AI and automation for law, accounting and consulting firms

Start inside the firm, with work your own staff can check. Precedent search and document chasing are good first projects because the people using the output are the people who would spot a mistake. Client-facing drafting comes later, once you have a review habit and an acceptable use policy everyone has read.

The first step we recommend is an AI readiness audit. The snapshot version costs $3,000 to $5,000 and takes 1 to 2 weeks. We interview partners and support staff, look at how matters and client files are stored, and hand you a ranked list of opportunities with a build cost and the risks attached to each.

Be ready for some of that list to involve no AI at all. Chasing documents against a checklist, routing inquiries by practice area and sending reminders are rules. We build those as plain automation because it is cheaper to run and easier to audit. A model is the right tool where the input is unstructured text: past matters, contracts, statements and free-text emails.

If partners want the whole team using AI tools well before any build, a half-day workshop for up to 25 people costs $3,000 to $6,000 and covers your policy, safe handling of client data and how to check a model's output against the source.

This is the shape of a first engagement for a firm of 20 to 150 people building a precedent search assistant. The dates move; the order does not.

The plan stops at week 5 if the pilot misses the bar agreed in week 2. You keep the findings and the code either way.

After week 13, you can move to an automation care plan at $1,500 to $6,000 a month or take the system in-house with the runbook. The contract review assistant blueprint makes a sensible second project, because it reuses the same index and permission model. For accounting firms, the invoice processing pipeline shows the same extract, check and review pattern applied to supplier documents.

Confidentiality is designed in from the first week, as 5 controls you can inspect.

• The search index reads the access rules from iManage, NetDocuments or SharePoint, so the assistant cannot show a document the user could not open themselves.
• Every answer links to the file and passage it used. This is what retrieval-augmented generation means in practice, and it lets a lawyer check the source in seconds instead of trusting a summary.
• Anything bound for a client passes a named reviewer. Our note on human in the loop explains how the approval rules are set.
• Data retention and training terms for each model provider are reviewed with your IT lead before any client file is sent, and the decision is recorded.
• The assistant can read but cannot send email, move files or change records. A document planted with hidden instructions (a risk called prompt injection) has nothing to act on.

On professional conduct, US lawyers often start with the ABA's Formal Opinion 512 (opens in a new tab) from July 2024, then check their state bar. Accountants and consultants have their own bodies. We do not interpret that guidance for you. We build the controls your general counsel or ethics partner asks for, and document them so they can be reviewed.