# AI agents for professional services > What AI agents are in practice, where they work in consultancies and professional services firms, the risks and controls they need, and build vs buy. Source: https://www.assume-ai.com/ai-agents-for-professional-services An AI agent is software that is given a goal and a set of tools, and decides for itself which steps to take to reach it. In consultancies and professional services firms, the agents that pay back today are narrow ones: they draft the proposal, process the data, set up the project or answer the enquiry, and a person checks the work before it reaches a client. This guide covers what agents are in practice, where they work, the risks and controls, and whether to build or buy. ## What is an AI agent, in practice? The word agent is used for almost anything with a language model in it, so it helps to start with a clean definition. Anthropic, the company behind Claude, separates two designs. Workflows are systems where language models and tools are orchestrated through predefined code paths. Agents are systems where the model dynamically directs its own process and decides which tools to use. Its advice to builders is to find the simplest solution possible and only add complexity when it is needed. That distinction matters more in a professional services firm than almost anywhere else. Your work is checked, signed and sometimes regulated. A workflow that takes a client brief, pulls the right templates, drafts a proposal and hands it to a partner for review is predictable, testable and easy to explain. A fully autonomous agent that decides on its own what to send to whom is none of those things, and most firms do not need one. Be wary of the label as well. Gartner has warned of agent washing, the rebranding of existing chatbots and automation tools as agents, and estimates that only around 130 of the thousands of vendors selling agentic AI are offering the real thing. When a supplier says agent, ask what the system is allowed to do without a person, which systems it can write to, and how you would see what it did yesterday. In practice, most useful agents in a firm combine the same parts: a clear goal, a small set of tools it is allowed to use, your firm's templates and house rules written down as skills, a point where a person approves anything that leaves the building, and a record of every step it took. ## Where do AI agents work in consultancies and professional services firms? Agents earn their place first on work that is repeatable, rule-bound and checked by a person anyway. These are six places where they tend to pay back, each with an example from our own client work where we have one. ### Proposals and bids Turning a brief into a first draft in the firm's own structure and voice. [Clear AIMS](https://www.assume-ai.com/work/clear-aims-proposal-generator) asked us for a proposal generator, and it took the average proposal from two to three hours of work down to five to ten minutes. ### The analytical grind Processing large volumes of data that a consultant would otherwise work through line by line. At [The Litmus Partnership](https://www.assume-ai.com/work/litmus-partnership-transformation-programme) a benchmarking platform now handles 2,000 lines in around 45 minutes, where 200 to 300 lines took eight hours, and monitoring work that took 30 to 40 hours takes two to three. ### Project set-up and the back office The administration every firm runs and nobody chose. For [Coaching Without Borders](https://www.assume-ai.com/work/coaching-without-borders-transformation-programme), one form now sets up a new project in four systems in about ninety seconds, and survey responses become a finished slide report automatically. ### The firm's own knowledge Years of proprietary tools, talks and articles, made available on demand. [The Consultancy Growth Network](https://www.assume-ai.com/work/consultancy-growth-network-ai-board-advisor) now gives every member an AI board advisor trained on more than 400 hours of the network's own material. ### Enquiries and bookings Answering, qualifying and booking new enquiries at any hour. This is where agents act with the most independence, so it needs the tightest rules on what they may promise. [Classmaids](https://www.assume-ai.com/work/classmaids-ai-agents) is our clearest example, below. ### Client files and first drafts Summarising a client file, preparing a first-pass report or pulling the facts out of a long document. Work like this suits the agents and skills built in an [Enterprise AI Deployment](https://www.assume-ai.com/services/enterprise-ai-deployment), around the workflows a firm's teams run every week. ## What does a working set of AI agents look like? [Classmaids](https://www.assume-ai.com/work/classmaids-ai-agents) was one of our first clients. It is a Florida-based cleaning company rather than a professional services firm, but it is the clearest example we have of agents doing customer-facing work end to end, and the pattern carries straight across to any firm that takes enquiries. The brief was an AI receptionist that could give quotes, book into the team's calendar and keep the team updated, followed by an AI automation roadmap, delivered in five sprints over eight weeks. What went live was a set of cooperating agents rather than one: a 24/7 AI receptionist connected to the CRM, giving instant quotes and taking real-time bookings on every call; an agent that answers and qualifies enquiries from review sites; a website booking widget and a text message agent working from the same quotes, availability and bookings; and automations that send call reports by email and follow-up quotes. Each agent has a narrow job and works from the same source of truth, the CRM, rather than its own copy of the facts. Calls are now answered around the clock and the admin work across the business is automated, so the owner can spend her time on the high-value work. In a consultancy or a law, accounting or surveying firm, the same design can answer and qualify new enquiries, book the first call and brief the partner, without ever giving advice. The other examples in this guide are mostly workflows rather than free-roaming agents, and that is deliberate. At The Litmus Partnership, every figure was collected from the people who use the tools, and the programme saved 268 hours a month. None of it asks the client to accept a lower standard of output. ## Where should AI agents not act alone yet? The failures so far have a common shape: a model produced something plausible, nobody checked it, and it went out under a professional's name. In June 2025 the Divisional Court of the High Court of England and Wales, in Ayinde v London Borough of Haringey, warned lawyers that they have a professional duty to make sure fictitious citations produced by generative AI do not reach the court, and that those who fail risk severe sanction. In October 2025 Deloitte agreed to partially refund the Australian government for a report that contained a fabricated quote from a court judgment and references to research papers that do not exist. Customer-facing agents carry the same risk in a different form. In Moffatt v Air Canada, a Canadian tribunal held the airline liable for incorrect advice its website chatbot gave about bereavement fares, and called its argument that the chatbot was responsible for its own actions a remarkable submission. The regulators who have approved AI-led services have drawn the line in the same place. When the Solicitors Regulation Authority authorised Garfield.Law in May 2025 as an AI-driven law firm, it noted that the system will not propose relevant case law, a high-risk area for language models, that it will only take a step the client has approved, and that named regulated solicitors remain accountable for its output. So, for now, keep agents away from three things unless a qualified person checks every output: citing authorities or sources, giving advice a client will rely on, and sending anything external without approval. ## What controls does an AI agent need in a professional services firm? The security community has a name for the central risk. The OWASP Top 10 for Large Language Model Applications calls it excessive agency, and traces it to three root causes: excessive functionality, excessive permissions and excessive autonomy. Its mitigations include keeping an agent's tools to the minimum, executing actions in the user's own context and requiring user approval before high-impact actions. The UK Information Commissioner's Office adds the data protection view: organisations remain responsible for the compliance of the agentic AI they develop, deploy or integrate, and it lists agents connected to databases they do not need among the signs of a poorly implemented system. In practice that comes down to six controls, all of which should be in place before an agent goes live: - Least privilege. The agent can read and write only what its task needs: one folder, one mailbox, one table, read-only wherever possible. - Human approval. Anything that leaves the firm, changes a client record or spends money waits for a person to approve it. - Data minimisation. Client data stays on an enterprise platform configured to your data governance policies, and the agent sees only the fields it needs. - An audit trail. Every step is logged, so you can see what the agent did, with what data, and why. - An impact assessment before use. A written assessment of each AI tool, with every risk rated before and after mitigation, as in the worked example we completed with [Hillbreak](https://www.assume-ai.com/work/hillbreak-impact-assessment-workshop). - A named owner. One person is accountable for each agent, its instructions and its results, and can switch it off. If your firm answers to a professional body, add its rules to the list. Our [AI governance implementation](https://www.assume-ai.com/services/ai-governance-implementation) work, from £3,000, writes the policies, impact assessments and audit trails that sit behind these controls; for Hillbreak we assessed its policy against the [RICS standard](https://www.assume-ai.com/work/hillbreak-ai-governance-foundations) on the responsible use of AI in surveying practice. ## Should you build or buy AI agents? Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. The way to avoid joining them is to decide what the agent is for, and what it is worth, before deciding how to get it. There are three routes, and most firms end up using more than one of them. ### Configure: agents on your AI platform Claude Cowork and ChatGPT Enterprise both let a firm run agents and skills on an enterprise platform. This is the right route when the work lives in documents, email and spreadsheets. Our [Enterprise AI Deployment](https://www.assume-ai.com/services/enterprise-ai-deployment) puts either platform on EU-based data hosting, configured to your data governance policies, with six to twelve agents and skills built around your real workflows, for £20,000 excluding VAT. ### Build: agents inside your own systems Build when the agent has to act inside your own systems, such as a CRM, a practice management tool or a finance package, or when it carries your firm's intellectual property and has to behave the same way every time. A [30-day sprint](https://www.assume-ai.com/services/30-day-sprint), from £6,000, builds one automation, assistant or agent for one clear workflow; a [90-day sprint](https://www.assume-ai.com/services/90-day-sprint), from £15,000, builds and embeds several. You own what we build. ### Train: your people build their own The cheapest agents are the ones your own people write. In our [Claude Cowork training](https://www.assume-ai.com/work/hillbreak-cowork-training) for Hillbreak we built six custom skills for its own work, along with the projects and guides the team works from. [Tailored AI team training](https://www.assume-ai.com/services/ai-team-training) starts from £5,000. ## How should a professional services firm start with AI agents? Start with the work, not the technology. The order we recommend applies to a single agent as much as to a programme: - Pick one workflow that is repeatable, rule-bound and already checked by a person, such as proposals, data processing, project set-up or enquiry handling. - Count the hours it takes today, and put a cost against them, so you know what the agent is worth before you build it. - Put the controls in place first: the platform, the permissions, the approval step and a written impact assessment. - Build the smallest version that does the job, usually a workflow rather than a fully autonomous agent, and test it with the people who will use it. - Measure the result with those same people, then decide whether to widen its job or move on to the next workflow. If you are not sure where your firm stands, the free [AI maturity assessment](https://assess.assume-ai.com/ai-maturity) (some people call it an AI readiness assessment) takes about two minutes and gives you a scored report. For a wider view of AI across a firm, see our guide to [AI for professional services](https://www.assume-ai.com/ai-for-professional-services), and for why agents need checking as much as trusting, our post on [AI agents being trusted faster than they are being checked](https://www.assume-ai.com/blog/ai-agents-are-being-trusted-faster-than-they-are-being-checked). ## Sources for the facts in this guide Every outside fact on this page comes from the sources below, checked on 21 October 2026. Facts about AssumeAI's work come from our published case studies and service pages. - [Anthropic, Building effective agents (December 2024)](https://www.anthropic.com/engineering/building-effective-agents): the definitions of workflows and agents, and the advice to find the simplest solution possible. - [Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (25 June 2025)](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027): the cancellation forecast, its three causes, and agent washing. - [OWASP, LLM06:2025 Excessive Agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/): the three root causes and the mitigations. - [Information Commissioner's Office, ICO tech futures: Agentic AI](https://ico.org.uk/about-the-ico/research-reports-impact-and-evaluation/research-and-reports/technology-and-innovation/tech-horizons-and-ico-tech-futures/ico-tech-futures-agentic-ai/): organisations remain responsible for the data protection compliance of agentic AI. - [Ayinde v London Borough of Haringey, 2025 EWHC 1383 (Admin)](https://caselaw.nationalarchives.gov.uk/ewhc/admin/2025/1383): the Divisional Court's warning on fictitious citations. - [Associated Press via CTV News, Deloitte to partially refund Australian government for report with apparent AI-generated errors (7 October 2025)](https://www.ctvnews.ca/business/article/deloitte-to-partially-refund-australian-government-for-report-with-apparent-ai-generated-errors/). - [McCarthy Tetrault, Moffatt v. Air Canada: a misrepresentation by an AI chatbot (2024 BCCRT 149)](https://www.mccarthy.ca/en/insights/blogs/techlex/moffatt-v-air-canada-misrepresentation-ai-chatbot): liability for a website chatbot's advice. - [Solicitors Regulation Authority, SRA approves first AI-driven law firm (6 May 2025)](https://www.sra.org.uk/sra/news/press/garfield-ai-authorised/): the safeguards on Garfield.Law. ## Questions ### What is the difference between an AI agent and an automation? An automation follows steps someone wrote in advance. An AI agent is given a goal and some tools, and decides for itself which steps to take and in what order. Most of the systems sold as agents today are really workflows with a language model inside them, and for most professional services work that is the better design: it is cheaper, easier to test and easier to explain to a client or a regulator. ### Are AI agents safe to use with confidential client data? They can be, if the agent runs on an enterprise platform configured to your data governance policies, can reach only the data its task needs, and keeps a record of what it did. The UK Information Commissioner's Office is clear that organisations remain responsible for the data protection compliance of the agentic AI they develop, deploy or integrate, so the controls are yours to set, not the vendor's. ### Who is liable when an AI agent gets something wrong? The firm that deployed it. In Moffatt v Air Canada a Canadian tribunal held the airline responsible for what its website chatbot told a customer, and rejected the argument that the chatbot was a separate entity. In professional services the duty is sharper still: the work goes out under the firm's name, so a person has to check anything an agent produces before it reaches a client, a court or a regulator. ### How much does it cost to put AI agents into a professional services firm? At AssumeAI, a 30-day sprint that builds one automation, assistant or agent for one clear workflow starts from £6,000. An Enterprise AI Deployment of Claude Cowork or ChatGPT Enterprise, with six to twelve agents and skills built around your real workflows, is £20,000 excluding VAT. A 90-day sprint for several workflows starts from £15,000. ### Should we build our own AI agents or use the ones in Claude or ChatGPT? Start with the platform if the work happens in documents, email and spreadsheets your people already handle: agents and skills configured inside Claude Cowork or ChatGPT Enterprise cover a great deal of it. Build when the agent has to act inside your own systems, such as your CRM, practice management or finance tools, or when it carries your firm's intellectual property and has to behave the same way every time.