AI and trainers: making collaboration work thoughtfully

Tasks an AI tool can support

• Sug­gest an­swers on re­quest, sub­ject to ser­vice avail­ab­il­ity and checks for ac­cur­acy.
• Sug­gest re­ph­ras­ings or al­tern­at­ives for the train­er to review; their rel­ev­ance de­pends on the in­struc­tions and con­text provided.
• Help ana­lyse col­lec­ted ques­tions, sub­ject to col­lec­tion set­tings, con­fid­en­ti­al­ity rules and data qual­ity.
• Gen­er­ate draft sum­mar­ies, ref­er­ence sheets or quizzes for review and ap­prov­al before dis­tri­bu­tion.

Responsibilities and tasks where the trainer’s judgement remains central

• Ob­serve group dy­nam­ics and decide, in con­text, wheth­er to re­dir­ect them.
• In­ter­pret verbal and non-verbal cues cau­tiously, without re­du­cing learners to auto­mated in­fer­ences.
• Draw on pro­fes­sion­al ex­per­i­ence, con­tex­tu­al­ised cases and prac­tic­al nu­ances.
• Take re­spons­ib­il­ity for edu­ca­tion­al de­cisions and ex­plain those choices to the com­mis­sion­ing or­gan­isa­tion.

Divide the work appropriately: who does what?

An AI tool can sug­gest an ini­tial ex­plan­a­tion or re­ph­ras­ing; the train­er checks the con­tent and handles ques­tions re­quir­ing an as­sess­ment of the con­text. Eval­u­ate the value of this di­vi­sion of work along­side veri­fic­a­tion time, errors and learners’ needs. See the AI prompt toolkit for trainers.

The effect on sup­port time must be meas­ured in each con­text: com­pare a baseline period with a de­ploy­ment period, doc­u­ment­ing the volume of re­quests, the tasks trans­ferred and veri­fic­a­tion time. Any time freed up can be re­as­signed to design or learner sup­port. See the of­fer­ings for independent trainers and businesses.

Decide which answers can be delegated

Al­loc­ate tasks based on the con­sequences of an error, not just how easy it is to gen­er­ate an answer. Re­ph­ras­ing an ap­proved defin­i­tion, in­ter­pret­ing an in­di­vidu­al situ­ation and making an as­sess­ment de­cision do not carry the same risks. The more an answer af­fects a right, safety or a judge­ment about a person, the more im­port­ant it is to provide human in­ter­ven­tion before it has con­sequences.

For each pro­posed use, spe­cify the audi­ence, per­mit­ted sources, ex­cluded ques­tions and the person who makes the final de­cision. An as­sist­ant lim­ited to ex­plain­ing learn­ing ma­ter­i­als should not be presen­ted as cap­able of resolv­ing a con­flict or cer­ti­fy­ing com­pet­ence. This bound­ary must appear in learner in­struc­tions and in the train­er’s own prac­tices when re­using a gen­er­ated sug­ges­tion.

The de­liv­er­able is a del­eg­a­tion sheet: the task being sup­por­ted, ex­pec­ted output, con­sequences of an error, checks re­quired and a non-AI al­tern­at­ive. Rule out a use if no one can verify its out­puts under real op­er­at­ing con­di­tions. Ap­prov­al that is man­dat­ory in theory but im­possible in prac­tice is not a re­li­able ar­range­ment. This cri­terion helps you reject some forms of auto­ma­tion without re­ject­ing as­sist­ance al­to­geth­er.

Test the limits before expanding conversations

Pre­pare a set of test ques­tions sep­ar­ate from the ex­amples used to design the as­sist­ance. In­clude a re­quest covered by the course, an am­bi­gu­ity, an in­cor­rect premise and an out-of-scope ques­tion. The ex­pec­ted out­come may be an ex­plan­a­tion, a re­quest for cla­ri­fic­a­tion or re­fer­ral to a person: an­swer­ing everything must not become the qual­ity cri­terion.

For each trial, retain the ques­tion, con­text provided, answer re­ceived and reason for ac­cept­ance or re­jec­tion. Ex­am­ine ac­cur­acy, but also fi­del­ity to the ma­ter­i­als, lan­guage level and the ab­sence of in­ap­pro­pri­ate advice. A fluent answer that sup­ple­ments the course with an in­ven­ted rule must be cor­rec­ted or blocked ac­cord­ing to the ar­range­ments in place. Rerun im­port­ant cases after changes to sources or set­tings.

When pre­par­ing these trials, favour con­struc­ted situ­ations rather than learner con­ver­sa­tions con­tain­ing identi­fy­ing in­form­a­tion. France’s data pro­tec­tion au­thor­ity, CNIL, high­lights the prin­ciples of pur­pose lim­it­a­tion and data min­im­isa­tion: col­lec­ted data must serve a defined ob­ject­ive and be lim­ited to what is ne­ces­sary. This re­mind­er does not itself au­thor­ise col­lec­tion; cla­ri­fy re­spons­ib­il­it­ies, in­form­a­tion for the people con­cerned and re­ten­tion ar­range­ments with the ap­pro­pri­ate spe­cial­ists.

Make human handover workable in practice

An es­cal­a­tion route is only useful if learners know where to go and someone can take over the issue. Define a chan­nel, a person re­spons­ible for re­ceiv­ing re­quests and the in­form­a­tion genu­inely needed for han­dover. De­pend­ing on the con­text, a sum­mary writ­ten by the learner may be enough; there is no need to cir­cu­late an entire con­ver­sa­tion auto­mat­ic­ally. State real­ist­ic re­sponse ar­range­ments without im­ply­ing that the train­er is always avail­able.

The train­er must be able to cor­rect an ex­plan­a­tion without blam­ing a learner who relied on it. Pre­pare a cor­rec­tion mes­sage stat­ing what has changed, why and which ref­er­ence to use from now on. If an error re­veals a gap in the ma­ter­i­als, cor­rect the re­source, not just the isol­ated answer. A learn­ing-in­cid­ent log can record the prob­lem and its res­ol­u­tion without be­com­ing a file used to judge in­di­vidu­als.

Application: a question requiring a decision

The team keeps an an­onymised or en­tirely re­con­struc­ted case sheet: the ori­gin­al ques­tion, the limit en­countered, the ex­pec­ted re­fer­ral and any change to the course. It adds this to the test set. Col­lab­or­a­tion thus be­comes ob­serv­able through the qual­ity of the han­dover, rather than simply the volume of an­swers pro­duced. This scen­ario de­scribes neither an avail­able auto­mat­ic fea­ture nor a client result.

Key takeaway

The useful ques­tion is which tasks a tool can sup­port, with what con­trols and human re­spons­ib­il­ity. Any prom­ise of freed-up time or edu­ca­tion­al im­prove­ment must be veri­fied in the con­text of use.