Edit page

AI usage guidance

This document provides guidance about what AI tools it’s okay to use and what you’re allowed to do with them. If you want to do something that isn’t covered here, let us know and we’ll provide some guidance.

This document does not cover building AI into the services we build for clients. For that, see our Development policy.

Core principles #

You are responsible

Clients come first

Be careful what you share

You are responsible for what you share with AI and what it does with that information. Do not share any sensitive, confidential, or proprietary information unless you can be sure it will not be logged. If you are inputting any sensitive information into an AI tool, it must be an account that is owned by dxw or a client.

AI tools #

Approved tools are okay #

You are allowed to use the AI tools listed below. These are tools we have company accounts for.

If you are using an AI tool not on this list, or you have an account that is not administered by dxw, you need to talk to Fred immediately.

Client AI tools and data #

Any tools provided by a client (e.g. Microsoft Copilot) are fine, so long as you only use them with client systems/data.

Putting client data in a tool not provided by that client is prohibited unless you have permission from them. See the section on data below.

If a client is offering you a GitHub Copilot licence, please let Joseph know before accepting.

Local AI models #

Use of local AI models that do not externally communicate is unrestricted, but subject to normal ISMS rules around untrusted software.

Prohibited tools/models are not allowed #

New tools require permission #

If you would like to use a new tool, you need to ask for permission first before putting in any information that might be in any way sensitive. That means you can evaluate tools with dummy data without asking for permission, but you need to ask for permission before using them for real.

You are encouraged to experiment, but safely. AI is new and it’s constantly evolving. You are encouraged to experiment with AI tools to see if they’re useful. Bear in mind that while they might not be quite good enough now, in six months they might be ideal.

Our clients expect that we’ll be able to help them make decisions about how and when to use AI, so the more experience we have of trying it out, the better.

Using tools #

Internal = Slack, Bikeshed, internal docs, drafts/planning docs clients won’t see, code for internal repos/uses, learning & development
External = client work, code and commits for client repos, our public website, the Playbook, public presentations

  Internal External
Explaining things Usually okay Caution
Reviewing things / providing feedback Usually okay Usually okay
Generating ideas Usually okay Usually okay
Generating text Caution Caution
Generating images and diagrams Usually okay Not allowed
Generating video Usually okay Not allowed
Generating code Usually okay Caution
Generating code-related text Usually okay Usually okay
Performing actions (being ‘agentic’) Caution Not allowed
Analysing data Caution Caution
Making decisions Not allowed Not allowed
Assisting with writing Caution Caution

Asking AI for help #

Explaining things #

Asking AI for help by getting AI to explain things to you is fine.

If an AI needs access to content like a codebase in order to do this, then check carefully to make sure it is secure and no secrets will be leaked.

Reviewing things #

Asking AI to review things to provide feedback, suggestions etc. is usually fine. Be careful about what you are sharing for review.

Generating content #

Ideas #

Generating ideas, topics, structures, questions, outlines, etc., anything that’s not the finished product, is usually fine.

Text #

AI-generated text (i.e. finished or near-finished paragraphs) must be carefully reviewed. External-facing content should be primarily human-authored. For code-related text, see the separate section below.

Images #

AI-generated images are fine internally but are not allowed externally. Note that they are much more energy intensive than text generation.

Videos #

AI-generated videos are fine internally but are not allowed externally. Note that they are much more energy intensive than text generation.

Code & prototyping #

AI-generated code is usually fine, but must be carefully reviewed in line with normal dxw ways-of-working.

Code-related text #

Generating code-related text is usually fine, but must be carefully reviewed for accuracy. The volume of content generated should be kept small.

Typical use cases:

This text is an exception to the rule that external-facing content must be primarily human-authored.

Performing actions (‘agentic’ behaviour) #

This is an area where there is a lot of risk. You are allowed to let AI perform actions on your behalf, but only within controlled contexts. You must be familiar with the Lethal Trifecta (access to your private data, exposure to untrusted content, and the ability to externally communicate) and the risks associated with it.

Analysing data and making decisions #

This is an area where there is a lot of risk. We know that AIs have biases and that automated decision making can be unfair and discriminatory. It’s not our area of expertise and we should exercise a lot of caution.

Analysing data #

Using AI to analyse data internally is generally fine, but be very careful of biases.

You can also use AI on client/project data, for example to generate transcripts from user research sessions, summaries from stakeholder interviews etc. but you must ensure that this is done in a way that does not leak any confidential information or breach agreements with clients.

If you are developing AI-enhance services for clients, see development policy.

Making decisions #

Using AI for automated decision-making is not allowed.

Assisting with writing #

Always-on AI writing tools may only be used if they do not log anything. Grammarly is specifically prohibited under our Software Policy.

Using AI tools paid for by dxw for personal use #

For AI tools where usage is unmetered, you may use them for personal use.

For AI tools where usage is metered, you may make reasonable use of them for personal use, so long as the costs are low.

For GitHub Copilot there is currently no way to hold a personal licence in addition to a dxw one. If you are using GitHub Copilot extensively outside of work, please let Joseph know.

Bear in mind that your usage, including prompts and responses, may be monitored by dxw.

Sharing data with AI #

Connecting AI tools to internal systems #

Giving AI tools access to our internal systems is dangerous. If an AI reads untrusted content then it can be compromised by an attacker, and any action it takes could be malicious. AIs sometimes make mistakes even when there is no malicious intent, and can do things like inadvertently delete data when trying to be helpful. It is important that you understand the risks and manage them appropriately.

Wherever possible:

You can give approved AI tools access to read the following internal systems:

If you want to give access to a system not on this list, you need permission first. If you want to give access to a client system you must get approval from the client.

Sharing sensitive data with AI tools #

You are allowed to share dxw’s sensitive data with approved AI tools, but must do so carefully. Wherever practical you should limit the amount of data you are sharing, for example by redacting or anonymising the data.

You are not allowed to share client data with AI tools unless permission has been given by that client, or you have suitably anonymised that data.

When an AI has access to private data and the ability to communicate externally, and is exposed to untrusted content, this is called the ‘lethal trifecta’ and attackers can easily trick the AI into leaking private information. Make sure that you are suitably restricting access to prevent this.

Documents and data within dxw are classified as follows:

Classification Description Examples Risk
Unclassified the most common classification used, for when none of the following apply. The default if no label is used. Public code repositories Low – generally fine
Company Confidential information owned by dxw which would be of value to those outside the company, such as competitors, and whose loss or theft would potentially damage the company. All public Slack channels Google Calendar (public visibility) Productive Most of Google Drive High – make sure the AI tool does not leak data externally
Management-in-Confidence internal documents whose circulation within dxw needs to be restricted. All private Slack channels Google Calendar (full visibility) Gmail Marked documents on Google Drive High – make sure the AI tool does not leak data externally or internally
Client Confidential or Commercial in Confidence information owned by dxw or its clients, which needs to remain confidential between dxw and the client. All project materials including private code repositories Bids, contracts, other commercial documents Zendesk Client-specific Slack channels Very high – not allowed without client permission or redaction

Who to talk to when #

If you need help using a tool:

If you need access to a tool we are already using:

If you want to use an existing tool a new way (e.g. give it permission to read Slack):

If you want to start using a new tool:

Any queries about things not covered here:

Roles and responsibilities #

Aspect People
AI strategy: Deciding what our approach to AI should be Alex Y
Workplace technology policy and guidance: Deciding what’s allowed, making sure people have the tools they need, reviewing costs Joseph, Adelle
Managing licences and access Fred, Joseph (GitHub Copilot only)
Supplier management: Checking terms, particularly around privacy Gurps
ISMS: Making sure our policies and processes meet ISMS requirements Neil, Joseph, ISMS group

Useful websites #

Last updated: 10 August 2026