Capacity, not cuts
We free up time for the work only people can do. We don't take on projects designed to replace your team.
Balance Lab / How we use AI
How we use AI, and where we won't
We help mission-driven organisations use AI to free up time for their mission. This page sets out the rules we hold ourselves to, so you can decide if our way of doing things fits yours.
We can't erase all of AI's harmful impact. We can keep it as low as possible.
The way we work with AI is designed to reduce its harm as far as we can, while still giving mission-driven organisations the chance to do more with their time.
Striking that balance, between helping the people making a difference do more and limiting the harm AI causes, is something we navigate every day.
We also fully respect organisations that decide not to use AI at all.
What we stand for
Four values behind every build
People decide about people
AI can draft, summarise and sort. Decisions about who gets help stay with humans, and anyone talking to an AI can reach a person.
Use only what's needed
We pick the smallest tool that does the job, and avoid AI where simpler fixes work.
Yours to keep
Your data stays protected, your rules stay written down, and your tools stay portable. We take no fees from AI vendors, so our advice is independent.
Red lines
What we won't build, for anyone
These are not negotiable, whatever the budget. If a project needs one of them, we'll say no and explain why.
Scoring or ranking people
No tools that decide eligibility, risk, fraud or priority of support. The Dutch childcare benefits scandal showed what that can do to families.
Emotion recognition
No reading the emotions of staff, volunteers or the people you serve.
Unsupervised crisis support
No chatbots for mental health, abuse or crisis without a human handover.
Made-up people
No generated images or stories presented as real people.
Hidden AI
No tools that talk to people without saying they are AI.
Replacement projects
No builds where the main goal is cutting jobs.
Commitments
Nine things we do on every project
Each one comes with a check you can ask to see. We report on them in every results report, alongside the hours saved.
Right-size every build
The smallest model that does the job. No AI where a rule, template or spreadsheet formula works. No generated images or video unless you ask for them.
You'll see: the smallest setup that does the job, written down in your handover pack.
Keep AI out of decisions about people
AI drafts, summarises and handles routine questions. Where a build touches people's requests, we test it with a varied sample, in every language you work in, with staff who know the community.
You'll see: the test sample and what we changed.
Data classes before tools
In week 1 we sort your data into public, internal, personal and sensitive, and agree which tools may touch each. Personal data only goes into business-tier tools with a data processing agreement and no training on inputs.
You'll see: a one-page data map, and a flag if you need a data protection impact assessment (DPIA).
Capacity, not cuts
The people who do the work co-design the change. Where a tool could monitor staff, we check whether works council consent is needed.
You'll see: time saved reported next to what the team will do with it.
A human route, always
Anyone talking to an AI tool is told so and can reach a person. Public-facing tools hand crisis, health and legal questions to staff.
You'll see: the tested list of topics the tool hands over.
A named person signs off
Anything that leaves your organisation, such as funder reports, donor messages or translations, has a named reviewer built into the workflow.
You'll see: whether sign-off still happens, checked at month 3.
Portable builds, no vendor fees
We take no referral fees or commissions from AI vendors. Prompts, workflows and data exports are documented so you can switch provider, and we offer European or open-model options.
You'll see: a handover pack you could give to anyone.
We say where we used AI
If we used AI in our own work for you, for example to draft a report or analyse survey answers, we tell you, and a person on our side checks it.
You'll see: a note on each deliverable.
Harm counts in every verdict
Our week 6 and month 3 verdicts of keep, adjust or drop cover harm as well as hours saved. If something isn't safe, we recommend dropping it, even if it saves time.
You'll see: incidents, complaints and red flags in the results report.
Your own AI policy
We help you write your own rules
Our rules are a starting point, and your organisation will have its own. Every readiness session includes a short workshop that turns your values into a one-page AI policy. The sprint then finalises it with your team.
- 1 · 10 minWhere are we using it?
A no-blame round on the tools people already use.
- 2 · 15 minWhat do we stand for?
What your values mean for AI, in plain words.
- 3 · 20 minTraffic-light sort
Real and proposed uses sorted into go, with rules, and not here.
- 4 · 10 minData classes
Which tools may touch which data.
- 5 · 15 minDraft the page
Purpose, rules, sign-off, disclosure and owner on one page.
- 6 · 5 minOwner and review date
Who owns it, who approves it, and when it's reviewed.
An example sort
Adapted from the Data Ethics Decision Aid (DEDA) by Utrecht Data School, a dialogue format widely used in the Netherlands.
Go
- Drafting internal docs and meeting notes
- Summarising public research
- Searching your own policies
- Rewriting text in plain language
With rules
- Funder reports, with a named reviewer
- Donor emails, disclosed where needed
- Translations, with a native check
- A public FAQ bot, labelled, with a human route
Not here
- Deciding who gets support
- Case notes in consumer tools
- Generated images of the people you serve
- Crisis conversations without a human
Questions we hear
About AI and ethics
Will you ever say no?
Yes. We won't build tools that score or rank people for support, hide that they are AI, or exist mainly to cut jobs. Sometimes the best fix isn't AI at all, and we'll say so.
What if we're not sure AI fits who we are?
That's a reasonable place to start. We can begin with your red lines and the smallest, lowest-risk use, and "not now" is always an acceptable answer.
Which AI tools do you use?
Whatever fits your data, budget and values, including European and open options. We take no commissions, and everything we build can move with you.
Our data is sensitive. Can we still work with you?
Yes. We start with workflows that don't touch sensitive data, and put rules and contracts in place before anything that does.
Our donors might not like it.
Hidden use is the bigger risk. We help you write a short public AI statement and keep AI away from decisions about people, where public support is lowest.
Does this cover the EU AI Act?
Our training helps you meet the AI literacy duty for staff, and our rules on telling people when they talk to AI match the Act's transparency duties. We're not lawyers, and we'll tell you when you need one.