What AI can do · Legal & accounting
AI in Law Firms and Accounting Practices
Your profession has more rules about this than most, and they matter: the firms that got into trouble were not punished for using AI, they were punished for not checking it. Worth knowing where the profession actually is: AI use among legal professionals reached 79 percent in 2025, while use of legal-specific AI tools fell to 40 percent from 58 percent. People are not deciding whether to adopt. They already have, on consumer tools, without governance.
14 things AI is doing here
- 01Winning work3
- 02The work your practice area actually does5
- 03Doing the work5
- 04Looking after clients1
Proven 6Working 7Early 1
Hear one working
This is intake, answering the night of the accident.
Not a recording and not a video of one. Press call and describe an accident the way a real caller would, to a personal injury firm we set up to show you. It takes the intake in order, runs the conflicts question, and books the free consultation without giving one word of advice.
- Say you were rear-ended on the freeway and see what it asks first.
- Ask what your case is worth. It will not tell you, however you ask.
- Ask about a divorce and watch it refer you out instead of stalling.
After-hours line
BlackSig Injury Law
Claire, new client intake
0:00
It is a demo, so the firm is made up. Calls are capped at three minutes.
Tap any of these to see what it does and what it takes. Add the ones that sound like your business and we will show you the system they add up to.
Winning work
Becomes Client Intake System
Answer the enquiries that currently go unanswered
Picks up the phone or the web chat after hours and while your intake person is on another call, takes the situation and the details, and books the consultation.
- What changes
- How many prospective clients reach a human at all. In Clio’s 2024 study, only 40 percent of firms answered a call from a prospective client, down from 56 percent in 2019, and of the firms that missed it, only 20 percent called back.
- What it takes
- A number to route, a script covering what you do and do not take on, a conflicts hand-off, and a hard rule that it never gives advice.
For exampleA prospective client calls at 6:30pm after being let down by another firm. They get a real conversation, a conflicts hand-off and a Tuesday consultation.
The boring one, and the one most reliably paid for. Note that it must never answer a question that amounts to legal or tax advice, which is a rule you set, not a setting.
Follow up properly instead of once
Reads each enquiry, scores it against your own criteria, drafts a reply specific to their situation, and keeps following up until they book or say no.
- What changes
- How many enquiries get a real answer. In the same Clio study, among firms that did reply, 84 percent replied within eight hours but only 18 percent gave clear next steps or any cost information.
- What it takes
- Your intake criteria written down, which most firms have never done, and a partner willing to let a good draft go out after a glance.
For exampleThe enquiry that came in during a trial gets a real reply the same day, and three follow-ups over two weeks, rather than one email and silence.
Bill the hours you already worked
Watches the work as it happens, documents opened, emails sent, calls made, and proposes the time entry with a narrative a client will accept, instead of you reconstructing Tuesday on Friday.
- What changes
- Hours that were worked and never recorded, and write-downs caused by narratives clients argue with. Clio’s 2025 benchmarks put utilisation at 38 percent and realisation at 88 percent, which compounds to about 2.6 billed hours from an eight-hour day.
- What it takes
- Passive capture running on every fee earner’s machine, which is a cultural fight in some firms, and a billing partner who will accept drafted narratives.
For exampleFriday afternoon, the time entries are already drafted from the week’s actual documents and calls, with narratives a client will not query.
Under ABA Formal Opinion 512, you cannot bill the client for the hour the AI saved you. This use case is about capturing real work, not inflating it.
The work your practice area actually does
The generic list stops here. These are the jobs that only exist in certain practices, and they are usually the ones eating the most unbilled hours.
Becomes Matter Workflow System
Build the medical chronology, not a summary
Turns a records production into a defensible dated table: every treatment, provider, finding and gap, with a page cite against each line.
- What changes
- The completeness of the record a case is valued on, and the days a paralegal spends building it. A summary is prose; a chronology is evidence.
- What it takes
- The records in machine-readable form, a defined column set, and a paralegal who verifies the cites.
For exampleFour thousand pages of records come back as a chronology with the three-month treatment gap defence counsel would have found first.
Keep the lien ledger straight before disbursement
Tracks every lien and reimbursement claim on a settlement, drafts the reduction requests, and chases the statuses, in the right order of priority.
- What changes
- Time from settlement to disbursement, and how many liens actually get reduced. It is the part of a personal injury practice nobody markets and everybody bleeds on.
- What it takes
- The correspondence, the plan documents, the provider statements and your case management system. The order matters: Medicare first by statute, Medicaid second, then plan claims where the plan document, not the lien letter, is what governs.
For exampleA settled case sits for eleven weeks waiting on a lien nobody chased. Next time the ledger flags it at week one.
The right answer here is a limit, not a capability. Funds subject to someone else’s claim sit in trust until resolved, and disbursing early is how lawyers lose licences rather than cases. AI builds the ledger and drafts the letters. It never touches the disbursement.
Draft the privilege log and keep the certification yours
Proposes log entries from the document set, with the basis for each withholding, ready for an attorney to check.
- What changes
- The single most hated task in litigation, and the one most likely to be done at 2am before a deadline.
- What it takes
- The review platform, a written protocol for what gets reviewed by a person, and a partner who understands the signature is theirs.
For exampleA 900-entry log comes back drafted, and the team spends its time on the eighty contested entries rather than the other 820.
Your signature on a discovery response is a certification you made a reasonable inquiry. Any workable protocol reviews every contested entry by hand and samples the rest.
Follow the money in a financial disclosure
Reads years of statements and builds the flow of funds: transfers between accounts, spending that does not match the declared income, the assets that appear late.
- What changes
- What gets found at all, in a case where the other side controls the paperwork.
- What it takes
- Statements in readable form and someone who knows what a suspicious pattern looks like.
For exampleThree years of statements show a monthly transfer to an account that was never disclosed.
Substantiate the credit study the way the form now wants it
Assembles the business component detail a research credit claim requires, in the structure the current form demands, from the project records you already keep.
- What changes
- Whether a claim survives examination, and how much of the study is rebuilt from scratch each year.
- What it takes
- Project and payroll records, and a preparer who signs off on the technical conclusions.
For exampleThe credit study gets assembled from timesheets and project notes rather than a fortnight of partner interviews.
Doing the work
Becomes Drafting and Review System
Read the whole stack and pull out what matters
Works through a contract set, a discovery production, a lease portfolio or a box of client records and puts the specific facts, dates, clauses or figures into a table.
- What changes
- The hours a junior spends reading to find things, and how much gets found at all. This is the clearest win in both professions, because the task is finding defined things rather than writing new ones.
- What it takes
- Documents a machine can read, and a defined list of what you are extracting, which is the actual work. Scanned paper with handwriting is still the ceiling.
For exampleSixty leases, and by the end of the day you have every rent review date, break clause and assignment restriction in one table.
In accounting the same thing shows up as Excel-native audit matching, already standard equipment in many audit practices.
Draft the routine document from your own precedents
Produces the first version of the engagement letter, demand letter, discovery response, standard motion or tax advice letter.
- What changes
- Time from instruction to first draft. Review time does not fall, and firms that assume it will are the ones that get burned.
- What it takes
- Your own precedent bank. Drafting from a generic model produces generic documents a partner rewrites entirely, which is a net loss.
For exampleThe engagement letter for a new matter comes back drafted from the version your firm actually uses, not from a generic model.
Routine and templated work only. Not for anything novel, adversarial or high value.
Get oriented on an unfamiliar question fast
Answers a research question in plain language against a licensed legal or tax database and comes back with citations.
- What changes
- Time to a first orientation. Among tax professionals, 60 percent now use AI for research at least weekly, up from 33 percent a year earlier.
- What it takes
- A paid research subscription, and the discipline to open every citation.
For exampleA client asks about an area you touch twice a year. Half an hour later you have the shape of the answer and the citations to check.
This is the use case that gets people sanctioned. A Stanford study found leading legal research tools still produced wrong answers between 17 and 33 percent of the time. The free-model version of this is how fake cases end up in filings.
Tie the numbers to the evidence
Links each supporting document to the workpaper, agrees figures across statements and source evidence, and flags what does not match.
- What changes
- The largest block of junior time in a small audit, and the audit trail itself, because the link between a number and its evidence becomes recorded rather than a pencil tick.
- What it takes
- Workpapers and supporting documents in the same place, and a methodology that documents where automated procedures were used.
For exampleThe tie-out that took a junior three days is done overnight, and the eleven figures that do not agree are the only ones anyone looks at.
Code the transactions and explain the breaks
Proposes the ledger code for each transaction based on how you coded that client before, matches the ledger to the bank, and drafts the explanation for each exception.
- What changes
- Bookkeeper minutes per client per month, and the size of the exception list a human has to look at. The value is in the flagging; the matching was already automated years ago.
- What it takes
- Clean bank feeds and a stable chart of accounts. Firms that standardised their charts get far more out of this than firms with one per client.
For exampleThe month-end reconciliation comes back with 900 lines matched and the six genuine breaks explained in plain English.
The "30 percent faster close" and "80 percent less bookkeeping labour" numbers circulating this year trace back to no primary source at all.
Looking after clients
Becomes Client Care System
Answer "where is my return" without a partner writing it
Drafts the reply to the routine client email, where is it up to, what do you need from me, what happens next, from the matter or job record, for a human to send.
- What changes
- A large, invisible and entirely unbillable block of correspondence.
- What it takes
- Client and job status living in one system the tool can read. Where status lives in someone’s head or a spreadsheet, this does not work.
For exampleThe fourth "any update?" email of the day comes back drafted with the real status from the job record, ready to send.
Drafting, not sending. Nothing goes out unread.
What you just built
Nothing picked yet.
Add anything above that sounds like your business. They group into the systems we build, so by the time you reach the form you are sending us a scope rather than a question.
Working with client data
How we handle privileged and client information.
Your rules are stricter than most, and they are the first thing a managing partner asks about. Here is how we work, before you ask.
- Nothing client-identifying goes into a consumer toolSystems that touch client matters run on infrastructure with the right agreements behind them, not on somebody’s personal AI account.
- Tax information gets its own treatmentPutting identifiable return information into a general AI service can be a criminal disclosure without specific signed consent naming the recipient. We design around that rather than hope.
- The human signature stays where the rules put itCertifications, disbursements and filings are signed by the person the rule names. We build the draft and the check, never the signature.
- You get the protocol in writingWhat was reviewed by a person, what was sampled, and what the system did. That document is what makes the work defensible later.
Being straight with you
What AI is not doing here yet.
Client data cannot go into a general AI tool
For tax work specifically, putting identifiable client return information into a general-purpose AI service can be a criminal disclosure under IRC section 7216 without specific signed consent. ABA Formal Opinion 512 may also require a client’s informed consent before their information goes in, and says boilerplate in the engagement letter is not enough.
It cannot be the one checking the work
In the sanctions cases so far, what got punished was supervision, not use. In one, two of the three lawyers sanctioned never touched AI: one was fined for signing a filing he had not read. Another firm had a written AI policy and it did not save them.
There is no single answer on disclosure
Whether you have to tell a client you used AI depends on your state or jurisdiction, and the guidance genuinely differs. Anyone who gives you one confident national answer has not read it.
Where to start
Which of these is worth doing first in your business?
That is the whole question, and it is what a roadmap answers. Tell us how your business runs and we will come back with what we heard and where we would start.
See another kind of businessMessage sent.
We'll get back to you within 24 hours.