A single opening for a delivery rider, sales assistant or warehouse picker can pull in applications from three places at once: a job board, a messenger (WhatsApp or Telegram, depending on the market) and the careers form on your website. Each channel lives in its own window and arrives in its own format, so the recruiter spends the day moving messages around instead of choosing people. Some applicants never hear back, and the strongest ones accept another offer while they wait.
This article is a working blueprint for high-volume, frontline hiring: how to merge applications into one queue, which fields a candidate record needs, how to write knockout questions, what to hand to AI and where the decision must stay with a person. For a broader view of AI across HR, from screening to onboarding, see AI agents for HR. Here we cover only the first step: from application to interview invitation.
What each channel actually sends you
Before automating anything, look honestly at what arrives from each source and in what shape. That determines how much work the AI has to do.
| Channel | What arrives | Typical problem |
|---|---|---|
| Job board (LinkedIn, Indeed, Bayt, hh.uz in Tashkent) | A structured CV: experience, education, sometimes salary expectations and a cover note | Lots of one-click applications to everything; CVs that are generic or out of date |
| WhatsApp or Telegram | Free text, voice notes, a photo of a document, sometimes a PDF; often in mixed languages | No structure; messages get buried; unclear which role the person means |
| Website form | Name, phone, role, maybe a file | Too little to screen on; without a quick follow-up the candidate goes cold |
The takeaway: job boards give you plenty of data and plenty of noise, messengers give you a live, motivated person with no structure, and the website gives you a contact with almost no information. So the AI's job differs by channel: filter in the first case, parse and clarify in the second, ask the missing questions in the third.
One queue, one candidate record
The main win at this stage is not "smart selection". It is that every application ends up in one place in one format: a CRM, a spreadsheet or your applicant tracking system. Each candidate gets a record with the same fields:
- role and location (branch, area, shift);
- source channel, so that later you can see which channel brings people who actually start work;
- contact details and preferred language;
- answers to the knockout questions (see below);
- a two- or three-line summary of experience, written by the AI from the original text;
- status: new, clarifying, suitable, not suitable, invited, not responding;
- a link to the original message or CV, so a recruiter can check where the AI got its conclusion.
On the technical side, most job boards offer either an employer API or email notifications for new applications; access terms depend on your plan, so confirm them with the board's account manager first. Messenger applications come in through a bot or a business account rather than a recruiter's personal phone. Note that the WhatsApp Business Platform only allows free-form replies within 24 hours of the candidate's last message; after that you need pre-approved message templates, which matters for follow-ups. The website form posts straight into the same queue.
Knockout questions: how to write them
Knockout questions are three to five conditions without which someone simply cannot do the job. Not "nice-to-have skills", but hard limits. Good ones are verifiable and never ask the AI to judge personality.
Example for a delivery rider:
- Can you work in the areas we cover (list them)?
- Do you have your own vehicle and a valid licence, or do you need one of ours?
- Which shifts suit you: mornings, evenings, weekends?
- Do you have the right to work here, and when could you start a trial shift?
What must never appear: age, gender, marital status, religion, nationality or anything unrelated to doing the work. In the UK, the Equality Act 2010 protects these characteristics at the hiring stage, and Gulf and Central Asian labour laws have their own anti-discrimination rules. Once a screening rule is written into an AI configuration, it exists in writing and can be produced as evidence. Let the hiring manager and HR define the criteria together, not whoever happens to set up the system.
What the AI does and what stays with people
| Step | What the AI does | What stays with a person |
|---|---|---|
| Receiving | Confirms receipt within minutes, in the candidate's language | — |
| Parsing | Extracts experience, area and availability from a CV, text or voice note; fills the record | Spot-checks that parsing is correct |
| Clarifying | Asks any knockout questions the CV does not answer | Decides the questions and their wording |
| Sorting | Assigns a status with a short reason | Reviews "suitable" and borderline cases, makes the call |
| Inviting | Offers open interview slots, sends the address and what to bring | Runs the interview |
| Rejecting | Drafts a polite message from a template | Approves sending, especially at the start |
The AI never issues a final "not suitable" on its own. It sorts and explains; a person confirms. This is also a legal point: under UK GDPR Article 22, people have rights around decisions with significant effects made solely by automated means, so keeping a human in the loop is the safer design anywhere. After a few weeks, once it is clear the obvious cases are handled correctly, you can limit manual confirmation to borderline ones.
What the bot conversation looks like
A candidate messages the hiring bot at 10:40 pm, long after the recruiter has gone home:
Candidate: Hi, are you still hiring riders?
Bot: Hi! Yes, the rider role is open. Would you prefer to continue in English or Arabic?
Candidate: English is fine
Bot: Great. I'll ask four short questions to see whether the role fits you. Which areas can you work in?
Candidate: Al Barsha, JLT is ok too
Bot: Do you have your own motorbike and a valid licence?
Candidate: yes both
Bot: Which shifts suit you: mornings, evenings, weekends?
Candidate: evenings and weekends
Bot: Thank you. Your answers have gone to our recruiter, who will contact you tomorrow during working hours. You're welcome to send a CV as a file or photo, but it's optional.
What matters here. The bot never promises a job or says "you're a match"; that is a human decision. It offers a language choice instead of guessing. It says when the next step will happen. And it does not ask for passport or visa copies: those belong to onboarding, not screening.
Rejections, silence and personal data
Candidates complain less about rejection than about silence. Replying to everyone is a simple way to protect your employer reputation, and in frontline hiring word travels fast through community chats. But sending messages outward is exactly where automation tends to fail. We learned this ourselves when our outreach system bypassed its own approval rule; the incident review is public. The lesson applies to hiring: mandatory approval before sending has to be enforced in code, not left as a team agreement.
A few rules for rejections:
- do not explain the reason in detail if it concerns personal qualities; "we have chosen other candidates for now" is enough;
- offer to keep the person on file for future roles, but only with their explicit consent;
- do not send rejections at night or seconds after an application arrives: an instant "no" reads as indifference.
CVs are personal data. In the UK that means UK GDPR; the UAE and Saudi Arabia have their own data protection laws, as does Uzbekistan. Before launch, agree with your lawyer where the candidate database is stored, which consent text the bot shows, and how long you keep data on people you did not hire.
A step-by-step pilot plan
- Pick one role with high application volume and clear requirements. Avoid starting with a role where everything depends on the interview.
- Measure your baseline: applications per week, hours until a candidate gets a first reply, how many reach an interview.
- Write the knockout questions and reply templates in every language you hire in, and have a native speaker proofread each one.
- Connect channels one at a time: usually the messenger first, where the chaos is worst, then the website, then the job board.
- For the first two weeks, HR confirms every decision and flags where the AI got it wrong. Those examples become material for refining its instructions.
- Compare against the baseline and decide whether to roll out to other roles.
How agents that take on this kind of work are built is described on our AI agents page. For a related front-desk process, see how an AI chatbot replaces first-line support, and more in the Operations section.
FAQ
Can the AI handle CVs and messages in mixed languages?
Current language models handle English, Arabic, Russian and many other languages, including messages that switch between them, but quality varies by language and drops with slang and abbreviations. That is why the record keeps the original message and a person spot-checks parsing at the start. Voice notes are transcribed first, and errors can creep in at that step too.
Could the AI screen out a good candidate?
Yes, if the criteria are too narrow or a CV is written unconventionally. So the AI does not reject anyone itself; it labels applications as suitable, not suitable or unclear. At the start, HR reviews every borderline case and every rejection.
Should we drop job boards if we have a messenger bot?
No. Different channels can bring different people, and after a couple of months the source field will show which ones produce hires who actually start. The point of the setup is one queue for everything, not picking a single channel.
How much does this cost?
It depends on the number of channels and roles, and where the data needs to go: a CRM, a spreadsheet or an applicant tracking system. The first conversation is for looking at your application flow and working out the scope. You can reach us via the contact page.