AI copilot for recruitment agencies: what it can (and can’t) do

Last reviewed: 5 October 2026. This guide describes Marvin as it works today. It is general information, not legal advice.

TL;DR

• An AI copilot is useful when it works on your own data: your candidates, clients, jobs, notes and pipeline. Without that, it is a chatbot.
• The jobs it does well in an agency: search in plain English, match and score candidates with an explanation, summarise a record, draft a message or an interview brief.
• The questions to ask any vendor: does it send on its own, which actions does it confirm first, and does it see more than the recruiter who uses it?
• In Marvin, the Copilot lives in Desk. It drafts outreach but does not send it: the recruiter does. In Desk it shows a confirmation card before sensitive actions, and it acts with the signed-in recruiter’s permissions.
• A recruiter, not the tool, makes and records every decision about a candidate.

What “AI copilot” means in a recruitment CRM

Many AI features in recruitment software write text from a prompt and stop there. A copilot goes further: it reads and acts in your CRM. You ask “who in our pool has led a finance transformation in the North West?” and it runs the search, opens the records and answers with names you can click.

For a UK agency, three things decide whether that is useful or risky:

1. What it can read. The answer is only as good as your data: up-to-date CVs, notes after each call, pipeline stages that reflect reality.
2. What it can do. Reading is low risk. Writing to records, moving candidates in a shared pipeline or sending messages are not.
3. Whose rights it uses. An assistant that sees every record in the firm, whatever the user’s role, breaks your own access rules.

What Marvin’s Copilot does

Marvin’s Copilot is built into Desk, the ATS and CRM where your consultants work. You write a request; it chooses the functions to call, shows you what it is doing and answers with links to the records.

Search and read

The Copilot searches your candidates, contacts, companies and jobs, including with Boolean queries and your saved searches. It reads a candidate’s details, timeline and notes, a company record, a job, a meeting recap or the attachments of an email. It can also list stale records and give pipeline figures.

With your connected LinkedIn account, it can also run a LinkedIn search or look up a profile. Importing a profile into your database is a separate step, and it asks you first (see below).

Match, score and compare

• Matching: candidates for a job, jobs for a candidate, or people similar to a profile you like.
• Scoring with an explanation: candidates for a job, with a score that separates fit from history signals and says why each one is ranked where they are.
• Comparison: two to five candidates side by side for a role, with a score, strengths, gaps and a recommendation for each.

These are suggestions for a recruiter to check, not a verdict. A score is built from what your database says about a person; a thin profile gets a weak score for reasons that have nothing to do with the candidate.

Prepare and draft

• A summary of a candidate’s recent activity, and the next action it suggests.
• A brief before an interview or a client meeting, built from notes, timeline and CV.
• Interview questions for a candidate and a role.
• Message drafts: an email, a LinkedIn message, invitation or InMail, or a WhatsApp message, written from the record and the role.
• Duplicate detection and tag suggestions to keep the database clean.

Each draft arrives on a card. The recruiter reads it, edits it and sends it in one click. The Copilot does not send these messages itself.

What it asks before doing, in Desk

Actions fall into three groups. The examples below are not a complete list.

Kind of action	Examples	What happens in Desk
Read and analyse	Search your database, read a record, score, compare, summarise, search LinkedIn or look up a profile	Runs straight away
Small writes you asked for	Add to a list, create a list, create a reminder, add feedback	Runs straight away
Sensitive actions	Enrol in a sequence, move a candidate to another stage, accept or refuse an application, send an assessment, post a job to job boards, import or enrich a LinkedIn profile, enrich contact details, create or update a record	Confirmation card: nothing happens until the recruiter clicks

Importing or enriching profiles, changing the pipeline the whole team sees, starting outreach and actions that can email a candidate, such as refusing an application with its refusal email, go through that card. When the Copilot reaches an action it does not recognise, it falls back to asking. Some actions are never offered to it at all: permanent deletions, data protection operations such as anonymisation or mass exports, payments, secrets and API keys, and file uploads.

Whose permissions it uses

The Copilot acts as the recruiter who is signed in. It calls Marvin’s functions with that recruiter’s session, so their role, the records they are allowed to see and the business rules of each action all apply. A record hidden from that recruiter is refused to the Copilot. If their rights cannot be determined, the Copilot refuses rather than showing the whole firm.

In practice, the Copilot cannot do anything in Desk that the recruiter could not do by hand.

Which AI runs it, and where the data goes

By default, the Copilot runs on the AI providers Marvin contracts with. Your data is hosted in France (Scaleway, Paris). Our sub-processors, including AI providers, are listed in our DPA.

An administrator of your firm can also plug in its own AI provider key (BYOK), from OpenAI, Anthropic, Mistral or Google Gemini. When it is enabled, that key runs all of Marvin’s AI features for your firm, the Copilot included, and the contract with that provider is yours: assess it like any other sub-processor.

If you prefer to work from Claude, ChatGPT or another assistant you already pay for, Marvin also has an MCP server: see our guide to MCP for recruitment agencies.

What it does not do

• It does not send the emails, LinkedIn or WhatsApp messages it drafts: the recruiter does.
• It does not decide to reject a candidate or make a placement: it refuses an application only when a recruiter asks and confirms.
• It does not delete records permanently or run data protection operations.
• It does not replace your notes and data quality: it works from what is in Marvin.

Using a copilot under UK GDPR and the Equality Act

A copilot changes how you process candidate data, so treat it like any new processing.

• Keep a human decision. Under the UK GDPR, as amended by the Data (Use and Access) Act 2025, a significant decision based solely on automated processing needs safeguards: information, a way to make representations, human intervention and a way to contest it. A recruiter who reviews the score and records the decision avoids that situation. Our guide to AI, MCP and UK GDPR covers these rules and the checks to run on the providers involved.
• Watch for bias. A score learns nothing new about a candidate; it reflects what is in the record and the words in the brief. Avoid briefs and criteria that point to a protected characteristic. Our guide to the Equality Act 2010 in recruitment explains what an agency is liable for.
• Update your privacy notice to say that AI tools help your consultants search, summarise and draft, and with which providers.

How to try it on your own desk

The quickest test is a real role. Load a job, ask the Copilot for the ten best matches in your pool with the reasons, compare the top three, then ask for a draft approach to the first one. Check the reasons against what you know of those candidates.

If you are setting up a new agency, our guide to starting a recruitment agency in the UK covers the rest of the groundwork. To see the Copilot on your own data, book a demo.

Frequently asked questions

What is an AI copilot in a recruitment CRM?

An assistant built into the CRM that you ask in plain English. It can search your candidates, clients and jobs, compare and score candidates against a role, summarise a record and draft messages, using the data the CRM already holds. In Marvin, it lives in Desk and calls the same functions a recruiter uses.

Does Marvin’s Copilot send messages on its own?

No. It prepares a draft email or LinkedIn or WhatsApp message on a card. The recruiter reads it, edits it if needed and sends it in one click. Actions that can email a candidate, such as refusing an application with its refusal email, only run after the recruiter confirms them on a card.

Which actions need a confirmation?

In Desk, the Copilot shows a confirmation card before sensitive actions, for example enrolling someone in a sequence, moving a candidate to another pipeline stage, accepting or refusing an application, posting a job to job boards, importing or enriching a LinkedIn profile or contact details, and creating or updating a record. Reading, LinkedIn searches and small actions you ask for explicitly, such as adding a candidate to a list or creating a reminder, run straight away.

Can the Copilot see records a recruiter cannot see?

No. It acts as the signed-in recruiter, with their role and visibility rules. A record hidden from that recruiter is refused to the Copilot too.

Can we use our own AI provider key?

Yes. By default the Copilot runs on the AI providers Marvin contracts with, which are listed in our DPA. An administrator can instead plug in your firm’s own key from OpenAI, Anthropic, Mistral or Google Gemini. That key then runs all of Marvin’s AI features for your firm, the Copilot included.

Can an AI copilot reject candidates for us?

It should not decide. Marvin’s Copilot scores and explains; it refuses an application only when a recruiter asks it to and confirms the card. Under the UK GDPR, a significant decision based solely on automated processing, such as rejecting an application, needs specific safeguards. Keeping a recruiter as the decision-maker is the simplest approach.