Lesson 1: What AI Can and Cannot Do in Agency Recruitment

Why start here

AI assistants such as ChatGPT, Gemini, Claude and Microsoft 365 Copilot can write a job advert in seconds. They cannot tell you whether the client's hiring manager is really flexible on salary, or whether your candidate is quietly considering a counter-offer. In this level you will build a task map: a simple sort of your own working week into what AI can draft, what it can help with, and what must stay fully human. Every later level builds on this map.

What AI assistants do well

These tools produce fluent text from patterns in their training data and from whatever you give them. In agency work they are reliably useful for:

• First drafts: job adverts, candidate update emails, interview question lists, LinkedIn posts.
• Restructuring: turning rough call notes into a tidy summary, or bullet points into a client email.
• Starting points: a first Boolean search string, or an outline for a market update.
• Learning: explaining an unfamiliar job title, technology or sector before a client call.

A useful comparison is a quick, well-read new trainee who has never met your clients or candidates. The work comes fast and reads well, but you check all of it.

What AI assistants cannot do reliably

Learn these four limits now. Level 4 covers how to check output in detail.

1. They make things up. A model can state wrong facts, salaries, laws or statistics, or invent sources, and sound completely confident while doing it. OpenAI's own help guidance warns that ChatGPT can produce incorrect or misleading answers and tells users to check important information.
2. Their knowledge can be out of date. A model's built-in knowledge stops at a training cut-off. Salary bands, legislation and market conditions may have changed since then.
3. They lack your real context. The tool does not know the client's culture, the hiring manager's unstated preferences or the candidate's circumstances unless someone tells it, and some of that information must never be pasted in (Level 3).
4. They can repeat bias. In November 2024 the UK Information Commissioner's Office reported that some AI recruitment tools it audited let recruiters filter candidates by protected characteristics, or guessed gender and ethnicity from candidates' names.

What the evidence says, with dates

When you quote a statistic, say who produced it, when, and who was surveyed. Then say what it does not show.

• LinkedIn research, released 7 January 2026: 93% of talent acquisition professionals surveyed planned to increase their AI use in 2026, and 59% of recruiters said they already use AI to find "hidden gem" candidates.
  • Shows: AI use is mainstream and growing.
  • Does not show: that the AI output was accurate or fair. These are self-reported plans from a mainly in-house, international sample, not UK agencies specifically.
• Bullhorn GRID 2026 Industry Trends report (nearly 300 UK & Ireland recruitment professionals): agencies using AI at any stage of recruitment were four to eight times more likely to have grown revenue than lost it, yet only 7% said AI was embedded throughout their workflow.
  • Shows: growing agencies tend to use AI, and few agencies have fully integrated it.
  • Does not show: that AI caused the growth. Growing agencies may simply have more budget for tools. The survey was also run by a software vendor.
• ICO audit outcomes report, 6 November 2024: audits of AI recruitment tool providers, carried out between August 2023 and May 2024, led to nearly 300 recommendations on fairness, collecting less data and transparency.
  • Shows: real risks exist in some tools.
  • Does not show: that every AI tool has these faults.

Task: build your AI task map

Do this with last week's diary open.

1. List every task you did last week, such as briefs, adverts, searches, screening, calls, CV formatting and business development.
2. Ask three questions about each task:
   • Is the output mainly text that I can read and check before anyone else sees it?
   • Does it depend on context only I have, such as relationships, tone or confidential details?
   • Does it decide or significantly affect a person's chances?
3. Sort each task into one of three columns:
   • AI can draft (I edit and approve): advert first drafts, email templates, interview question lists, Boolean starter strings.
   • AI can assist (I lead): summarising my own notes, preparing for a client meeting, learning about a new sector.
   • Stays fully human: shortlisting and rejection decisions, getting candidate consent before a CV goes out, offer negotiation, sensitive conversations such as health, adjustments or redundancy, and relationship calls with clients.
4. Note the check each AI task needs, for example "verify salary figures" or "confirm against the client brief".
5. Review the map after a week and move any task where the AI saved no time or needed heavy correction.

Example

Priya, a technical resourcer, puts "write advert for a DevOps role" under draft, "research Kubernetes before a client call" under assist, and "decide which 8 of 240 applicants go forward" under human. The AI can help her read faster. The shortlist decision stays hers.

Key takeaways

• Sort tasks into draft, assist and fully human. Decisions about people stay human.
• The four main limits are made-up facts, out-of-date knowledge, missing real context and inherited bias.
• Every AI output gets a human check before it reaches a client or candidate.
• Quote statistics with source, date and sample, and say what each one does not prove.
• A survey showing a correlation, such as "AI users grew faster", does not prove that AI caused the result.