Published · Updated
The real question isn't what AI can do
Every demo shows what AI can do. The question that decides whether a firm adopts it is narrower: what can it do without asking, who can stop it, and how would you prove — months later — what happened and who approved it.
Those are governance questions, and they deserve answers before the first workflow goes live. AI that operates inside firm-defined boundaries is adoptable; AI that can't demonstrate its boundaries isn't, no matter how impressive the output.
Permissions that apply to people and agents alike
FernLex uses role-based access control across nine roles — managing partner, attorney, paralegal, legal secretary, intake coordinator and more — sixteen resource types (cases, clients, documents, drafts, billing, filings, approvals, AI agents and others) and six permission levels. The same boundaries govern every user and every AI agent, so an agent can never reach a record its human counterpart couldn't.
- View — read the record.
- Create — add new records.
- Edit — modify existing records.
- Delete — remove records.
- Export — take data out of the system.
- Manage — configure the resource itself.
Policy checks before every action
Permissions are the fence; the policy engine is the checkpoint. Policies are evaluated at four levels — org, client, case and action — and every agent action is checked before it executes, not audited after the fact. Agents also stay inside their defined jobs: Marcus routes work, Atlas analyzes matters, Harper qualifies leads — no agent operates outside its function.
- Client-level and case-level governance rules scoped to individual engagements.
- Usage budgets and cost controls — daily and monthly limits per agent, plus firm-wide caps.
- A master AI switch: firm-wide pause and disable for all AI operations, effective immediately.
Approval gates where the stakes are highest
Governance is only real if there are actions AI cannot complete alone. In FernLex, drafting agents operate in draft mode — nothing is finalized, sent or filed without attorney sign-off. Financial AI is similarly bounded: it cannot create billable time without human input and cannot send invoices or execute transactions. Suggestions land in a review queue; manual tracking always overrides AI suggestions; and AI activity itself is never billed to clients.
The audit trail
Every action — human and AI — is recorded in a tamper-evident audit log. Records are sequentially numbered and checksum-verified, so any alteration is detectable. The log captures login tracking and policy decisions alongside ordinary activity, lists AI employees as actors in their own right, and is retained for up to seven years. When a client, a bar counsel or an opposing party asks what happened, the answer is a record, not a recollection.
Confidentiality and data protection
Governance sits on top of ordinary information security: encryption in transit and at rest, field-level encryption for sensitive values like SSNs and passport numbers, cryptographically signed sessions, and region isolation with US and EU data center choices. Your firm's information is never sold, never used for marketing, and never used to train AI models. The platform is designed to meet ABA Model Rule 1.6 and 1.1 confidentiality expectations, state bar technology competence standards, and GDPR/CCPA requirements, with the NIST Cybersecurity Framework as its stated control baseline.
Frequently asked questions
What does it mean to evaluate policies at org, client, case and action level?
Rules can be set firm-wide (every matter), narrowed to a single client, tightened for one case, or applied to a specific action type. Each agent action is checked against all applicable levels before it runs — so a firm-wide rule can't be bypassed by a case-level setting, and vice versa.
Can AI ever act without a human approving first?
Within its scope, yes — that's the point of delegation. But consequential actions are gated: drafts are never finalized, sent or filed without attorney sign-off, and financial actions require human input. Every ungated action is still logged and reversible.
What exactly is in the audit log, and how long is it kept?
Every action by every user and AI employee, plus login events and policy decisions. Records are sequentially numbered and checksum-verified for tamper detection, and retained for up to seven years.
Is our firm's data used to train AI models?
No. Your data is never used to train AI models, never sold, and never shared for marketing. Each firm's data is isolated, and you can export it at any time.
Can we turn AI off completely?
Yes. A master AI switch provides firm-wide pause and disable controls for all AI operations, with immediate effect across every agent — useful during audits, incidents or simply a policy change.
Which compliance frameworks does FernLex reference?
ABA Model Rules 1.6 and 1.1, state bar technology competence standards, GDPR and CCPA, with the NIST Cybersecurity Framework as the stated control baseline for security documentation.