Childcare worker / early-years educator — tasks, one by one

Watching children and keeping them safe

Supervision here means predicting what a toddler is about to do and physically stepping in within a second, across a room of children who move unpredictably. Cameras and video analytics can record a room and flag some events after they happen, but they cannot catch a falling child or take a toy out of a mouth. Licensing rules in most places require a set number of qualified adults physically present for each group of children, so even a capable monitoring system would not reduce the staff on the floor without a change in those rules, and parents choosing a nursery are paying precisely for a trusted adult in the room.

The protection here comes largely from ratio rules, which are a policy choice and are periodically loosened when providers are short of staff; it does not show that cameras and remote viewing will stay marginal, only that they add rather than replace. A licensing change that let monitoring count towards a ratio, or a dated deployment where one adult supervised more children because of a system, would move this.

Feeding, changing, sleep and hygiene

These are dexterous tasks performed on a small body that wriggles, cries and needs reassurance at the same moment, which is the hardest case for robot manipulation and the one where a mistake is least tolerated. Machines already help around the edges: dishwashers, steam sterilisers and floor-cleaning robots take some of the cleaning, and bottle-preparation machines exist for homes. None of them changes, dresses or settles a child, and the routine is also where children learn eating, sleeping and toilet habits, which is teaching rather than handling.

Hard to automate is not the same as well paid or well staffed: the cleaning and kitchen share of the job can still be cut back by equipment or moved to other staff, which changes the hours a setting pays for without touching the care itself. A household robot that could reliably dress or change an infant, shown working outside a demonstration, would change this judgement; cheaper cleaning equipment would not.

Emotional support, behaviour and spotting concerns

Young children form attachments to specific adults, and much of what a carer does here works only because the child knows and trusts that person. Noticing a concern depends on knowing what is normal for this particular child and on watching them across weeks of ordinary play. Screening questionnaires and software that scores them exist, and speech or movement analysis may later flag patterns, but a flag still has to be weighed against the child in front of you and then delivered to a worried parent, which is a conversation about trust and tact that parents expect from a person and that safeguarding duties assign to people.

This does not establish that the conversation with parents stays face to face: messaging apps already carry much of it, and a drafted message can soften or blur a concern that needed to be said plainly. Nor does it establish that screening stays with the carer; a validated tool that flagged developmental delay from routine video, adopted by settings and accepted by parents, would move part of this task towards augmenting.

Planning activities and early learning

Writing a plan is text work against a known framework, which is exactly what language models do quickly: given the children's ages and interests they produce activity ideas, song lists and plans mapped to learning goals in seconds, and many carers are likely to use general chat assistants for this already. Leading the activity is a different task: holding the attention of a group of toddlers, adapting when it falls flat, and noticing who is joining in. The plan gets faster; the person who reads the room and is accountable to the inspector for it does not change.

Faster planning does not show that the time saved returns to the children; it may simply be absorbed by more documentation or thinner staffing. It also does not show that generated plans are good: plans that look complete but ignore the actual children in the room are an easy failure to miss. Evidence that settings reduced planning or non-contact time because of such tools, or that inspectors accepted generated plans as they are, would change the reading.

Observation records and daily reports to parents

The logging moved onto tablets and childcare apps before language models arrived: meals, sleep and nappies are tapped into a form and pushed to parents. What language models add is the written part, turning a carer's quick notes and photos into a readable observation linked to learning goals, or a friendly summary of the day. The record still depends on someone having watched the child, and medication and incident records carry legal weight, so a person checks and owns what goes out. It is the task most likely to change form soon, and the part most exposed to a draft that sounds right about a child nobody actually observed.

This judgement is about form, not about headcount: no carer is hired to write reports, so faster reports do not remove a post, but they can be used to justify more children per adult or less paid non-contact time. It does not establish that parents value a generated summary as much as a written one. A rule or regulator's guidance on machine-written records about children, or evidence that settings let a tool send updates without a person reviewing them, would change it.

Attendance, sign-in and admin

Most of this is structured and repetitive, and childcare management software already does it alone: parents sign children in and out on a tablet or phone, attendance feeds billing and funding claims, invoices and reminders go out automatically, and booking and waiting lists run online. Chat assistants can now answer routine questions about opening hours or fees. What is left is the exception: the unexpected person at the door to collect a child, which is a safeguarding decision a person makes.

Little of a carer's day goes on this, and in small settings and home-based care it is often done by the owner, so its automation removes office hours rather than care posts. It does not show that collection checks can be automated: identity verification at the door is a place where an error is a child leaving with the wrong adult, and a dated deployment where collection was released without staff confirmation would be the evidence to watch.