Industries
AI for professional services
Permission-aware knowledge search, contract review and audit support for law, accounting and consulting firms. A person checks every citation before it reaches a client.
The sector
What's different here
Law, accounting and consulting firms sell expertise that is mostly written down already, in precedent banks, matter files, workpapers and past reports. Finding the right document is often slower than reading it, which is why knowledge search is usually the first thing that pays off.
The failure mode is well documented. Courts have now dealt with hundreds of filings containing AI-fabricated citations, and even legal-specific research tools hallucinated on a meaningful share of test queries. Every citation needs a human check, and the system has to make that check quick.
Confidentiality shapes the architecture. Ethical walls, DMS permissions, privilege and client outside-counsel guidelines all decide who may see what, and retrieval has to enforce them on every query instead of relying on a prompt.
17–33%
of benchmark queries produced hallucinations in Lexis+ AI, Westlaw AI-Assisted Research and Ask Practical Law AI, across 200+ preregistered queries. The vendors disputed parts of the method.[1]
712
court decisions worldwide dealt with AI-hallucinated content by late Dec 2025, about 90% of them written in 2025. An estimate from researcher Damien Charlotin's database, as reported by Bloomberg Law.[2]
22%
of professionals say their organisation has a visible, defined AI strategy. Thomson Reuters survey of 2,275 professionals across legal, tax, accounting and risk.[3]
25.1%
faster, with 12.2% more tasks completed and about 40% higher quality, for consultants using GPT-4 on tasks inside its capability. On a task outside it, AI users did worse. Field experiment with 758 BCG consultants, 2023.[4]
Workflows
Where AI does useful work
Each of these runs inside systems the sector already uses. Steps that carry real risk wait for a person.
Permission-aware knowledge search
Fee earners search precedents, matter files and know-how and get answers that link to the source documents.
- 1Sync documents and their access controls from the DMS
- 2The user asks a question in plain language
- 3Retrieval filters by the user's DMS permissions and ethical walls
- 4The answer cites the documents and pages it used
- 5It declines when the sources don't support an answer
Knowledge and risk teams approve which libraries are indexed, and a fee earner reviews any answer that goes to a client.
DMS (iManage, NetDocuments) · Microsoft 365 · Walls and conflicts (Intapp) · Identity (Entra ID)
Contract review and due diligence
A data room is read against the firm's playbook and turned into a risk table with page-level citations.
- 1Ingest the data room and classify documents
- 2Extract clauses such as change of control, assignment, termination and liability caps
- 3Compare each clause to the playbook
- 4Produce a risk table with citations to the source pages
- 5An associate verifies before anything reaches the client
An associate verifies every row of the risk table, and a partner signs off the report.
Virtual data room · DMS · CLM (Ironclad, Icertis) · Playbook library
Audit document matching
Invoices, purchase orders and ledger entries are matched across the full population, and anomalies are flagged for the audit team.
- 1Load the client's transaction population and supporting documents
- 2Match invoices to POs and to the general ledger
- 3Flag anomalies and unmatched items with the reason
- 4Draft workpaper notes for the matched items
- 5The auditor reviews exceptions and concludes
The engagement team reviews every exception and owns every conclusion in the workpapers.
Audit platform (Caseware) · Client accounting (NetSuite, Xero, QuickBooks) · Document storage · Data analytics tools
Time entry narratives
Draft time entries from calendar, email and document activity for the fee earner to approve.
- 1Collect the day's calendar, email and DMS activity
- 2Group activity by matter
- 3Draft narratives in the client's billing-guideline format
- 4The fee earner edits and approves each entry
The fee earner approves every entry, and only time actually spent is billed.
Practice management (Clio, Aderant, Elite 3E) · Microsoft 365 · DMS
Research and report drafting
Interviews and desk research synthesised into draft reports, with every reference checked before delivery.
- 1Collect interview notes and source documents
- 2Synthesise themes with links back to the sources
- 3Draft sections of the report or deck
- 4Check every reference against its source
- 5The engagement lead approves the final version
The engagement lead approves the final document after the reference check.
Microsoft 365 · Research platforms (Westlaw, Lexis+) · Knowledge base · Reference manager
Systems and rules
What the design has to respect
The systems we expect to connect to, and the regulation that decides what an AI system may do with the data in them.
- DMS (iManage, NetDocuments)
- Practice management and billing (Clio, Aderant, Elite 3E, Intapp)
- CLM (Ironclad, Icertis)
- Research platforms (Westlaw, Lexis+)
- E-discovery (Relativity)
- Tax and audit (CCH, Thomson Reuters UltraTax, Caseware)
- Client accounting (Xero, QuickBooks, NetSuite)
- Microsoft 365
- ABA Formal Opinion 512
- The ABA's first ethics guidance on generative AI (Jul 2024). It covers competence, confidentiality (including when informed client consent is needed), communication, supervision, and fees for time actually spent.
- SRA rules (England and Wales)
- The SRA authorised Garfield.Law as the first purely AI-based firm on conditions: it may not propose case law, it acts only with client approval, and named solicitors remain accountable.
- Privilege, walls and client guidelines
- Legal professional privilege, ethical walls, conflicts and outside-counsel guidelines that restrict AI use all limit what a system may retrieve, and for whom.
- Audit standards
- PCAOB and IAASB standards on audit evidence and documentation, plus independence rules, apply to anything an AI produces for an audit file.
- GDPR and court rules
- Client personal data falls under GDPR, and some courts have standing orders requiring disclosure of AI use in filings.
Risks
What goes wrong, and how we design for it
- Fabricated citations
- Answers come from the firm's own documents and licensed databases, every citation links to its source, and a person verifies them before anything leaves the firm.
- Breaching ethical walls through search
- DMS permissions and walls are enforced at retrieval time, and we test with accounts that sit on the wrong side of a wall.
- Privilege or confidentiality loss through third-party tools
- Models run in the firm's own tenant or under terms that exclude training and retention, checked against client guidelines.
- Over-trust on tasks outside the model's ability
- Each workflow has its own evaluation set, and tasks that fail it stay manual.
- Efficiency that cuts into billable hours
- We involve the firm's pricing team at scoping, since a faster review changes what can sensibly be billed by the hour.
Where to start
Permission-aware search over your own DMS
Firm data stays inside the firm's tenant, which addresses confidentiality and privilege, and answers come from documents the firm trusts instead of the open web. The value shows quickly as time to find precedent. Playbook-based contract review with a human sign-off is the natural second project.
- Connect one practice group's precedents and know-how from iManage or NetDocuments, with permissions
- Enforce DMS access controls and ethical walls at retrieval time, tested with restricted accounts
- Require a citation to document and page for every answer, and decline when the sources don't support one
- Build an evaluation set of real questions with answers agreed by professional support lawyers
- Measure time to find precedent, citation accuracy and weekly active users
In the sector
Publicly reported examples
What others in the industry have announced. These are not our clients.
A&O Shearman and Harvey[5]
Allen & Overy (now A&O Shearman) rolled out Harvey to more than 3,500 lawyers across 43 offices after a trial of about 40,000 queries. The firm said the output needs careful review by an A&O lawyer.
Garfield.Law authorised by the SRA[6]
The SRA authorised Garfield.Law as the first purely AI-based firm to provide regulated legal services, on condition that it does not propose case law and takes steps only with client approval.
Deloitte Australia report refund[7]
Deloitte refunded part of an A$440,000 government contract after a 237-page report was found to contain fabricated references and a fabricated court quote. It disclosed that it had used Azure OpenAI GPT-4o in drafting.
Use cases
Related use cases
- Company knowledge searchStaff ask a question in plain language and get an answer drawn from your own documents, with citations, limited to what each person is already allowed to see.
- Document processingInvoices, forms and contracts are read, checked against your own records and posted to the right system, with low-confidence fields and failed checks sent to a person.
- Back-office workflowsReconciliations, approvals and onboarding are prepared by agents with the evidence attached, then routed to the right approver. Nothing is posted or paid without a person signing off.
Sources
- 1.Stanford HAI, AI on trial: legal models hallucinate in 1 out of 6 (or more) benchmarking queries (2024)
- 2.Bloomberg Law, AI faked cases become core issue irritating overworked judges (Dec 2025)
- 3.Thomson Reuters, Future of Professionals 2025 (Jun 2025)
- 4.BCG, There's a time and place for generative AI (Oct 2023)
- 5.Law Gazette, Magic circle firm rolls out ChatGPT-type platform (Feb 2023)
- 6.SRA, Garfield.Law authorisation news release (May 2025)
- 7.CFO Dive, Deloitte refunds report after AI errors for Australian government (Oct 2025)
Figures are as reported by each source on its publication date. Examples are public industry cases, not AI Sewa clients.
Next step