That outputs are grounded in the record, and that a person verified them
Every answer carries the document and the page it came from, and the page opens. Drafts are drafts until a person approves them.
Security and data ownership
The question is not what we built. It is what you can demonstrate.
ABA Formal Opinion 512 does not ask whether your AI vendor is good. It asks whether your firm can show what the vendor retains, whether it trains on your data, what its security protocols and liability limits are, whether it screens for conflicts among your clients, and where you go for relief if it breaches. Most security pages answer a different question. This one answers that one.
SOC 2 Type II at the company level
A separate database per tenant
Your AI accounts, not ours
US data residency
Mapped, item by item
What you have to show, and what we give you to show it with.
That you tested the tool's accuracy on a sample of your own matters before relying on it
The eight-week implementation includes an extraction trial on your own live binders, and the coverage self-score on every extraction tells you what it could not find.
Whether the vendor retains your inputs after the engagement ends, or asserts rights over them
Your data lives in your own tenant database. If your firm configured it, generated it, uploaded it or produced it, it is yours, by contract.
Whether your inputs train a model or surface to another user
The AI provider accounts are in your firm's name under standard paid API terms, which exclude prompts and responses from model training. That protection is between your firm and the provider, not something we negotiate on your behalf.
The vendor's security protocols and its limitations on liability
SOC 2 Type II report available under a non-disclosure agreement. Liability terms are in the agreement, not in a marketing claim.
That access is supervised, and that AI-produced material is identifiable in the file
Every tool call, admin action, approval and export is logged with the user and the timestamp. Deterministic steps and model calls are visually distinguished everywhere they appear.
For agentic systems, exactly what firm systems the agent can reach, and that nothing leaves without human review
Connectors are typed and per-tenant with field-level access controls. Every write to your case system is staged, whitelisted and approved by a person. Nothing sends to a third party without an approval.
That you have a written firm AI policy and delivered training
Administrator deep-dive, working sessions on your own book, a recorded training library inside your tenant, and 90 and 180 day refreshers, all part of implementation.
That billing reflects actual cost, with no undisclosed markup on passed-through AI expense
Zero markup on model consumption. Your firm holds the provider accounts and receives the raw provider invoices.
Sources for the demands above: ABA Formal Opinion 512 (29 July 2024), the State Bar of California 2026 Practical Guidance, Texas Ethics Opinion 705, Florida Ethics Opinion 24-1, and 22 NYCRR Part 161, effective 1 June 2026. We are not your ethics counsel and this is not legal advice. It is a map of the questions we are asked most often, and our answers to them.
Architecture
A database per customer, and deny by default.
Your own PostgreSQL database
Provisioned when your tenant is created. Physical isolation, not row filtering, and no shared customers table anywhere. A dump of another customer's database contains nothing of yours.
Single sign-on
SAML and OpenID Connect against your own identity provider, so multi-factor authentication and your own conditional access policies apply to Mayeshal exactly as they apply to everything else.
Role-based access, deny by default
Grants are resource, action and scope. Filtering happens in the query, so a record a user is not entitled to see is never retrieved rather than merely hidden.
Permission gate before every tool call
The check runs before the tool executes, not after it returns.
Our people are not in your tenant day to day
Support access is granted by you, time-bound, and logged. The platform console we operate and the tenant console you operate are separate systems.
Encryption
TLS on every external connection with strict transport security. At rest, AES-256 on the managed database cluster and its backups, with credentials and provider keys separately encrypted at the column level.
US data residency
Droplets and the managed database cluster run in the United States.
The retention question
The provider relationship is yours, which is why the answer is simple.
Most vendors have to explain their own data retention policy to you, because your prompts pass through their account. Ours do not. Your firm's AI provider accounts are created in your firm's name and billed to your firm's payment method. Standard paid API terms already exclude prompts and responses from model training, so when your counsel asks whether your client's medical records are training a model, the answer is a term in a contract your firm signed directly with the provider, and you can read it.
You also get the raw provider invoices, and an in-platform cost meter that records every call. Two independent records of the same thing, one of which we do not produce.
Each agent has its model pinned by identifier. A provider shipping a new version does not silently change how your firm's letters read.
The test to apply to every vendor
If you left tomorrow, what could you take with you?
Ask that of everyone on your shortlist, including us. Our answer, in five parts, and it is a contractual commitment rather than a feature we are describing:
- All of your configurations in structured formats.
- All of your stored data in JSON, CSV and PDF.
- A full copy of your tenant database, under a data transfer agreement. Because you have your own database rather than rows in a shared one, this is a straightforward operation rather than an extraction project.
- Documentation of the configuration state as it stood.
- A transition period of read access, so you can verify all of the above before anything is switched off.
The definition of what is yours is deliberately blunt: if your firm configured it, generated it, uploaded it or produced it, it is yours. The orchestration engine, the connector framework, the permission model and the source code are ours. There is no third category.
We are also honest about what this is today. The exit package is a documented process
we run, not a button in the console. The self-serve version is on the roadmap and is listed on the
platform page under what is not built yet.
Stated honestly
Certified, aligned, and on the roadmap. In three separate columns.
SOC 2 Type II
At the Micronetbd company level. The report is available under a non-disclosure agreement, covering Security, Availability and Confidentiality. Mayeshal is built and operated under those controls. Mayeshal is not yet separately in audit scope as its own system, and we will not imply that it is.
HIPAA, GDPR, CCPA and ISO 27001
HIPAA-ready posture, with protected health information tagged and access-controlled by default and business associate agreements available on Enterprise. GDPR and CCPA, with the customer as controller and Micronetbd as processor. ISO 27001 aligned in practice; not certified. Certification is on the roadmap and we will say so until the certificate exists.
Partner credentials
Salesforce Silver (Ridge) Consulting Partner. Filevine partner. Litify partner. DocuSign consulting partner.
Continuity
Four hours and one hour, stated as targets.
Our recovery time objective is four hours for a full database restore and our recovery point objective is one hour of maximum data loss, in practice much less through transaction log recovery. Application services are stateless and redeployable from images in minutes.
Targets, not measurements. We are calling those targets rather than measurements,
because that is what they are. If a timed restore test result matters to your evaluation, ask on the
call and we will tell you exactly where that stands.
Get started
See it read one of your own matters.
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.