Do you pay extra for AI in recruitment software?

Nearly every agency I speak to now has AI on their quote somewhere, and almost nobody can tell me what it will cost them in a normal month. That is a reasonable state to be in. AI is charged in at least three different shapes across the market, the shapes behave completely differently as you grow, and a demo never shows you which one you are looking at.

This piece is about how to read those shapes, how to estimate your own number before you sign, and which AI features are worth a line on the bill at all.

Why AI is charged separately in the first place

AI costs the vendor money every time it runs, which is different from the rest of the software. Once a CRM screen is built, ten thousand people opening it costs the vendor almost nothing extra. Every time a model reads a CV or drafts an email, somebody pays a supplier for that run, so vendors either price it as usage, wrap it in a per-seat add-on, or absorb it into the licence and manage the volume with limits.

That is the whole reason for the confusion, and it is worth knowing because it tells you what a vendor is optimising for. A vendor charging per action has an incentive that grows with your usage. A vendor including it has an incentive to make the AI cheap to run and to cap the heaviest cases. Neither is wrong, and you should know which one you have agreed to.

The three shapes of an AI charge

AI on a recruitment CRM quote is almost always one of three things: a per-seat add-on, a consumption charge counted by the action or by credits, or included in the licence with fair-use limits. The differences show up in your behaviour rather than on the first invoice.

A per-seat add-on is predictable and it punishes uneven adoption. If half your team uses AI and you are paying for all the seats, your effective cost per actual user is double the quoted figure. Check whether the add-on can be bought for some seats rather than all of them, because many cannot.

Consumption charging is honest about cost and it changes how people work, usually for the worse. The moment a consultant knows a summary costs something, they start rationing it, and a feature that gets rationed stops being part of the process. I have seen teams leave a genuinely useful tool switched off for a month because nobody wanted to be the person who ran up the bill.

Included-with-limits is the easiest to plan around, and the limits are the thing to read. Ask what the limit is per user per month, what happens when it is reached, and whether bulk work counts against the same pool as everyday work. A limit that is generous for individual actions can disappear the first time somebody runs a thousand CVs through it.

What counts as one AI action

The unit is defined by the vendor and it is rarely defined the way you would guess. Parsing one CV might be one unit, or one unit per page, or one unit for the parse and another for the summary it produces. A conversation with an assistant might be one unit for the whole exchange or one per message. Matching a job against your database might be one unit for the search and one per candidate scored.

Ask for the unit definition for the five things you would actually do most, in writing, with an example. If the answer is vague, price your estimate at the least favourable reading, because that is what the invoice will do.

There is a second question hiding behind this one, which is whether retries cost you. When the AI produces something a consultant is not happy with and they run it again, ask whether that second run is billed. In a busy week that difference is not small.

How to estimate your AI bill before you sign

Take one real week from your own agency and count the events, then multiply against each vendor's unit price. You need five numbers: CVs received, adverts or job descriptions written, outreach emails drafted, calls or meetings you would want summarised, and candidate searches run against your own database. Every one of those is already countable from your current systems or from a tally sheet on one consultant's desk for a week.

Multiply by four for a month and by your headcount, allowing for the fact that adoption will be uneven at first and heavier later. That gives you a range rather than a number, which is the honest answer, and a range is enough to see whether one quote is in a different league.

Then do the comparison that actually decides it. Work out what the same volume of work costs you in consultant hours today at their loaded hourly rate. A feature that costs real money and removes more expensive time is a good purchase, and the only way to know is to have both figures in front of you.

Which AI features are worth paying for

The AI worth paying for removes typing or reading that happens many times a day, and the AI that rarely repays a per-action charge is the impressive kind that runs once in a while. Parsing a CV into a record, drafting the first version of an advert, summarising a long call, sorting inbound replies so the interested ones surface first, and searching your own database in plain language all clear that bar, because each one happens constantly and each one currently costs a consultant minutes.

The features to be careful with are the ones that make a judgement you will have to check anyway. Scoring candidates against a role, predicting who will accept, ranking a shortlist. These can be genuinely useful, and their value depends entirely on whether the software tells you how sure it is. A score with no confidence attached has to be checked every time, which means you have bought the work rather than removed it.

Our AI reports how sure it is on every judgement and hands anything below three-quarters confidence back to a recruiter rather than writing it to the record. That design costs us more to run than simply answering every time, and it is the difference between a feature that saves time and one that quietly creates a checking job.

Charging shape	Predictable	What goes wrong	What to ask
Per seat add-on	Yes	You pay for seats that never use it	Can we buy it for some seats only
Per action or credit	No	Consultants ration it and adoption dies	Define a unit, and are retries charged
Included with limits	Mostly	Bulk work eats the monthly pool	What is the cap, and what happens at it
Included, no stated limit	Yes, until it is not	Terms change at renewal	Is the limit in the contract or the blog

The questions to put to any vendor

Five questions get you everything you need, and all five have short answers if the vendor has thought about it. What is the unit, and what does one of my real tasks cost in units. Can I cap spend, and what happens when the cap is hit. Are retries and failed runs charged. Does the AI tell me how sure it is, and what does it do when it is unsure. And does using AI more heavily ever change my per-unit price, up or down.

Put them in an email rather than a call. The answers you get back in writing are the ones you can hold at renewal, and the speed of the reply tells you something as well.

My own view

I would rather pay a clearly higher licence with AI included than a lower one with a meter running, and not because the arithmetic always favours it. It is because I have watched what a meter does to a team. Consultants are sensible people who do not want to be the reason the bill went up, so they hesitate, and a feature that gets hesitated over never becomes a habit. You bought software to change how the work gets done, and that only happens when nobody has to think about the cost of using it.

---

Anshika leads Customer Success at Recruitly.