Company
Agents you can put in front of an auditor.
Meridian Systems, Inc. builds narrow, governed AI agents for HR and finance operations. Each agent does one job, runs on your data under your security model, logs every action, and stops at a person for anything consequential.
- Founded
- 2026
- Legal entity
- Meridian Systems, Inc.
- Offices
- New York, London, Dublin
- Data regions
- US-East, EU-West
Why Meridian
Three bets, made in 2026.
The first bet is narrow, governed agents. The market spent two years building assistants that could attempt anything and guarantee nothing. HR and finance cannot buy that. A payroll error is a regulatory event, and a misposted accrual is a restatement. So every Meridian agent is scoped to one workflow, with a written contract for what it reads, what it may do, and what it must hand to a person.
The second bet is that the system of record matters more than the model. The Registry holds every agent, its version, its permissions, its approvers, its evaluation scores, and every run it has ever made. Models will keep changing. The record of what ran, on whose data, approved by whom, is what an auditor asks for, and it has to survive the model underneath it.
The third bet is pricing. We charge for completed work in credits, not for seats. It costs a customer nothing to give a team access, and it costs us if an agent burns compute without producing something the customer keeps. That puts our incentive next to the outcome instead of next to headcount. The rates and the calculator are public.
Principles
Six rules we build against.
They are enforced in the platform, below the agent, so no prompt and no builder can switch one off.
How the rules are enforced- 01
One workflow per agent
An agent scoped to a single job can be specified, evaluated, bounded, and rolled back. A general assistant can be none of those things.
- 02
A named person approves
Every consequential action routes to an accountable human before it takes effect. The approver is a field in the record, not a setting an administrator can quietly clear.
- 03
Nothing happens off the record
Inputs, model version, tool calls, approver, and outcome are written to the Registry and exported to your SIEM. If it is not in the log, it did not happen.
- 04
Your data is not our training set
Customer data never trains a model, ours or a provider's. Retention for agent inputs and outputs is configurable down to zero days.
- 05
Ship behind an evaluation
Each agent carries a test set drawn from real workflows. Releases are gated on it, and the scores move with the release notes in the changelog.
- 06
Charge for work, not seats
Consumption pricing means we are paid when an agent finishes work you keep. A seat licence would pay us whether or not anything got done.
How we build
Four commitments in the architecture.
Not a roadmap. These are properties of the platform today, and the reasons an agent release can be blocked.
Model-agnostic
The Gateway sits below the agent, so a workspace can enable, restrict, or replace providers without rewriting an agent. Bring your own model, or run only models we host in your region.
Domain-tuned models
General models do not know what a retro pay adjustment is, or why a credit memo lands in the wrong period. We tune on HR and finance work, using licensed and synthetic data, never customer data.
Human approval in the path
Approval is part of the workflow definition, not a notification bolted on afterwards. An agent that cannot find an approver stops and says so rather than proceeding.
Evaluation gates
Every agent has a held-out test set, a regression suite, and a published accuracy threshold. A release that drops below the threshold does not roll out, regardless of the ship date.
Leadership
Who owns what.
Three remits, written down. Meridian publishes the accountable owner for scope, for models, and for trust.
- CO
Claire Okafor
Chief Executive
Sets the company's scope, including which workflows Meridian will not take on.
- TL
Tomas Lindqvist
Head of AI Research
Owns the domain-tuned models and the evaluation sets that gate every agent release.
- PR
Priya Raghunathan
Chief Trust Officer
Owns security, privacy, and the audit evidence customers hand to their regulators.
Offices
Three cities, two data regions.
Where we sit and where your data sits are separate questions. Production data is processed in the region you select at provisioning, US-East or EU-West, and it does not leave it. Office location does not change that.
Subprocessors and regions- New YorkHeadquarters
- Commercial, trust, and applied research. Incorporated in Delaware.
- LondonEngineering
- Registry, Gateway, and Data Fabric. UK and EMEA solutions engineering.
- DublinEU operations
- Customer engineering for the EU-West region and EU data protection.
Press and contact
- Press
- Briefings, executive commentary, and the brand assets. Send a request.
- General
- hello@meridian.example. A person replies within one business day.
- Security
- Report a vulnerability or request the SOC 2 report on the security page.
- Working here
- Six roles are open across the three offices. See careers.
Get started
See one workflow run end to end.
A 45-minute session on your systems, including the approval step a person signs. You leave with a written plan and a credit estimate.