAG agents
Document ingestion with permissions, hybrid keyword and vector retrieval with reranking, and grounded answers with citations or an honest not-found.
Every system ships with guardrails, PII and PHI protection and prompt injection defence. For regulated environments we deploy open-source LLMs on-prem.
How deployment works
Twelve weeks from one workflow to production, most of it at your desk
Forward deployed engineers embed with the team that owns the work. Brass marks the weeks they are onsite.
Week 0
Remote
Scoping
We pick one workflow with a number that has to move, list the systems it touches, and agree what done looks like. Two calls, one written scoping note.
Weeks 1–2
Onsite
Discovery, embedded
Two to four engineers arrive at your office. They sit with the team that owns the workflow, shadow the work, map the exceptions, and get access sorted.
Weeks 3–8
Onsite
Build, in your environment
A working agent is running on your systems by week four. From then on it is reviewed weekly by the people who will use it. Evals are built from your real cases.
Weeks 9–12
Onsite
Production and handover
The agent runs on live volume with human review on every risky action. Your engineers pair with ours. Runbooks, evals, dashboards and on-call move to your team.
Ongoing
Remote
Support, then the next workflow
Your team owns the system. We stay reachable, review model changes with you, and come back when you are ready to point the same approach at the next workflow.
The platform underneath
Every agent runs on the same shared layer, deployed in your cloud
The first agent is a project. The fifth is a platform problem. The Unmatrix platform is the evaluation, permission, review and audit layer that every agent inherits, installed in your VPC and owned by your platform team.
Bring your own models and data
Routes between frontier and small models through your own endpoints. Connects to your document stores and warehouses with your permissions.
Evals on every change
Test suites built from your real cases run on every prompt, model or code change, so nothing regresses quietly.
Review queues and audit logs
Risky actions wait for a person. Every read, decision and write is logged with the model version and exportable to your SIEM.
Unmatrix platform · your-vpc
Results
Measured on the number we agreed in week zero
- median time to the first agent on live volume
- 6 weeks
- a week onsite during discovery and build
- 3+ days
- of agents ship with evals, audit logs and human review
- 100%
Northwind Insurance
“They sat with our claims team for ten weeks. The agent they left behind handles first-notice intake the way our best adjuster does, and our engineers run it.”
Name to come
Chief Operating Officer
Atlas Manufacturing
“Every other vendor sent a deck. Unmatrix sent two engineers who asked for a desk and SAP access on day one. We were on live volume by week nine.”
Name to come
VP, Supply Chain Systems
Meridian Health
“Our security team approved the design because it was built inside our VPC with our permissions model. The audit log was there before the demo was.”
Name to come
Chief Information Security Officer
The difference
Enterprise AI fails in the gap between the demo and the floor
Most of the market sells the demo. Forward deployment is how we close the gap.
What most vendors provide
- A strategy deck and a roadmap
- A pilot that never meets your ERP
- A remote team and a weekly status call
- Agents with no audit trail or review path
- Handover as a PDF
What Unmatrix provides
- Engineers embedded in your office, on your systems
- Production on live volume inside twelve weeks
- Evals, permissions and audit logs in the first build
- Your code, your cloud, your team on call
- One number that has to move, reported honestly
Where we work
Regulated, operational, system-heavy
The first workflow we usually start with, by industry.
- Banking
- KYC refresh and lending file preparation
- Insurance
- First notice of loss and claims triage
- Government
- Permit and benefits case processing
- Healthcare
- Prior authorisation and referral queues
- Fintech
- Dispute investigation and KYB onboarding
- E-commerce
- Returns resolution and guided selling
Questions
What buyers usually ask
What does forward deployed actually mean?
Our engineers work from your office, inside your environment, alongside the team that owns the workflow. They are not a remote delivery team with a weekly status call. They are at the desk when the exception happens.
How many engineers, and how often are they onsite?
Two to four engineers per engagement. During discovery and build they are onsite at least three days a week. During handover it tapers as your team takes over. We adjust to your working pattern and travel policy.
Who owns the code, the prompts and the evals?
You do. Everything is built in your repositories and your cloud accounts. At handover there is nothing to migrate because nothing ever lived anywhere else.
Can this run inside our VPC or on-premises?
Yes. We deploy into your AWS, Azure or GCP accounts, and can work with private model endpoints such as Bedrock, Azure OpenAI or a self-hosted model. Data does not leave your boundary.
Which models do you use?
Whichever fit the task and your procurement. We usually start with a frontier model for reasoning-heavy steps and smaller, cheaper models for classification and extraction. The platform routes between them and lets you swap later.
Next step
Pick one workflow. We will be at your office in two weeks.
A thirty-minute call to find the right first workflow, followed by a written scoping note. No deck, no pilot.