Tesora Workbench: the agent harness behind actuarial work

We rebuilt actuarial work on the frontier AI models.

When AI is taught to think like an actuary and to produce full citations and reproducible work, you spend more time on results and less time checking work.

Power user? Call our MCP server from within your own Claude or Codex workflows.

If your team already works in Claude or Codex, it does not have to switch. Your rating models, documents, and review queue are available there directly.

Your book sits on its own machine.

Most software like this keeps every customer in one database and tags each row with who it belongs to. We gave each customer a machine of their own instead, because the first thing an actuary asks about a loss run is where it goes.

Established actuarial software has good methods built in, but it cannot take an instruction. General AI models will do anything you ask, but they make mistakes, like using the wrong LCM or trend-to date, without noticing or flagging them.

Tesora is built for both: the approved methods for each study, a source for every figure, and an audit before anyone sees the result. AI that demos well is easy to build. AI you can trust with work an actuary has to sign is not.

How Tesora compares.

Every number is traced to its source.

Two checks make sure of it, and you can verify both.

Common questions.

Try to break it.

Send us an analysis with a mistake in it and see which audit agent catches it.

Spreadsheets and filing PDFs. Flexible and familiar, but slow, and every update is manual.

Works well for the workflows it was designed for, and breaks down outside them.

Typically two engineers and a year to build, then ongoing maintenance for your team.

Ask for a change to the rater you already filed and get it back tested.

We already have Claude or Copilot. Why Tesora?

Those are models, and we route to them. Tesora adds what is built around them: vaulted data, a citation on every figure, and audit agents that catch censored data and mismatched aggregations.

Which model are we paying for?

Whichever one gets each step right at the lowest cost. We handle the routing, and because we are paid for results, keeping model costs down is in our interest too.

What do we get back?

The finished deliverable and the full trail behind it, so your team can reproduce it next quarter in your own format. A consultant gives you only the deliverable.

How does this fit with our current tools?

Excel, Python, R, or a mix of all three. Nothing has to be migrated first. Onboarding is mostly Tesora learning how your team presents its work, so reports come back in your format with only your judgment left to add.

Where it runs

Hosted by us on your deployment’s endpoint. The address is on the app’s MCP page, and you sign in the first time you use it.

Who it runs as

You. Anything started from Claude or Codex runs with your permissions, no more and no less.

What it will not do

Promote or approve. A rater built this way still goes through your normal approval process.

The environment

You get one virtual machine, a service account that only it uses, and a data disk of your own. Your team signs in with your SSO.

Where the data lives

The database and the data lake both live on that disk. When a trial ends we delete the disk and the machine, and its nightly backups expire two weeks later.

Getting it in

We can connect to your systems, and we do not have to. Most trials start with files your team uploads and nothing else.

Agents and the internet

Agents and workflows run in hardened sandboxes with web search and fetch switched off. The sandbox sets that when each session starts, so it does not depend on the model behaving.

Our own actuarial benchmarks

There was no standard way to measure whether AI is good at actuarial work, so we built one, graded against what a reviewer would accept. We also post-train models on the actuarial workflows we see every day.

Paid for results, not tokens

Each step goes to the lowest-cost model that gets it right. A model vendor has no reason to do that for you.

Called in when internal builds stall

In-house AI projects usually get built, run over budget, and then fail IT review because governance was never part of the design. Tesora has governance built in from the start.