01 · The case
Why AI in recruiting, and why now.
Roughly 60% of a recruiter's week goes to work that is repetitive and automatable — sourcing admin, screening, scheduling, note-taking, reporting. That figure is not an argument for fewer recruiters. It is an argument about where a scarce, expensive resource is currently being spent.
The shift underway is not "AI replaces recruiters". It is narrower and more uncomfortable: recruiters who orchestrate AI outperform recruiters who don't — on speed, on volume handled, and on the quality of the hiring-manager conversations they have time for. The gap between those two groups is widening inside the same team, on the same requisitions, at the same salary.
~60%
of a recruiter's week goes to repetitive, automatable work — sourcing admin, screening, scheduling, notes, reports.
10–20 hrs / week
realistically recoverable with disciplined AI workflows — reinvested in candidates and hiring managers.
High-risk
how the EU AI Act classifies AI used in employment decisions. Capability has to arrive with governance attached.
Zero theory for its own sake
Eight hours, six modules, and a practice block after every one. Nobody leaves a module having only watched it done.
The training is not about tools. Tools change every quarter.
It is about a repeatable prompt structure, a set of saved workflows, and the judgement to know which decisions never leave a human — three things that survive every vendor cycle.
Figures throughout this paper are directional planning estimates drawn from the programme. Every team builds its own baseline in the measurement module — a number you did not measure yourself is a number you cannot defend to a CFO.
02 · Audience & outcomes
Who it is for, and what they can do by 18:00.
Designed for TA Specialists and Senior TA Specialists with two to seven years of experience — recruiters who already know how to hire and now need to know how to orchestrate. It assumes no technical background and no prior AI use.
| By the end of the day | What that means in practice |
|---|---|
| Prompt like a pro | Use a repeatable six-part prompt structure for any recruiting task, instead of chatting hopefully at a text box. |
| Source faster | Generate Boolean and X-ray strings, candidate personas and sourcing strategies — then iterate them against real results. |
| Screen smarter | Analyse CVs, compare candidates against the same agreed criteria, and draft shortlist memos that survive a hiring-manager challenge. |
| Communicate better | Personalised outreach, candidate updates and rejection notes at scale — in the recruiter's own voice, not the model's. |
| Run better interviews | Structured guides, behavioural questions and anchored scorecards in minutes rather than evenings. |
| Report like a leader | Pipeline analytics, dashboards and executive-ready hiring updates generated from the ATS export. |
The session shape
| Time | Block | What happens |
|---|---|---|
| 09:00 · 45m | Module 1 · AI fundamentals | LLMs, prompts, limits — the 20% of theory you actually need |
| 09:45 · 20m | Practice 1 | Break a model on purpose: run the same prompt three times, force a hallucination, catch it |
| 10:05 · 45m | Module 2 · Prompt engineering | The six-part recruiting prompt, plus live rewrites of the room's own prompts |
| 10:50 · 30m | Practice 2 | Rebuild one of your own real prompts through the six parts; swap with a colleague and critique |
| 11:20 · 15m | Break | |
| 11:35 · 80m | Module 3 · AI across the lifecycle | Stage by stage: pain points, AI plays, hours recovered, human checkpoints |
| 12:55 · 20m | Practice 3 | Map your own desk: where do your hours actually go, and which two stages do you fix first |
| 13:15 · 45m | Lunch | |
| 14:00 · 120m | Module 4 · Five hands-on labs | Five core workflows built on live requisitions — JDs, sourcing, screening, interview kits, reporting |
| 16:00 · 45m | Module 5 · Tools & automation | The landscape, vendor evaluation, and no-code automation |
| 16:45 · 30m | Practice 4 | Vendor teardown role-play against the five criteria, then map one automation end to end |
| 17:15 · 45m | Module 6 · Responsible AI + your 30-60-90 | GDPR, EU AI Act, governance, adoption roadmap, KPIs and the seven-day homework |
Nine hours elapsed, 09:00 to 18:00, of which eight are instruction and practice and one is breaks. Three hours forty of that eight is hands-on — the five labs plus a practice block closing every taught module.
Why the day is eight hours and not six. A six-hour version fits, and it produces recruiters who have watched the workflows rather than run them. The two additional hours buy a practice block after every module, so nobody leaves having only seen it done — which is the difference between a team that is enthusiastic on Friday and a team still using the workflows in week six.
03 · Fundamentals
The minimum viable theory.
Enough to use the tools well and to hold your own in a vendor meeting. No more than that.
What actually happens when you send a prompt
You send text
Instructions, context and examples — everything the model can see. Nothing outside the prompt exists to it.
The model predicts tokens
It generates the most probable next words from patterns learned in training. It is not looking anything up.
You get a draft, not a verdict
Fluent, fast, often excellent — but not a database and not a decision-maker.
Three numbers that matter
Tokens
Chunks of text of roughly four characters. Models read and price by token — a CV is about 800–1,500 tokens.
Context window
The model's working memory for one conversation. Everything you paste must fit, and it forgets nothing inside the window and knows nothing outside it.
Temperature
Outputs vary run to run. Same prompt, slightly different answer — by design, and precisely why review matters.
And one habit
Treat the first output as draft one. Refinement — "shorter, more senior tone, add a risks section" — is where the quality actually comes from.
The eight terms you need in a vendor meeting
| Term | What it means, in recruiting terms |
|---|---|
| LLM | A large language model — GPT, Claude, Gemini. The engine behind generative text tools. |
| Generative AI | AI that creates new content — text, images, structured data — rather than only classifying. |
| Embeddings | Text converted to numbers so machines can measure similarity. Powers semantic CV-to-job matching. |
| RAG | Retrieval-Augmented Generation: the model answers using your documents — ATS notes, JDs — which sharply reduces invention. |
| AI agent | AI that plans and executes multi-step tasks with tools (search, email, ATS) rather than returning single answers. |
| MCP | Model Context Protocol — an open standard letting AI assistants connect safely to your systems. |
| Reasoning model | Models that work step by step before answering. Better for comparison, scoring and analysis. |
| Hallucination | Confident, fluent, wrong. The model fills gaps with plausible fiction — the single risk to manage hardest. |
Where these systems fail — and the habits that keep you safe
Failure modes to expect
- Invented facts — employers, degrees, statistics, salary data, "sources"
- Outdated knowledge — training data has a cutoff; market data may be stale
- Instruction drift — long tasks lose constraints set early in the prompt
- Bias reproduction — models mirror biased patterns in training data and in your inputs
- Sycophancy — it agrees with your framing unless you ask for counter-arguments
Your safety habits
- Verify every fact you will act on — names, numbers, claims, links
- Ground the model: paste the real JD, the real CV, the real data. Don't let it guess
- Ask for uncertainty explicitly — "flag anything you're not sure about"
- Never let AI make the hiring decision. It drafts, you decide
- No personal candidate data in unapproved tools
04 · Prompting
The six-part recruiting prompt.
One structure, memorised once, reused for every recruiting task. It is the single highest-leverage thing in the programme, and it needs no tool beyond the assistant the team already has.
| Part | What goes in it |
|---|---|
| 1 · Role | "You are a senior technical recruiter in iGaming…" — sets expertise and tone. |
| 2 · Context | Company, team, role, market, constraints. Paste the real JD, the intake notes, the CV. |
| 3 · Task | One clear verb: draft, compare, extract, rank, summarise, generate. |
| 4 · Format | Table, bullets, email, JSON, scorecard — with length limits. |
| 5 · Constraints | Tone, language, what to exclude, compliance rules, "flag uncertainty". |
| 6 · Examples | One good example beats three paragraphs of instruction. |
The same task, two prompts
Weak
"Write a LinkedIn message to a Java developer."
- Generic, template-smelling output
- No role, market or candidate specifics
- Sounds like every other AI InMail that candidate received this week
Strong
"You are a senior iGaming recruiter. Context: hiring a Backend Java Engineer (Vert.x, high-throughput betting platform) in Sofia, hybrid, budget X. Candidate: 6 yrs at a payments scale-up, posts about performance tuning. Task: 90-word LinkedIn message. Format: hook referencing his content, one concrete challenge of the role, soft CTA. Constraints: no buzzwords, no 'rockstar', British English. Example of my voice: [paste past InMail]."
The second prompt is longer to write once and then reusable forever as a template. That is the whole economic argument for prompt structure: the effort is front-loaded and the output compounds.
05 · Lifecycle
Eight intervention points, four human gates.
AI compresses every stage of the lifecycle. Judgement stays human at four gates: shortlist decisions, interview evaluation, offer decisions, and any message that shapes a candidate's experience.
| Stage | The AI play | Time recovered | Human checkpoint |
|---|---|---|---|
| Intake & job descriptions | Generate intake question sets; transcribe the intake call and extract must-haves, nice-to-haves, red flags and sell points; draft the JD from real intake data in your template; produce the LinkedIn and internal-mobility variants from the same source. | 2–3 hrs per role | Hiring-manager sign-off on requirements before sourcing starts |
| Talent mapping & sourcing | Boolean with full synonym trees, iterated live ("too narrow, relax titles"); X-ray strings for LinkedIn, GitHub, conference sites; candidate personas; target-company lists as a research starting point. | 3–4 hrs per week | Recruiter validates result quality before outreach; company lists verified against live data |
| CV screening & comparison | Extract a structured profile per CV; compare candidates against the same agreed criteria in a side-by-side matrix with evidence quotes; draft shortlist memos; generate phone-screen guides personalised to each CV. | 4–5 hrs per week | Recruiter makes every advance/reject decision — AI output is decision support only |
| Outreach & communication | Messages drafted from the candidate's actual profile in your voice; three-touch sequences each with a different angle; warm, specific rejection notes; status and interview-prep emails from pipeline data. | 2–3 hrs per week | Every message read and personalised before send — you sign it, you own it |
| Interviews | Structured guides per competency with follow-up probes; behavioural and technical question banks mapped to the scorecard; AI notetakers with consent; draft debrief summaries aggregating scorecards. | 1.5–2 hrs per role | Interviewers score independently before seeing AI summaries; consent required for recording |
| Offers, reporting & post-hire | Offer summary and business justification from comp data and interview evidence; weekly pipeline updates per hiring manager from the ATS export; executive dashboards; silver-medalist nurture sequences. | 2–3 hrs per week | Numbers reconciled against the ATS before anything reaches an executive |
The realistic ceiling is around 18 hours per recruiter per week — and it arrives in months, not days, once workflows are genuinely embedded rather than demonstrated. The point of recovering it is not to run leaner. It is to reinvest it in what AI cannot do: hiring-manager trust, candidate relationships, and closing.
06 · Labs
Five workflows, built on your own live roles.
The two-hour lab block is the largest of the day's five practice sessions, and it is not a demo. Participants bring live requisitions and real materials, and leave with working assets. Every lab ends the same way: save your prompt.
The anatomy of a repeatable workflow
Goal → Inputs → Prompt → Tool → Review → Ship & save. One outcome and one owner; real materials rather than invented ones; the six-part structure saved as a template; a tool appropriate to the data sensitivity; a human checking facts, tone and fairness; and then the prompt filed in the team library.
| Lab | Participants bring | They leave with |
|---|---|---|
| 1 · Job description in 15 minutes | A live req with intake notes, their JD template, and one JD they consider excellent | A ship-ready JD, its LinkedIn variant, and a saved jd-draft-v1 template |
| 2 · Sourcing strategy + Boolean, 20 min | A hard-to-fill live role and access to LinkedIn to test strings | A tested Boolean and X-ray set, a persona doc, and a target-company hypothesis list |
| 3 · CV screen & shortlist memo, 20 min | Three sanitised CVs (names, photos, age, address removed) and the agreed screening criteria | A defensible comparison matrix, a shortlist memo, and the screening prompt saved |
| 4 · Interview kit, 20 min | The finalised JD from Lab 1 and their rating scale | A structured guide, anchored scorecard and one-page interviewer cheat sheet |
| 5 · Weekly reporting, 15 min | A sanitised pipeline export and one past weekly update they wrote by hand | Friday's reporting done, a reusable template, and two hours back every week |
A workflow run once is a trick. A workflow saved, named and shared in the team library is an asset.
It improves with every refinement, and it onboards the next recruiter for free.
07 · Tools
Six categories, and how to evaluate any vendor in twenty minutes.
| Category | Representative options | What it is for |
|---|---|---|
| General assistants | Claude, ChatGPT, Gemini, Copilot | Drafting, analysis, comparison — the daily workhorse |
| Research & sourcing | Perplexity, Deep Research, SeekOut, hireEZ, Juicebox | Market intelligence and people search |
| Interview intelligence | Metaview, BrightHire, Otter, Fireflies | Transcription, summaries, coaching — consent required |
| ATS/CRM native AI | Ashby, Greenhouse, Lever, Recruit CRM, LinkedIn Recruiter | Matching and drafting inside the system of record |
| Automation platforms | Zapier, Make, n8n, Lindy | Connecting ATS, email, calendar and Slack into workflows |
| Content & decks | Gamma, Canva AI, Notion AI, NotebookLM | Employer branding, documents, knowledge bases |
Master one general assistant deeply before adding point solutions. Around 70% of the value in this programme needs nothing else — and a team that buys three tools before mastering one usually ends up using none of them well.
The vendor conversation
- Data & complianceWhere is data processed and stored? Is our data used for model training? Is a DPA available? Is the GDPR and EU AI Act posture documented?
- Real capabilityDemo on OUR data, not their demo data. What breaks at our volume? Which claimed features are generally available versus roadmap?
- IntegrationNative ATS/CRM connection or manual export? API or MCP access? Who maintains it when the ATS updates?
- Bias & auditabilityIs there bias-testing evidence? Can we audit why it scored or ranked a candidate? Is human override built in?
- ROI & exitPrice per recruiter against hours saved. Pilot terms. What happens to our data if we leave?
08 · Automation
From single prompts to hands-free workflows.
The pattern is always the same four steps, and the third one is not optional:
Trigger → AI step → Human gate → Actions. A new application lands in the ATS; the model extracts the profile, scores it against criteria and drafts a screen summary; the recruiter reviews it in Slack and approves or adjusts with one click; the ATS updates, the hiring manager is notified and the candidate email is queued.
Three starter automations — build in this order, because it is also the value order
Interview scheduling
Calendar link, reminder and prep-email sequence. Pure admin, zero judgement — automate it fully and never think about it again.
Weekly reporting
ATS export → AI summary → a draft waiting in your inbox every Friday at 08:00, for review before sending.
Application triage
Extraction and criteria pre-screen, then a recruiter approval gate, then ATS and notifications. Never auto-reject.
09 · Responsible AI
Capability without governance is a liability.
GDPR — data protection
- Candidate data is personal data: lawful basis, minimisation and retention limits apply to AI use too
- No candidate personal data in unapproved consumer AI tools — anonymise before prompting
- Art. 22: candidates have rights around solely automated decisions. A human must meaningfully review rejections
- Transparency: be able to explain how AI was used in the process if asked
EU AI Act — employment AI is high-risk
- AI for recruitment, screening, evaluation and selection sits in the high-risk category
- Obligations flowing to deployers: human oversight, monitoring, transparency to affected people, staff AI literacy
- Manipulative and social-scoring uses are banned outright; emotion recognition in the workplace is prohibited
- Vendor due diligence is your obligation too — ask for their conformity documentation
Six golden rules for the whole team
| Rule | In practice |
|---|---|
| Human decides, always | AI drafts, structures and suggests. Advance, reject and offer are human decisions only. |
| Anonymise before you paste | Strip names, photos, age and contact details from CVs before any prompt. |
| Approved tools only | Sensitive data stays in tools with a DPA and a no-training guarantee. |
| Audit for bias | Spot-check AI shortlists against human ones; challenge prestige-signal drift. |
| Verify, then send | Every fact checked and every message read before it leaves under your name. |
| Log and share | Save prompts to the library; note where AI assisted in the process. |
This is training guidance, not legal advice. Involve Legal or your DPO when defining the organisation's AI policy — and note that the AI Act's high-risk obligations have their own timetable, which is worth confirming as it stands when you plan.
10 · Adoption
The 30-60-90 roadmap.
Training that ends at 18:00 with no roadmap produces enthusiasm and no change. The last block of the day is the plan each participant leaves with.
Personal wins
Master the six-part prompt on daily tasks. Run Labs 1, 2 and 5 on live roles. Start a personal prompt library. Baseline your time — where do the hours actually go now?
Team habits
Shared prompt library with a naming convention. A team tool policy: approved tools and the anonymisation rule. First automation live — scheduling or reporting. A weekly fifteen-minute "prompt of the week" ritual.
Scale & prove
Full lifecycle coverage on two or three pilot roles. Measure against the baseline: hours, time-to-fill, hiring-manager satisfaction. Governance checklist signed off with HR and Legal. Present results and the expansion case to leadership.
11 · Measurement
Measure it, or it didn't happen.
| Dimension | What to track |
|---|---|
| Efficiency | Hours saved per recruiter per week against baseline · time-to-shortlist and time-to-fill · admin share of the week |
| Quality | Submit-to-interview and interview-to-offer conversion · quality of hire at 90 days · shortlist diversity against the pre-AI baseline |
| Experience | Candidate NPS and response sentiment · outreach reply rate · hiring-manager satisfaction with updates and kits |
| Adoption & safety | Percentage of recruiters active weekly with AI workflows · prompt library size and reuse rate · governance checklist compliance, incidents = 0 |
Two of these matter more than the rest for the business case. Hours saved is what finance will ask for, and it is only credible if the baseline was measured before the training. Shortlist diversity against the pre-AI baseline is what protects you: it is the number that shows, with evidence, that the efficiency gain did not quietly cost you fairness.
What participants take home
- AI is a drafting and analysis engineJudgement, decisions and relationships stay yours.
- The six-part prompt plus your real materials beats any tool you can buyStructure outperforms subscriptions.
- Workflows saved to a shared library compoundOne-off prompts evaporate.
- Human review gates and anonymisation keep you fast AND compliantThey are not in tension.
- Start with reporting and JDs this weekMeasure against your baseline from day one.
The seven-day homework is deliberately small: run one lab workflow on a live role end to end, save three prompts to the team library, log an hours-saved estimate for the week, and bring one failure or surprise to the follow-up session. The failures are the most valuable material in the follow-up.
12 · Delivery
How Dreamwork delivers this.
This programme is part of Dreamwork's AI-Native TA Advisory practice. It is delivered by a practitioner rather than a trainer — talent acquisition led for a 10,000-employee multinational across four markets, a €30M staffing P&L, a 1,400-person greenfield build inside twelve months, and a sustained 38-day time-to-hire at multi-country scale.
That matters for a specific reason. A trainer teaches the tool; a practitioner knows which of these workflows survives contact with a real hiring manager on a contested requisition, and which ones quietly get abandoned in week three.
| Format | Shape | Best for |
|---|---|---|
| The core session | 8 hours, in person or remote, up to 12 participants | A TA team that needs capability fast, with working assets built by their own hands by 18:00 |
| Session + 30-day follow-up | Core day plus a 90-minute review at day 30 | Teams that want the adoption curve supported past the enthusiasm phase — where most training dies |
| Full advisory engagement | Process audit, AI workflow design, tool selection, governance | Functions rebuilding how they hire, not just how they prompt |
Every version includes the governance module. Capability without it is a liability we are not willing to hand a client — and in a high-risk regulatory category, it is the part that a board will eventually ask about.
"AI won't replace recruiters. Recruiters who use AI will replace recruiters who don't."
The training exists to put your team on the right side of that sentence — with the governance to make it defensible.