Underneath the vertical packs there is one platform: one login, one permission model, one audit log, one cost meter, one set of connectors. Adding a capability is a configuration action, not another vendor.
Every request enters one place. It resolves the caller's permissions first, picks the agent, retrieves only the data that caller is allowed to see, and coordinates the model calls. Agents never talk to a source system or an end user directly. When an agent hits work outside its own scope it hands off through the orchestrator, under the same user's permissions, and cannot escalate. A denied handoff is written to the audit log.
Deny by default. Every tool call is checked against the caller's grants before it runs, not after. Tags such as protected health information and confidential are enforced as constraints, and filtering happens in the query rather than in the display layer, so a record a user may not see is never retrieved in the first place.
Typed, per-tenant, with field-level access controls and separate sandbox and production environments. Every write to a source system enforces the calling user's own permissions in that system. An agent cannot write where the person could not.
Your system of record is copied into your own database. At one live firm that is 61 objects and 1,277,411 rows, verified at exact count parity on all 61. A five-minute tick keeps the busy objects current; a nightly full pass rebuilds everything, four objects at a time. Rows are never deleted, only labelled live, archived, deleted or purged, so a record that disappeared upstream is still answerable here. On that firm's data, a full rebuild takes about sixteen minutes.
Every document opened page by page and indexed, with a summary, the parties, the dates and the amounts extracted, and a citation you can open. Documents are held in your tenant's own storage after first read, so they stay openable when the source system is not.
One wrapper around every single model call. It records the model, the tokens in and out, and the dollar cost, before and after the call. There is no path to a model that bypasses it, which is why we can show you a per-document cost to four decimal places.
A vertical pack is a signed, versioned manifest plus its asset files. Installing one adds agents, workflows, custom objects, roles, reports and dashboards into your existing tenant. It does not create a second application, a second login or a second permission model. Uninstalling is symmetric: the agents suspend, and your objects and history stay exactly where they are.
Packs are signed before they are published, and the signature is verified before install. Install runs as an idempotent state machine with a dry run first, and every step is audited. One tenant can run more than one pack.
Ten capabilities, five roles, the document intelligence layer, the corpus-grounded demand engine and staged writeback. Running today at a Texas personal injury firm on Litify.
Fourteen configured agents, ten custom objects, six roles, five correlation rules across the systems an advisory firm actually runs, and a books-and-records retention posture. The pack is built, signed and published, and its dry-run install passes end to end. We have not yet run it in production for a client, and we would rather say so here than let you find out on the call.
If your practice area is not one of these two, the honest answer is that we would be building the pack with you as the first customer, which is a different conversation and a different price. Ask for it.
Your tenant gets its own PostgreSQL database. Not a filtered view of a shared one. Your AI provider accounts are created in your firm's name and billed to your firm's payment method at published rates, and you see the raw invoices. We add zero markup to model consumption, which means the amount we make does not change when your usage does.
That also makes model choice portable. Switching or mixing providers is a configuration change, not a vendor change, and each agent has its model pinned by identifier so a provider's release does not silently change how your firm's letters read.
Delivered in your implementation
Voice agents. Approval-gated outbound dispatch. The litigation suite. Legal research with citation verification. Production optical character recognition.
We would rather you read this here than discover it in month three.
The workflows run. There is no drag-and-drop editor for them yet, so changing one is a configuration task we do with you rather than one you do alone.
Prompts are versioned and every change is recorded, but editing them today goes through us rather than through a self-serve review-and-publish screen.
Failures are captured and visible. The one-click retry surface is not built.
Documents enter prompts as data rather than as instructions, which is the load-bearing protection, but we do not yet run signature scanning or strip active content on the way in.
The exit package is a contractual commitment we deliver as a process today. The button that generates it on demand is not built yet.
Thirty minutes, on a screen share, on a matter you pick. We connect nothing, install nothing and ask for no credentials to have that conversation. If it is not obviously useful in the first ten minutes, we will tell you so and you get your afternoon back.