# Shadow AI in professional services firms

> Shadow AI in professional services firms: what it is, how to find it, and how to fix it with approved tools, an AI policy and training that protect client data.

Source: https://www.assume-ai.com/shadow-ai-professional-services

Shadow AI is the use of AI tools at work that the firm has not approved and cannot see: personal ChatGPT or Claude accounts, browser extensions, meeting note-takers, AI features switched on inside software you already pay for. In a professional services firm it matters more than most, because what goes into those tools is usually client material. The fix is not a ban. It is finding what is in use, giving people an approved tool that is better, writing the rules down and training everyone on them.

## What is shadow AI, and how common is it?

Shadow AI is any AI tool used for the firm's work without the firm's approval or visibility. The term borrows from shadow IT, the older problem of staff signing up for software that nobody in charge knew about. The difference is what the tool does with what it is given. A shadow file-sharing app stores a document; a shadow AI tool reads it, may keep it, and on some consumer plans may use it to train future models.

It takes four common forms. Personal accounts on consumer chatbots, used to summarise a client report or redraft a letter. Browser extensions and meeting note-takers that someone installed because they were free. AI features that arrive inside software the firm already subscribes to, switched on by a supplier update rather than by a decision. And free tools whose terms allow the provider to train on what is typed into them.

It is widespread. Microsoft and LinkedIn's 2024 Work Trend Index found that 75% of knowledge workers already use AI at work, and that 78% of those users bring their own AI tools rather than waiting for their employer to provide them. In most firms of 20 to 200 people the question is not whether shadow AI exists, but how much client material has already passed through it.

None of this is usually malicious. People reach for these tools because they save real time on real work, and nobody has offered them a better approved option. That matters for the fix: shadow AI is a sign of unmet demand, and treating it purely as misconduct tends to drive it further out of sight.

## Why is shadow AI a bigger risk in professional services?

Professional services firms are trusted with information their clients would share with nobody else: financial positions, disputes, people matters, deal terms, unpublished strategy. When that material goes into a personal AI account, the firm loses control of it. There is no contract with the provider, no audit trail, no say over how long it is kept, and nothing to show a client, an auditor or a professional body if they ask what happened to it.

The terms differ sharply between consumer and business plans. OpenAI says it does not use data from ChatGPT Business, ChatGPT Enterprise or its API platform to train its models by default, while on the Free, Plus and Pro plans data sharing for training is on by default and each user has to opt out. Anthropic changed its consumer terms in 2025 so that chats on Claude Free, Pro and Max can be used for training if the user allows it, and kept commercial plans such as Claude for Work out of that change. The same model can therefore be safe or unsafe for client data depending only on which account it is used through.

Data protection law applies too. Where personal data is involved, the firm needs a lawful basis for the processing, and the Information Commissioner's Office says that in the vast majority of cases the use of AI will involve processing likely to result in a high risk, which triggers the legal requirement for a data protection impact assessment. Nobody carries out an impact assessment for a tool they do not know is being used.

The cost is measurable. IBM's Cost of a Data Breach Report 2025 found that one in five organisations studied reported a breach due to shadow AI, that organisations with high levels of shadow AI saw breach costs on average $670,000 higher, and that only 37% had policies to manage AI or detect shadow AI. For a firm whose business rests on discretion, the reputational cost of a client finding out can be larger still.

## Why does banning AI tools not solve shadow AI?

A ban feels like the safe answer, and sometimes a temporary one is right while a firm works out its approach. Samsung restricted staff use of generative AI tools in 2023 after engineers uploaded sensitive code to ChatGPT. But as a permanent policy a blanket ban has two problems.

First, it does not remove the demand. People who have found that AI saves them an hour on a report will keep using it, on a personal phone or a home laptop, where the firm can see even less than before. A ban just pushes usage underground. Second, it leaves the firm behind competitors who have found a safe way to use the same tools, and it tells the people most keen to improve how the firm works that their initiative is unwelcome.

The firms that get shadow AI under control do the opposite. They make the approved route easier and better than the unapproved one, and then they are clear about the rules.

## How do you find shadow AI in your firm?

Start by asking, and make it safe to answer. A short, no-blame survey asking which AI tools people use, for which tasks and with what kind of information will surface more in a week than any technical scan, and it tells you which tasks your approved tool needs to handle well. Then check the evidence the firm already holds:

- Expense claims and card statements, for AI subscriptions bought on personal or company cards
- The third-party apps people have connected to Microsoft 365 or Google Workspace with their work accounts, which your administrator can list
- Browser extensions installed on firm devices, especially note-takers and writing assistants
- The AI features available, and switched on, inside the software you already pay for, from your office suite to your CRM and practice management system
- Calendar invitations that include an AI meeting assistant nobody in the firm set up

Record what you find in a simple register: the tool, who uses it, for what, what kind of data goes into it, and which plan it is on. That register becomes the starting point for the risk register that sits behind your AI governance. If you want a quick, structured view of where the firm stands overall, the free [AI maturity assessment](https://assess.assume-ai.com/ai-maturity) takes about two minutes and benchmarks your firm against comparable firms.

## How do you fix shadow AI without slowing the firm down?

Shadow AI is fixed by three things done together. Any one on its own leaves a gap: a tool without rules invites misuse, rules without a tool invite workarounds, and neither works if people do not understand them.

### Approve a better tool

Give people one secure platform that is better than what they were using. [Enterprise AI Deployment](https://www.assume-ai.com/services/enterprise-ai-deployment) puts Claude Cowork or ChatGPT Enterprise on EU-based data hosting, configured to your data governance policies, with six to twelve agents and skills built around your real workflows, so people open it on day one and the useful work is already there. It is £20,000 excluding VAT, with a full refund if you are not happy for any reason.

### Write the rules down

Write down what may and may not go into AI, and who decides. [AI governance implementation](https://www.assume-ai.com/services/ai-governance-implementation) produces internal AI policies for your teams, external AI policies for clients and partners, impact assessments for what you deploy, and audit trails including a risk register and an incident log. It typically takes two to four weeks and starts from £3,000. The free [AI governance template](https://www.assume-ai.com/resources#governance) is the policy and controls template we use with clients, and our free [AI policy template](https://www.assume-ai.com/resources/ai-policy-template) gives you a written acceptable use policy in thirteen sections, with guidance notes, to adapt for your firm.

### Train people on them

Rules only work if people understand them and can see how the approved tool helps with their own work. [Tailored AI team training](https://www.assume-ai.com/services/ai-team-training) covers Claude, Cowork, Gemini and prompt engineering fundamentals, built around your real work, from £5,000. Firms with staff or clients in the EU should also note Article 4 of the EU AI Act, which has applied since 2 February 2025 and asks providers and deployers of AI systems to take measures on the AI literacy of their staff.

## How should a firm handle client data in AI tools?

AI data security in professional services comes down to a few decisions, made once, written down and applied consistently. The most useful is a simple classification of the information your people handle, with a clear rule for each class:

- Public information, such as published reports and your own marketing: fine in any approved tool
- Internal firm information, such as templates, methods and internal notes: approved business platforms only
- Client confidential material: approved platforms only, within the client's own terms, and only for the tasks your policy allows
- Special category and regulated data, such as health, criminal records or anything a client has restricted: only where an impact assessment has been carried out and the use signed off

Around that sit the controls. Use business or enterprise plans whose terms rule out training on your data, and set retention to match your own policy. Keep access tied to the people who work on each client. Carry out an impact assessment before a new use goes live. Keep an incident log, so that if something does go into the wrong place the firm records it, deals with it and learns from it. And tell clients how you use AI: an external AI policy answers the question before a client has to ask it.

On the data itself, we do not build on top of your business, we build inside it. Everything lives in your existing infrastructure and you own all of it, code included.

## What does good AI governance look like in a regulated profession?

Our [governance work for Hillbreak](https://www.assume-ai.com/work/hillbreak-ai-governance-foundations), a specialist ESG and sustainability advisory firm for the real estate sector, shows how this works where a professional body sets the rules. The RICS professional standard on the responsible use of AI in surveying practice took effect on 9 March 2026 and introduced specific, documented obligations. Hillbreak's existing policy was principles-based and a solid foundation. We assessed it against the standard, identified twelve new provisions and prepared a nine-document pack, taking the updated policy from four rules to thirteen.

A later [impact assessment workshop](https://www.assume-ai.com/work/hillbreak-impact-assessment-workshop) completed a worked example of all nine sections of the template together, with every risk rated before and after mitigation, so the team can carry out assessments themselves. Alongside the governance, [Claude Cowork training](https://www.assume-ai.com/work/hillbreak-cowork-training) gave the team an approved tool they knew how to use well: three workshops, with six custom skills built for the team's own documents.

That is the shape of the fix in any professional services firm: rules that fit the obligations you actually answer to, a tool people want to use, and the skills to use it properly. Our page on [AI consulting for professional services](https://www.assume-ai.com/ai-for-professional-services) covers the wider picture, from strategy to development.

## What can a firm do about shadow AI in the next 30 days?

A firm of 20 to 200 people can move from not knowing to having it under control in about a month, using this sequence:

- Week one: run the no-blame survey, check expenses, connected apps and extensions, and start the register of tools in use
- Week two: classify the information your people handle, and decide which tasks AI may help with and which it may not
- Week three: choose the approved platform, on a business plan with the right data terms, and write the internal AI policy that goes with it
- Week four: train every team on the policy and the tool, set up the incident log, and publish an external AI policy for clients

After that, review the register every quarter and whenever a supplier adds an AI feature. Shadow AI does not disappear for good, because new tools arrive every month, but a firm with a clear policy, a good approved tool and trained people will hear about the next one early, from its own staff.

## Questions

### What is shadow AI?

Shadow AI is the use of AI tools for work that the firm has not approved and cannot see: a personal ChatGPT or Claude account, a browser extension, a meeting note-taker, or an AI feature switched on inside software the firm already pays for. It is usually well meant. People use it because it saves them time and nobody has offered them a better, approved option.

### Can our staff put client information into ChatGPT?

Not into a personal account. Consumer plans such as ChatGPT Free, Plus and Pro, and Claude Free, Pro and Max, can use conversations to train models, depending on a setting each user controls, and the firm has no contract, no audit trail and no control over retention. Business plans such as ChatGPT Business or Enterprise and Claude for Work do not train on your data by default. Even on an approved platform, client material should only go in under a written policy, and personal data needs a lawful basis and, in most cases, a data protection impact assessment.

### Should we just ban AI tools?

A blanket ban rarely works. The demand does not go away, so usage moves to personal phones and home laptops, where you can see even less. A better answer is to approve one secure platform that does the job well, write down what may and may not go into it, and train people on the rules.

### Do we need a data protection impact assessment to use AI?

Usually, yes. The ICO's guidance on AI and data protection says that in the vast majority of cases the use of AI will involve processing likely to result in a high risk, which triggers the legal requirement for a DPIA. If you decide a particular use is not high risk, you still need to record how you reached that conclusion. Our governance work includes impact assessments for what you deploy.

### How long does it take to get shadow AI under control?

The governance itself, meaning internal and external AI policies, impact assessments, a risk register and an incident log, typically takes two to four weeks and starts from £3,000. Deploying an approved enterprise platform such as Claude Cowork or ChatGPT Enterprise on EU-based data hosting is £20,000 excluding VAT. Training runs from a single session to a sustained programme, from £5,000.