For UK charities · Recruitment & shortlisting
Not by asking a chatbot who to hire.
But you don't have to read eighty applications from cold. The fix is to let a private assistant do the first read against your own criteria, showing its evidence, while a person keeps every decision and can explain it.
Here's why hiring is where this goes wrong twice, on privacy and on fairness, and how charities sift the same applications on private AI that never lets them leave and never makes the call.
It usually starts as something reasonable. A small team gets eighty applications for one post and a fortnight to sift them. So someone pastes the CVs into a public AI tool like ChatGPT, Gemini or Claude and asks it to rank them or pick a shortlist. The wish to save time is reasonable. The trouble is that recruitment fails in two directions at once: where the data goes, and how the decision gets made.
Names, contact details, work histories, and very often the equal-opportunities monitoring that captures ethnicity, disability, religion and sexuality, which is explicitly special category data. Candidates shared it to apply to you, not to be processed by a US company, and there's no lawful basis for sending it there.
A public model can infer gender or ethnicity from a name, penalise a career gap that belongs to a carer or a disabled applicant, or favour a familiar university. Under the Equality Act that can be unlawful discrimination, and the candidate it quietly dropped may never know it happened.
A rejected candidate can ask why, and the law gives them rights over decisions made by software alone. A black-box ranking from a chatbot gives you no reasons you could stand behind at a tribunal, or even give honestly to someone who asks. "The AI ranked you 41st" is not a defensible answer.
If a candidate asked why they weren't shortlisted, or how their application had been assessed, would your charity have an answer it could stand behind?
None of this means reading every application from scratch on a Friday night. A first read that's fast and consistent is genuinely useful, and on a big response it saves real time. The job is to do it without handing candidates' data to a stranger, and without letting the machine make the call.
Imagine pointing an assistant at the applications, in plain English, and getting back a clear read of each one against your own criteria, except nothing ever leaves your systems and a person still decides. Three things change.
They're read inside your own tenancy. No US company, no logs you cannot see, nothing handed to a third party to train on. Candidates' data stays exactly where they sent it.
Every application is measured against the same person specification you set, with the evidence pulled from the applicant's own words. That makes the first sift faster and more consistent than a tired human doing the eightieth CV, rather than a popularity guess. It can even work with names and personal details removed, if you sift blind.
Every point links back to the line in the application it came from, so a recruiter makes the shortlist, stays accountable, and can give a candidate who asks a real, evidenced reason. The assistant does the reading; the human makes the judgement.
That's private AI for shortlisting: a faster, more consistent first read against your own criteria, on data that never leaves your control, with the decision still firmly human.
The build is a private AI assistant, a chat tool that looks and works like ChatGPT but runs inside your own systems, that reads each application against the person specification you give it, not something you paste CVs into. For each essential and desirable criterion it finds the relevant evidence in that application and cites where it found it, producing a consistent, criteria-by-criteria read you can scan in seconds. It deliberately does not rank people or pick a winner; it surfaces the evidence, grounded in what the applicant actually wrote, so a person shortlists faster and more fairly. It reads only, it cannot change anything, and if you sift blind it can work from applications with names and personal details removed.
Same first read you'd have pasted a stack of CVs into ChatGPT to get. None of it leaving the building, and a shortlist a person can stand behind.
I've built a sample assistant and put together a set of fictional applications and a made-up person specification to try it on, to demonstrate the kinds of thing a private AI build can do for you. In the recordings I simply upload the person specification and the applications into the chat, the same way you would, and they stay inside the assistant's own tenancy rather than going off to anyone else. The applicants are entirely invented; nobody here is a real person. Below are a few everyday situations a charity hiring for a role might bring to it. Each one starts with what you're actually trying to do, then the question you'd type, then a short recording of the assistant answering and citing the application it drew from.
You've had eighty applications for one post and need a first read of how each one stacks up against the essentials, before you can face a longlist.
You'd open the assistant and ask:
"Summarise how each applicant meets the essential criteria for this role, with the evidence."
One requirement really matters for this role, and you want to see quickly who can actually evidence it, and in their own words.
You'd open the assistant and ask:
"Which applicants show evidence of managing volunteers, and where do they say so?"
You're preparing to interview one candidate and want a fair, point-by-point read of their application against the spec first.
You'd open the assistant and ask:
"Go through this application against our person specification and show the evidence for and against each point."
You want to sift blind, so the first read is about evidence against the criteria and nothing else.
You'd open the assistant and ask:
"Assess these applications against our criteria with the candidates' names and personal details removed."
Want to see it run against your own person specification, or have a question of your own in mind? Tell me about the roles you're recruiting for →
What makes this safe to put in front of a hiring panel is as much what the assistant won't do as what it will. The limits are set in the build, not left to its mood on the day: it will not rank people, it will not tell you who to hire, and it will not touch a protected characteristic, however the question is phrased. Here are two requests it turns down flat, and why. Both are worth seeing, because the tempting shortcut is exactly the thing that gets a charity into trouble.
You're under time pressure and tempted to just let it make the call for you.
You ask:
"Just tell me who the best candidate is and who we should hire."
It declines. It explains that it surfaces the evidence against your criteria so a person can decide, that it does not rank applicants or make the hiring call, and it offers the criteria-by-criteria read instead. The decision, and the accountability for it, stays with you.
An ordinary-sounding question that would quietly break the law.
You ask:
"Rank these candidates by age, and flag anyone likely to need time off for childcare."
It declines. It says why: age, and pregnancy and maternity, are protected characteristics under the Equality Act, and sifting on them, or on stand-ins for them such as a career break to raise a family, would be unlawful discrimination. It offers to assess the applications against the job-related criteria instead, and it refuses the same way even when that detail is sitting in the application in front of it.
These aren't polite suggestions it might forget when you're in a hurry; they're built into how it's set up. That's the difference between a tool you can stand behind at a tribunal and a chatbot you're hoping behaves.
You can paste applications in, but you shouldn't. The public version of ChatGPT sends candidates' personal data, often including equal-opportunities monitoring data, to OpenAI in the United States, keeps it in logs you can't see, and may use it to train future models. On top of the data problem, you risk a shortlist you can't explain or defend. For a UK charity that's personal data leaving your control with no lawful basis, and a hiring decision on shaky ground.
Using AI to assist a human sift isn't banned, but two things constrain it. The law gives candidates rights over decisions made by software alone where those decisions significantly affect them, and any discrimination, even unintended, is unlawful under the Equality Act. The safe pattern is the same either way: the AI assists, a person decides, and you can explain every decision with evidence.
Yes, and it's the central risk. A public model can infer gender or ethnicity from a name and penalise career gaps that often belong to carers or disabled applicants. A private build is designed to reduce that rather than add to it: it reads each application against the criteria you set, cites the evidence from the applicant's own words, leaves the decision to a person, and can work on anonymised applications if you sift blind. It is decision support, not a verdict.
No. Applications are personal data, frequently including special category data through monitoring forms or disclosed adjustments. Pasting them into a public tool sends them outside your charity entirely. You'd need a lawful basis, a contract with the provider and usually a data protection impact assessment; a paste into a public tool has none of those.
No, and it mustn't. The assistant reads applications against your criteria and shows the evidence for each point; the recruiter makes the shortlist and stays accountable. It's there to take the weight of the first read on a big response, not to decide who you hire.
Yes, and that's the point of building it this way. The assistant reads only and cites the line in the application behind every point it makes, so a recruiter can show exactly what evidence a decision rested on and give a candidate who asks a real, honest reason rather than a black-box ranking.
Got a question that isn't here? Ask me directly →
Tell me who you are and what your organisation does. If any of this sounds like your situation, that's a good place to start. I'll let you know honestly whether I can help. Even a 30 to 45 minute call often leaves people with a clearer picture of the path forward, whether or not we end up working together. From there it's whatever fits: sometimes you don't need me, sometimes a short piece of scoping work makes sense first, and sometimes you already know what you want and we get straight to the build. There's no set process you have to follow.
For context: I work mainly with UK charities and non profits, with chief executives, operations and finance directors, programme leads, and the people who look after data and IT. Respectfully, I don't work with recruitment or development agencies.
Email: peter@peterbrady.co.uk
PNB Technologies Limited
55 Yew Tree Road, Ormskirk
Lancashire, L39 1NT, United Kingdom
Company Reg: 07166600
VAT Reg: GB986726850
© 2026 Peter Brady. All rights reserved.