An accounting firm and a law firm are companies where AI has the most to do and the most to break. The work is text and documents, which is what models do best. And what is easiest to paste into the wrong window.
I say this as someone who deploys agents on companies' documents and runs a company himself with an accountant and a lawyer on the other end of the e-mail.
Eurostat shows that professional, scientific and technical services have the second highest AI adoption in Europe, 40 in 100 companies, behind the information and communication sector at 63 in 100. Thomson Reuters surveyed more than 1,500 people from law, tax and accounting: 40 in 100 organizations use generative AI, up from 22 a year earlier. More than 80 in 100 users use it weekly. AI agents are used by 15 in 100, another 53 are planning to. And the number that says the most: only 18 in 100 organizations measure whether the AI investment works, and of those, three quarters measure only internal metrics such as time saved. Almost nobody measures client satisfaction or revenue impact.
Three processes to start with, from the safest:
- Transcription and notes from client meetings. The recording turns into notes and a task list that lands in the tool you work in. Zero integration with accounting systems, immediate savings, minimal risk. Weeks one and two.
- Classification and first drafts of e-mail replies. A consultant's assistant, not a bot for the client. AI proposes, a human sends. This is one of the most common uses of company APIs according to Anthropic's data.
- Extracting data from documents. Invoices, contracts, PDF orders turned into fields in the system instead of retyping. Vendors report that a three-person office processing 500 to 1,000 documents a month recovers 60 to 80 hours and reduces manual errors by about three quarters, at a cost of 20 to 100 dollars per user per month. Those are their numbers, so count on half. It still pays off.
Only after these three: an assistant on your own documents, procedures and past cases, so people stop searching through e-mails from two years ago. And generating quotes and letters at the end, when the knowledge base is in order.
Now the trap. In legal tasks without verification, models hallucinate in 17 to 33 percent of cases. That is not an argument against AI in a law firm. It is an argument for every answer having a cited source and a human checking it before sending. Demand it from the tool: an answer without a reference to the document and paragraph is worthless.
The second part of the trap is more dangerous. Most accountants and lawyers today use consumer ChatGPT instead of tools with a data processing agreement. A client's data pasted into a consumer window is a breach of confidentiality that no policy will fix. If you do one thing in your company this month, do this: introduce a tool with a contract and forbid pasting client data anywhere else.
Limitation: the processes above take load off specialists, they don't replace their judgment. The client pays you for responsibility, and that can't be rented from a model.
A question for you: do you know which AI tool your people pasted client documents into today?
Want to see how the agent handles your customers' questions? Enter your website address and e-mail in the demo form. In 20 minutes you get an agent trained on your website's offer, ready to be quizzed.
Sources
- Thomson Reuters, 2026 AI in Professional Services Report: https://www.thomsonreuters.com/en/reports/2026-ai-in-professional-services-report
- Eurostat, Use of artificial intelligence in enterprises (2025 data): https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises
- AdAI, Dsm, Truvisory: data from vendors of tools for accounting and law firms (self-reported)