The AI features a recruitment CRM should have in 2026

Two years ago, adding AI to a recruitment product meant putting a text box somewhere and letting a language model write paragraphs. Every vendor did it at roughly the same time, and it is now worth almost nothing as a point of difference. Writing is the cheapest thing these models do, and job adverts were never the bottleneck in an agency.

The features that matter now are the ones where the software does something on your behalf and then has to be right. That is a much harder problem and it is where products separate. Below is what to look for and the questions that tell you whether it has been built properly or bolted on.

Reading a CV properly, including the parts nobody typed in

The first thing a recruitment CRM should do with AI is read documents so your consultants do not have to retype them. That means a CV arriving by email becomes a candidate record with name, contact details, work history, skills and dates filled in, and the same for a job specification arriving from a client.

Parsing is not new, but the quality difference between an old rules-based parser and a modern one is large enough to change behaviour. Old parsers handled tidy CVs and mangled the rest, so consultants checked every field and were retyping anyway. A parser that is right most of the time and honest about the rest changes the job from typing to confirming.

Ask on the demo to parse the worst CV you have, with its two columns, its photograph and its dates written as "Jan 19 to present". Then ask what the product does with a field it could not read. The right answer is that it leaves the field empty and says so. A plausible value with nothing marking it as a guess is the thing that quietly poisons a database.

Sorting the inbox before a consultant opens it

The most valuable AI feature in a recruitment CRM is the one that reads inbound replies and sorts them, because every consultant already does this by hand every morning. A hundred replies to an outreach campaign contain a handful of interested people, a pile of polite declines, a few out of office messages, some asking a question and some asking to be removed from your list.

Doing that sorting by hand costs the first hour of the day and it costs it at the worst possible time, because the interested replies are the ones where speed decides whether you get the candidate. Software that reads each reply as it lands, works out what the person meant, and puts the interested ones at the top gives that hour back and improves the outcome at the same time.

This is also the feature where the difference between a chatbot and a classifier is easiest to see. Sorting a reply is a decision with a small number of possible answers, so the software should give a typed answer and a number saying how sure it is, rather than a paragraph a human has to read and interpret. If a vendor shows you this feature as a summary written in prose, they have built the demo rather than the tool.

An assistant that can see your data, and agents that do specific jobs

There are two useful shapes for AI in a CRM and they do different work. An assistant answers questions about your own data in plain language, and an agent does one defined job repeatedly without being asked each time.

The assistant is the one people picture. It is useful when it can actually see the account: which of my candidates have not been contacted this month, what did we agree with this client in June, show me everyone who interviewed for a role like this one. A general chatbot cannot answer any of those, because the answers are in your records rather than on the internet. Ours is called Simi, and the interesting part is not the conversation, it is the access.

Agents are less visible and usually more valuable. An agent watches for a thing happening and does a defined job when it does: reading an inbound CV and creating the candidate, keeping a job's fields filled from your own lists, writing up what was said on a call, preparing a brief before a client meeting. Each one is narrow, which is what makes it checkable. We run a set of these rather than one large system that claims to do everything, for the same reason an agency gives people job titles.

Search and summarisation over calls, interviews and meetings

A recruitment CRM in 2026 should let you search inside recorded calls and interviews by what was said, because otherwise the recordings are storage rather than information. Every agency now records far more than it can listen to. A consultant who wants to know whether a candidate mentioned a notice period has a one hour recording and no way in.

Searchable recordings change what recording is for. Instead of evidence kept in case of a dispute, it becomes a source you can query: find the moment salary came up, find whether anyone actually said the role was hybrid. We support searching recordings by what was said, and it surprises people on a demo because they had stopped expecting anything back from the pile.

Summaries are the other half. A summary that lands on the record as a note, automatically, is worth more than a better summary that a consultant has to go and fetch. The test is where the output goes, not how well it reads.

The feature that matters most: AI that says when it is unsure

The single most important AI feature in a recruitment CRM is the ability to decline, because a model will otherwise answer every question you ask it including the ones it cannot possibly know. That is not a flaw in a particular product, it is how these systems behave, and the only defence is a product that scores its own confidence and stops below a line.

Picture an email that says "Sara asked me to send this over" and an agency with a Sarah in permanent and a Sara Collingwood in contract. A system that always answers will pick one. It will pick it confidently, write it to the record, and the resulting line will look exactly like a line somebody verified. Multiply that by a couple of thousand emails a month and you get a database that is subtly wrong in a way no audit will ever find, because nothing in it is marked as a guess.

Ours reports how sure it is on every judgement and hands anything below three quarters back to a recruiter, with nothing written. That line costs us: it means a consultant still does some of the work. It also means that when the software does write something, it is worth reading.

Feature	What good looks like in 2026	What to ask on the demo
CV and job spec parsing	Fields filled from a messy real document, blanks left honestly	Parse the worst CV you own and look for invented values
Reply sorting	A typed answer per reply with a confidence score, interested ones first	Show me an ambiguous reply and tell me what happens to it
Assistant	Answers questions about your own account, not the internet	Ask it something only your data could answer
Agents	Narrow jobs that run without being asked, each one checkable	Name every agent and what it is allowed to write
Calls and interviews	Searchable by what was said, summaries landing on the record	Find a sentence inside an hour long recording
Confidence	A number on every judgement and a line below which a person decides	What is the threshold and what happens under it

How quickly does new AI capability reach you?

The gap between something becoming possible and it appearing in your account is a feature in its own right, and in a field moving this fast it may be the most important one. A product that was impressive two years ago and has not moved since is a product you will be stuck with for the length of your contract.

The way to check it is unglamorous. Ask when the vendor last shipped anything, ask to see the record of it, and ask whether that record is public and dated. We ship every week and publish a dated changelog in public, which is a low bar that surprisingly few vendors clear.

My own opinion, from building these things rather than selling them, is that the writing features are finished as a competitive matter and the judgement features are just starting. Anything that produces a paragraph for a human to read is easy and nearly free. Anything that makes a decision and writes it to your database is hard, and the only version of it worth having is the one that knows when to stop and ask.

---

Siva is an engineer at Recruitly.