Maturity models have done the AI conversation in talent acquisition a lot of good this year. They turned a vague argument about whether AI belongs in hiring into a clear one about what a company has actually deployed. They also gave leaders a sensible way to describe where they stand to a board.
A recent survey of around a hundred TA leaders at companies with 500+ staff found most of them at the first level of maturity. At that level, AI drafts and suggests while people make every decision. The same survey found interview scheduling to be the most planned investment for the coming year, and found very few companies willing to let AI near selection decisions.
While we agree with the above, where we differ is in the order these models suggest you do things in, and what that order does to coordination.
What the models assume
Almost every framework in this space describes the same ladder:
- At level one, AI advises and a person decides.
- At level two, AI acts within limits a person has set.
- At level three, it acts on its own within one area.
- At level four, it works across areas with little supervision.
This follows the advice to climb one rung at a time. Build confidence at each stage before trying the next. That advice rests on an assumption worth saying out loud, which is that partial capability gives you partial value. Half an implementation should deliver roughly half the benefit, which still leaves you better placed for the rest.
For most AI in talent acquisition, that holds. Think about resume scoring or job description design. What those systems produce is advice for a person to read. A system that gives reasonable advice half the time is still useful, and making it better moves you smoothly up the scale. The same is true of sourcing suggestions. You can stop anywhere on that ladder and get value that matches where you stopped.
Why coordination is different
Interview coordination does not produce advice. It is a chain of handoffs, and nearly all of its cost sits in the handoffs rather than the individual steps.
That changes what partial automation does. Automate half of an advice process, and you get half the advice. Automate half of a handoff chain and you keep every handoff you had. You also add two more, because the work now has to move into the tool and back out again.
Here’s what that looks like in practice: The system suggests times -> a coordinator reviews them -> the coordinator approves -> the system sends them out -> the candidate, interviewer, or an EA replies with a condition rather than a slot -> the reply lands in a mailbox -> the coordinator answers it. The coordinator then re-enters the outcome so the system can issue the booking.
Every one of those steps is a point where the work waits for a person. The process now has a state that lives in two places at once.
Compared with the manual process it replaced, that arrangement is often slower. The coordinator can no longer just pick up the phone, because doing so puts the system out of date. They have also gained a queue of approvals that did not exist before. This is not bad software. It is what “acting inside limits” means when the thing being limited is a negotiation.
Half the autonomy does not give you half the value
Coordination has a value curve that dips before it rises. Partial autonomy costs more than none at all. The benefit only arrives once the handoffs actually go away.
We see this most often in companies that did the sensible thing and rolled out limited automation across all their hiring. Six months later, the tool is used for simple screens and worked around for senior panels. Coordinator time has not really improved.
The step-by-step advice is good general guidance. You do not get sixty percent of the value from sixty percent of the autonomy. You get a second system to keep in step with the first.
How about multiple autonomy, treated as one?
The problem with most AI frameworks is that they often treat autonomy as a singular measure. So autonomy over selection and autonomy over coordination sit at the same level and carry the same risk rating. In practice they are nothing like each other.
Autonomy over selection means software making or shaping a decision. That brings real risk of discriminating, an obligation to explain how decisions were reached, and a genuine ethical weight. The caution in the research is completely justified. Very few companies plan to go near it next year, and we would not push them to.
Autonomy over coordination contains no decision about a person at all. Nobody is scored, ranked, advanced or rejected. The decisions are which of several imperfect arrangements to offer, who to write to first, and what to do when an interviewer pulls out. Get it wrong and you have sent an awkward calendar invitation, not produced an unfair outcome.
Putting those two things on one scale has an unfortunate effect. Teams delay the safest autonomous use case available to them while they work through the legal and ethical questions raised by the riskiest one. The caution is right. It is simply being applied to the wrong workflow, and the cost is measured in coordinator weeks.
What we suggest for 2027
Take the maturity model seriously as a description of where you are. But be more careful with it as a route map.
Work out which of your use cases produce advice for a person to weigh up. Advance those gradually, because that is exactly what step-by-step adoption is for. Then find the ones that are really about removing handoffs, and treat them differently. For those, pick a segment you can define, hand it over completely, and measure the result against the process it replaced rather than against the level you were on last quarter.
Coordination belongs in the second group. It is the least risky autonomous capability a talent team can buy, and the frameworks in circulation today make it look like one of the most advanced.
Evie was built in 2013 as an autonomous coordinator rather than a scheduling assistant. That means we have watched a fair number of companies try this in stages before handing over a segment completely. Across more than 300,000 interviews, including nine-round sequences, assistant-managed executive calendars and back-to-back cross-timezone loops, teams report scheduling roughly ten times faster, around 50 percent fewer reschedules, and close to three months of recruiter time returned each year.
You can see what a complete handover covers at evie.ai/how-it-works.
Leave a Reply