Agentic AI in Recruitment: What Simply Can Do

What is agentic AI, and why should you care?

Most AI tools in recruitment work like a search engine with a personality. You ask a question, you get an answer. Done. But what if your AI didn’t just answer questions but actually did things inside your systems?

That’s what agentic AI is about. Not a chatbot that generates text. An AI agent that lives in the ATS itself, knows the fields you defined there, can search candidates, generate CVs and create tasks. All from a chat window in the system that holds the records.

Simply Ask is exactly that. You type: “I just uploaded a vacancy, who’s the best match?” And it searches your candidates, scores them against the requirements the vacancy actually stated, and hands back a ranked list with the reasoning attached, including the parts that don’t match.

Sounds like science fiction? The reason it works is duller than the promise. Simply Ask isn’t reading a copy of your data through a connector. It reads the workspace it lives in, under the permissions of the person asking.

Simply Ask: your AI assistant inside your CRM

The real difference between Simply Ask and other AI tools? Context. Simply Ask knows the exact fields in your workspace, because the data model is metadata rather than code and it reads that model directly. Not a generic schema. Your organization’s fields. So if you added a field for “availability date” or “preferred travel distance” last month, Simply Ask knows about it.

What can it actually do?

• Search and match candidates based on a vacancy
• Generate CVs in your brand template via CV formatting
• Rewrite candidate introductions to match a specific vacancy. Relevant experience gets highlighted, less relevant work history becomes more concise, and the formatting fits what the client expects
• Create tasks without leaving the conversation you’re in
• Draft emails for hiring managers or candidates
• Process files like CVs and job descriptions via CV parsing
• Query data using natural language instead of building reports
• Research companies when you need context on a potential client

And then the part that separates it from a fetch-and-draft assistant: it can rebuild the product itself. Describe an object or a set of fields in plain language and Simply Ask generates the metadata for them, which means you change your data model by asking for it. Describe a goal and it drafts or edits a whole workflow. Before you ever see a drafted workflow, a validator on the server checks the node kinds, the connectedness, the dangling edges and whether every object and field reference actually resolves. Anything that fails goes back into the prompt rather than in front of you, so you never approve a workflow that cannot run.

Here’s the boundary, and it’s the important sentence on this page: none of that goes live by itself. Every write, a field on a candidate, a task, an email, a new object, a workflow, arrives as a proposal a person approves, and the approval goes into an append-only audit log with a name on it. There is no autonomous mode to switch on. There is also no credits model: you connect your own AI key, your provider bills you directly, and the cost ledger shows what each model and each action cost, in euros.

It also works outside Simply. There’s an MCP server, so Claude Desktop, Cursor or anything else that speaks MCP can query your workspace under the same permission rules, with a public API and HMAC-signed webhooks underneath it.

This isn’t a future roadmap item. This is what happens when you combine AI data extraction with an AI engine that reads the same configuration your recruiters work in.

AI Matching: from thousands of candidates to a ranked shortlist

“AI matching” is a term everyone uses. But almost nobody can explain how their matching actually works. At Simply we can, because transparency isn’t marketing. It’s a requirement, and it’s the whole reason the formula is published on a web page instead of described in a sales call.

The system works as a funnel. Each stage narrows the field, and each stage is something you can inspect afterwards.

Stage 1: the vacancy turns itself into requirements. A job description is prose, and prose can’t be scored. So it’s parsed into structured requirement rows first. Each skill row carries a requirement level, a minimum proficiency, a minimum number of years and an explicit weight, on a 0 to 100 scale: required weighs 100, preferred 60, nice-to-have 25. Nobody has to hand-weight a vacancy before it can be matched, and you can correct a row you disagree with.

That is a deliberate choice about where the weighting comes from. Plenty of matching engines ship a fixed set of criteria with fixed percentages behind them, defended with a citation to selection research. The trouble is that the research answers a different question: which predictors correlate with job performance across populations, not what this client asked for in this brief. Here the weights come from the vacancy. If the client says the certification is non-negotiable and the degree is a nice-to-have, that’s what the arithmetic uses, and you can point at the row that says so.

Stage 2: retrieval, on meaning as well as words. Finding the candidates worth scoring is its own problem. Keyword search finds the phrase somebody typed into a note two years ago. Semantic search compares your query against embeddings held in the same database, so “project management in construction” surfaces against a vacancy for “coordination of civil engineering projects” even though the words don’t overlap. Hybrid runs both and merges the two rankings. And all of it runs behind your permissions: a record you may not open is a record you cannot find.

Stage 3: the arithmetic. The score is one division you can do by hand. Add up the weight of the requirements the candidate satisfies, divide by the weight of everything the vacancy asked for. “Satisfies” isn’t a mood: the candidate has to have the skill and clear both thresholds, the minimum proficiency and the minimum years. A row they fall half a year short on contributes nothing at all, however well the rest of the CV reads. And a missing must-have comes back as a named gap in its own right instead of being averaged into a friendlier number.

Stage 4: which lane produced the number. Two lanes compute a score from identical structured input. A model reads the profile and the requirement rows; the deterministic lane runs the division above. Every result is tagged with the lane that produced it, so the screen can tell you where a number came from rather than implying a model was involved. A workspace with no model key connected still gets real scores, permanently, and not as a degraded mode you’re meant to upgrade out of.

The result: a ranked list where every number can be taken apart. You can see exactly why candidate A scores 82 and candidate B scores 74, and which requirement row is responsible for the gap. That’s the point of explainability and audit: you don’t just see who fits best, you see why, and you can defend it to your client and to a regulator with the same sentence.

EU AI Act: why this matters right now

Recruitment AI falls under Annex III of the EU AI Act. That makes it high risk. From 2026 onwards, users of recruitment AI need to meet strict requirements. Many tools aren’t prepared for this yet.

Simply’s answer isn’t a badge. It’s a set of design decisions you can check.

Explainability. Every match comes back as the requirement rows behind it, not as one opaque percentage. Which rows were satisfied, which were missed, and by how much. With readable reasoning attached.

Transparency. The scoring formula is published, on the matching page, in arithmetic rather than adjectives. Satisfied weight over total weight, with required at 100, preferred at 60 and nice-to-have at 25. Two people can look at the same score and argue about it usefully, which is the actual test.

Human oversight. The AI ranks. You decide. Always. Nothing in Simply moves a candidate out of a process on its own, and a low score doesn’t reject anybody.

Bias. The arithmetic runs on requirement rows, meaning skill, level and years, and nothing else. We’re not claiming a model has no bias, because that claim wouldn’t survive contact with a serious audit. We’re claiming you can see what drove a number, name the row that produced it, and overrule it.

Auditability. Every score records the lane that produced it, and every approval goes into an append-only audit log with a name on it, including the attempts that were denied.

For many agencies, this still feels abstract. But the deadlines are approaching. And when you need to prove that your matching system is fair and transparent, you don’t want to be starting from scratch. At Simply, that’s been built in from day one.

Want to know more about data protection? The platform and your candidate data are hosted in the Netherlands, on infrastructure Simply runs, with ISO 27001 underneath. The AI processing runs inside the EU as well, or on your own key with your own provider. The models themselves are bought in, which is a separate question from where they run — and worth asking any vendor as two questions rather than one, because the single sentence about Europe usually only answers the first.

What actually changes in your daily work?

Let’s be honest. Most recruiters spend the majority of their day on admin. Filling in CRM fields, formatting CVs, searching the ATS, typing emails. The real work (having conversations, building relationships) takes a back seat.

With Simply Ask, you flip that.

Scenario 1: New vacancy just came in.

You type: “Match candidates for the Senior Java Developer vacancy in Amsterdam.” Simply retrieves from your database within your permissions, scores what it finds against the vacancy’s requirement rows, and hands back a ranked list with the satisfied requirements and the gaps named. In seconds, not hours.

Scenario 2: Need a CV for a client.

“Generate a CV for candidate Jan de Vries in our brand template.” Simply grabs the profile, formats it using your template via CV formatting, and the output is ready to send.

Scenario 3: Writing an introduction.

“Write an introduction for this candidate, tailored to the vacancy at TechCorp.” Simply analyzes the candidate profile and the vacancy, highlights relevant experience, and generates a tailored introduction. Work history stays intact, but what’s relevant for this vacancy gets positioned more prominently.

Scenario 4: Email to hiring manager.

“Draft an email to the hiring manager at TechCorp introducing these three candidates.” Done. Adjust where needed, send.

This is the difference between a tool that gives you information and a tool that does the work. Combine AI summaries from your conversations with an agent that can take actions, and your day looks very different.

Why this is different from ChatGPT

Every recruiter has played with ChatGPT by now. Maybe had it write a job posting or draft an email. But that’s something fundamentally different.

ChatGPT doesn’t know your ATS. It doesn’t know which fields you have, which candidates are in your database, or what your CV template looks like. It’s a generalist. Simply Ask is a specialist that lives inside your system.

The difference? Context. Simply Ask knows that “available” in your workspace is a dropdown with three options. It knows you have 12,000 candidates in the Utrecht region. It knows what your branded CV looks like. Contextual recruitment makes answers not just faster, but better. The AI makes fewer mistakes because it knows the constraints.

There’s a second difference that matters more in a compliance review than in a demo. Retrieval is permission-filtered before any text reaches a model, and your authority is re-derived on every request rather than trusted from a session. A recruiter and their manager can ask the identical question and get different answers, which is correct rather than a bug. If you can’t see a salary expectation on the record, you can’t get it out of Simply Ask by asking nicely.

And that transparency isn’t optional. Every action Simply Ask takes is traceable. Every AI conclusion can be traced back to the source data. That’s not just nice for you. It’s a requirement under the EU AI Act.

One honest limit while we’re here: Simply Ask doesn’t know anything that never reached the workspace. A call nobody recorded, an email in a personal mailbox, a decision made in a corridor. It won’t invent them either. The quality of the answers tracks how much of the work actually lands in the system.

Curious how other teams use AI? Read 11 practical uses of AI in recruitment or why AI is not a threat to recruiters.

The recruiter stays in control

Simply Ask and AI Matching are tools. Instruments. They make you faster and give you better information. But the decision? That stays with you.

The AI ranks candidates. You choose who to call. The AI writes an introduction. You adjust it and send it. The AI suggests a task. You confirm.

That’s not a limitation. That’s the whole point. The best recruiters won’t be replaced by AI. They’ll be strengthened by it. Less time on admin, more time doing what they’re good at: connecting people with the right opportunities.

Want to see it against your own vacancies? Request a demo and bring a role you closed last quarter. Pointing the matching at your own database and seeing whether the person you actually hired comes back near the top is a more useful half hour than any feature tour.