Mentor Support in Self-Paced Online Learning

Pattern · All patterns
Evidence · 101 claims (53 for, 42 mixed, 6 against) · 80 studies (16 causal, 16 associational, 12 review, 11 quant-synthesis, 10 qualitative, 8 theoretical, 7 design), q1–q4 · 12 of 80 report an effect size · 91 claims rest on one study

Description and scope

Support for learners who work through online courses on their own schedule, from mentors, coaches or instructors who are not lecturing but are available to help: answering questions, giving feedback, helping learners plan and keep pace, and reaching out when a learner stalls. The decisions it covers are how much contact is scheduled and how much is on demand, what triggers outreach (inactivity, missed milestones, failed assessments), what a contact does (content help, feedback on work, planning and self-regulation, encouragement, referral), and how human mentors and automated or AI support divide the work. Grain size is a course, or a programme of self-paced courses with one mentor across them.

It applies the community of inquiry principle (teaching presence and social presence when there is no class meeting), the self-regulated learning principle (planning, monitoring and help-seeking that a self-paced learner must do alone), instructor accessibility and social presence. The course-shape decisions (template, module structure, discussion design) belong to online course design; this pattern is the people-and-outreach layer on top of a course. The study configurations below are evidence; the response-dependent policy is an untested design proposal. Use it to gather better evidence about a programme's learners, not to certify that a learner is "at risk" or "self-regulated".

What the evidence does and does not test. The wiki's causal evidence comes from four bodies of work, none of them a self-paced competency-based programme: (1) proactive student coaching and advising in US colleges (one randomized trial of a commercial coaching service, two matched comparisons of a city coaching programme, and a practice-guide synthesis of eight advising studies), which moved persistence by a few percentage points and did not detectably move completion or grades; (2) automated outreach (a randomized chatbot outreach trial in two large online undergraduate courses, a null text-messaging trial for college transition, and a pre–post check of analytics-triggered feedback emails); (3) planning prompts at the start of MOOCs (field experiments that raised completion); and (4) self-regulation supports in online and computer-based environments (meta-analyses of SRL interventions and metacognitive prompts). Evidence on teaching and social presence in online courses is correlational and qualitative. In the wiki, no study compares mentor contact schedules or doses, tests when to hand a learner from automated help to a person, or measures a mentor model inside a self-paced competency-based programme; the competency-based evidence recorded here is descriptive completion data. Treat those as open questions with proposed defaults, labelled below.

Inputs: establish the design brief

Field	Record or elicit
Learner and starting observation	Prior online or self-paced experience, prior knowledge of the course domain (novice, intermediate, advanced, or ?), work and caring commitments, the learner's own plan and expected pace, and their first-week activity (logins, first submission). Preserve unknowns; do not infer readiness from demographics.
Context and activities	Mode (online-self-paced, online-instructor-led, cohort or rolling enrolment), whether there are deadlines or only an end date, what help exists (mentor, course instructor, tutoring, peer forum, AI tutor or chatbot), each one's hours and response time, and the mentor's caseload.
Learner-valued goal	Ask what the learner wants from the course or programme and by when (a credential, a promotion, a skill for current work, finishing quickly). Record disagreement with the designer objective, for example a learner who wants speed when the programme wants mastery.
Designer objective	What counts as success: completion, pace, mastery of the competencies, or all three; which assessments are judged by a person (performance tasks, simulations, labs) and which are automated.
Outcome	Separate measures for activity (logins, submissions), progress (milestones, course completion, term-to-term persistence), learning (assessment results, judged by an instrument independent of the help given) and uptake of support (who used which help, and how often), each with a horizon.

A request to "add mentoring" or "improve retention" does not specify these inputs. Ask for them before choosing a contact schedule, an alert threshold or a predicted gain. Check first that outreach can say something specific (which activity to do next, what a failed attempt shows), and that someone has the time to answer the replies it produces.

Sequence and conditional policy

1. Before or in the first week: a planning contact. Ask each learner to write an open-ended plan for how they will work through the course, the obstacles they expect and an if–then response to each (about ten minutes, once). A mentor reads it and replies personally within the first week. Contact early: in one continuing-education programme, online attrition during the orientation week was twice that of the first week of on-campus classes, and once instruction had begun and learners had met the instructor, online and on-campus drop rates were about the same, so the excess loss came before anyone had made contact. Optionally add a short belonging message in the same contact (under 30 minutes, once): that doubts and early setbacks are common among new students and pass, written up in the learner's own words. In a 22-college trial this, delivered online before college, raised full-time first-year completion, but only where the institution offered students real opportunities to belong, so it is worth adding only alongside the personal reply and check-ins, not instead of them. The planning activity and the belonging message are evidence-based in other settings (see below); the personal reply is a proposal.
2. Weeks 1–4: scheduled check-ins plus signals. Proposed default, untested: a short scheduled contact at each of the first two milestones (about one a fortnight), each naming what the learner has done and the next concrete step. In parallel, define the signals that will trigger outreach (for example days without activity, a missed milestone, two failed attempts at one assessment) and write the thresholds and who answers the flags into the programme's records before launch. Set an alert budget (the share of learners mentors can actually contact in a week) and fit the threshold to it; in one replay of programming-course data, a plain consecutive-failures rule caught more persistent failures within a fixed 20% budget than a logistic-regression model did. Do not add scheduled contacts on the assumption that more hours mean more effect: in 60 studies of coaching teachers, coaching hours were not associated with larger effects (a transported finding; no learner-coaching dose study is in the wiki). Dropout does not concentrate at a single lesson in most online courses, so watch progress throughout, not at one checkpoint.
3. Whenever a signal fires: triggered outreach with specific content. Send a message that names what the data show and the specific activities or help to use (not a general "we noticed you've been away"), and offer a person. Automated outreach can carry routine reminders and course-specific nudges; a mentor takes any reply that concerns planning, motivation or a failed performance task. Record what was sent, to whom and what the learner did next. If the same signal fires again after a first contact, a person goes back and asks what the learner is trying to do before explaining again; encouragement alone, in one classroom study, steadied a learner without getting her past the point where she was stuck (proposed rule, drawn from qualitative classroom evidence).
4. On demand: divide content help. Routine questions and practice hints can go to automated or AI help available at any hour. Feedback on graded performance work (labs, cases, simulations, written analyses) comes from a mentor or instructor, and a mentor reviews work-in-progress where the platform lets them see it. Hand off from AI to a person when the learner asks for one, repeats an answer-seeking request without progress, fails the same task twice, or raises anything about pace, wellbeing or withdrawal (proposed rules, untested). When requests for a mentor queue up, answer learners who have not yet been helped before repeat requesters: in a simulation of course office hours this raised the share of students helped at all (82% against 70% under first-come-first-served at normal load) without changing how many requests were resolved or how long students waited.
5. Middle to late course: fade scheduled contact, keep triggered contact. Once a learner has met two consecutive milestones on their own plan, move them to milestone-only contact; keep signal-triggered outreach for everyone. Learners in one study wanted feedback outreach more often than twice a term, so do not fade to nothing.
6. Reobserve and audit. At each milestone check progress and learning separately, and at term end check who used which help and who received triggered outreach, by subgroup, against outcomes. Revise thresholds and messages from that audit rather than from mentor impressions alone.

Observation	Discriminating probe	Proposed action, with its uncertainty
No activity for the agreed number of days early in the course	Short message asking which of: unclear what to do next, too little time this week, content difficulty, reconsidering the course; check whether the learner ever started a first task.	If time or planning: revisit the plan and its if–then obstacles. If content: point to the specific resource or a mentor session. If reconsidering: a person contacts them. The categories are a proposal, not a validated diagnosis.
Steady activity but repeated failed attempts on one assessment	Look at what the attempts got wrong; ask the learner to explain their approach to one item.	Mentor feedback on the work, with a revise-and-resubmit, rather than more practice of the same kind. If attempts show answer-guessing, address strategy before content.
Heavy use of an AI tutor, requests shifting toward "just give the answer"	Ask the learner to solve one comparable item without the tool; compare with assisted performance.	If unaided performance lags: a person discusses how to use the tool (attempt first, ask for hints), and practice includes unaided items. Assisted performance alone does not show learning.
Learner on pace but reports feeling isolated or unsure of standards	Ask what they would want from contact (feedback, a peer, a person to check in).	Add a personal check-in or a peer connection; in the message, acknowledge the feeling, recast the difficulty as an ordinary part of learning, and connect it to where the learner is heading (a theoretical account, not a tested script). Presence evidence in the wiki is correlational; this branch may change satisfaction without changing completion.
The same signal fires again after a first contact (still inactive, or failing the same task)	Ask the learner what they are trying to do or reach on this task, rather than repeating the explanation; check whether the first contact addressed that.	A person returns with questions, then targets help at the intention the learner names. Reassurance alone may steady a learner without moving them on. Drawn from a single qualitative classroom study with AI alerts; untested in self-paced courses.
Class- or cohort-level signals improving while individual flags are quiet	Look at the trend across learners before contacting individuals.	Holding back individual contact can be a deliberate choice; keep triggered outreach for anyone who flags. One qualitative classroom episode; untested.
No reply to outreach	Check whether messages are delivered and read; try a different channel or a person.	Escalate to a person after one unanswered automated message. Whether escalation helps is untested in the wiki.
Fast progress, high scores	Check that assessments are independent of the help given.	Reduce scheduled contact and offer stretch work; do not withdraw access to help.

Choosing configurations from evidence

• What to expect from proactive human coaching. InsideTrack coaching has potentially positive effects on persistence [+M]: in the Bettinger & Baker randomized trial (seven lotteries, 3,527 college students), 66.4% of coached students were retained at 12 months against 61.4% of the comparison group (effect size 0.13). At 24 months degree completion did not differ detectably [~M] (35.2% vs 31.2%, effect size 0.11, p < .10, 1,346 students). The service cost about $390–$500 per student per semester [~M]. Plan for a persistence gain of a few percentage points, not for coaching to fix completion on its own. On dose, the only recorded evidence is about coaching teachers on their instruction: across 60 causal studies, more coaching hours were not associated with larger effects [~M] (exploratory meta-regression, hours inconsistently reported). It does not test learner coaching, but it gives no reason to buy more contact hours rather than better-targeted contact.
• Coaching in another programme, with weaker designs. Success Boston Coaching has potentially positive effects on progressing in college [+M]: two matched comparisons of Boston public-school graduates (second-year enrolment 83% vs 75%, effect size 0.30, and 82.6% vs 78.3%, effect size 0.17). The scale-up study found no effect on cumulative GPA [~M] (2.37 vs 2.36, effect size 0.01). Coaching moved enrolment, not grades, in both programmes recorded here. A WWC practice guide pooled eight studies and found comprehensive, integrated advising improves progression, achievement and degree attainment [+M], with no effect size printed on the claim page; "integrated" advising bundles several services, so it does not say which part did the work.
• Automated outreach in online courses. In a randomized trial in two large online undergraduate courses (1,568 and 915 students), chatbot outreach raised the share earning an A or B by four percentage points [+M] (from 61%). Effects differed by course [~M] (fewer D, F or withdrawals in one, fewer withdrawals in the other), concentrated in the middle of the grade distribution [~M], and did not carry to other courses or next-term enrolment [~M]. Even the work it moved was course-specific: practice-quiz completion rose by five percentage points in one course and assignments did not change in the other [~M]. Satisfaction is not a substitute outcome: 82% of survey respondents recommended the chatbot, yet attitudes showed no consistent treatment effect, and only about half answered, with response rates differing across groups [~M]. Part of the effect ran through people: the chatbot raised tutoring attendance by about two percentage points [+M] from baselines of 7% and 2.4%, carrying about 18% of the grade effect. So outreach that connects learners to human help is one configuration with a tested benefit; expect it to be course-specific and short-lived.
• Generic messaging is not enough. Text-message college-transition messaging did not increase persistence [-M] (GEAR UP evaluation), and the authors suggest, without testing it, that broad advisor caseloads limited what the messages could lead to [~W]. Outreach is only as useful as the help behind it.
• Analytics-triggered feedback emails. In one course, models trained on one semester identified at-risk students in the next with AUC above 0.8 [+M], and after a feedback email naming the activities they had not used, more than 30% of flagged students visited those activities within two weeks [+M]. That is a pre–post count with no comparison group and no learning outcome. Respondents wanted the feedback more often than twice a semester [+W], but only 9.27% of flagged students answered the survey.
• Do not take a dashboard vendor's retention figures at face value. The Course Signals early-warning system's reported retention (93.24% vs 69.4%) may be explained by confounding [-M]: students who took more courses both met the system more often and persisted longer. Georgia State's advising system paired prediction with advising capacity and grants, and reported narrowed graduation gaps [~W]; those are institution-reported outcomes with no comparison group, and the claim credits the advising capacity, not the prediction.
• Prediction models can miss some learners more than others. On a large open-university dataset (25,690 students), a standard logistic-regression model had unequal true-positive rates across subgroups [-M] by sex, age, deprivation and disability. Audit a model's misses by subgroup before letting it decide who is contacted. Adding demographic variables is not a fix: evidence is mixed on whether they improve prediction at all [~W], and a review of AI in higher education finds that early-alert systems and chatbots may serve marginalized students only with governance, transparent data practice and continuing human oversight [~W]. A simpler trigger can do as well: replaying a programming course's records at a fixed 20% alert budget, a consecutive-failures rule caught 75 persistent failures (precision 0.68) against 66 for logistic regression and 62 for a student-history model [+W]. That is one dataset and a replay, not a trial of outreach. Whoever answers the flags also needs time and skill: teachers named workload, habit, poor data and lack of skill as what kept them from data-informed practice, and none regularly collected data on their teaching [-W], and educators trying to act on analytics asked for training, time and institutional backing [+W].
• Planning prompts at the start. In three HarvardX MOOCs, open-ended planning prompts at the start increased completion by 29% [+M] (relative, no standardized effect size) and certificate payment by 40% [+M]; the effect was largest for students enrolled in traditional schools [~M], and plans focused on specific times forecast lower completion [~M]. Across 94 tests, if–then implementation intentions improve goal attainment [+S] (d = .65), and in two randomized online-course experiments (17,963 learners) an eight-minute plan-and-obstacles activity made learners in countries classed as individualist 32% and 15% more likely to complete, with no effect for those in countries classed as collectivist; the text analysis found it worked when the obstacle named was predictable and surmountable (work or family), not when it was lack of time. Use the obstacle the learner names, not their country, to decide what follow-up they need.
• Self-regulation support online. Across 15 studies (5,519 participants), SRL interventions in online and blended environments had a pooled effect of d = 0.65 on learning outcomes [+M]; metacognitive prompts in computer-based environments [+S] raised learning by g = 0.40, with larger effects when prompts were task-specific, adaptive and paired with feedback, and a 35-study review of online and MOOC settings found prompting effective but too varied for one effect size. In six MOOCs (4,831 learners), self-reported goal setting and strategic planning predicted reaching one's goal [+M] (54% higher odds per SD of goal setting), while help seeking predicted lower odds of certification [~M] (27% lower per SD). That last result is associational; learners who struggle are more likely to seek help, so it is not evidence that help harms, but it does say that help-seeking alone does not carry a struggling learner through. The same study found learners inclined to seek help less likely to pass an assessment straight after a lecture [~M]. An AI coach can carry part of the self-regulation work: a review reports a randomized trial with 86 K-12 students in which agentic coaching that prompted planning, monitoring and reflection raised SRL strategy use [+W] (p < .05, no effect size, strategy use rather than achievement). How much any learner gains from controlling their own path depends on prior knowledge and self-regulatory skill [~W], which is why the brief asks for both rather than assuming them.
• Who gives feedback on graded work. In a randomized comparison with 35 students, those given instructor feedback improved their lab scores more than those given AI-generated feedback [+M] (group coefficient −4.97 for AI, p = 0.04). The sample is small and the design is coded r1; treat it as a reason to keep people on performance-task feedback until a programme tests otherwise. Feedback in general improves learning by a medium amount on average (d = 0.48), most when it carries information about how to proceed (d = 0.99), and more than a third of feedback interventions reduce performance [~S]: the content of a mentor's message matters more than its existence.
• AI help and unaided performance. Reviews of one randomized trial report that unrestricted GPT-4 access raised assisted practice by 48% but left students 17% worse once it was withdrawn [-M], and in 112 students' sessions with an LLM, help-seeking shifted toward answer- and verification-seeking as a session went on [~M]. In K-12 AI-based instruction a review reports AI-generated information combined with teacher intervention associated with better performance, participation and confidence [+M], while real-time alerts sometimes overloaded teachers' attention [~M]. In a survey of 478 students, weaker effort regulation and more general help seeking went with more reported dependence on AI [~M] (associational). In one computing course, office-hours and forum use fell over the semesters when LLMs spread and rose again in the term the course built LLM use and retakes into its design [~M] (descriptive counts, no test), so track human-help uptake whenever AI help is added. When AI does the practice, an unaided check gives learners a reason to learn: in a small course that replaced lectures with a self-paced AI-assisted block, students ranked the weekly closed-book test second only to that block among components they would miss [+W] (18 respondents). These support step 4's hand-off rules as a direction; none tests a hand-off rule.
• Presence and personal contact. In a national survey of undergraduates moved online, personal instructor messages showed the largest association with satisfaction [+M] (68% vs 47% satisfied; motivation a major problem for 33% vs 53%). A meta-analysis summarised in a book chapter verified the importance of student–instructor, student–student and student–content interaction in distance learning [+M]. In an online case-based graduate course, co-regulation correlated more strongly with the three presences than self-regulation did [+M] (47 students, ρ = .65 with social presence). Distance students said their needs centre on course support and communication with tutors [+M], and online teachers said community online requires deliberate, planned effort [+M]. None of these measures completion or learning; they support scheduled personal contact as a satisfaction and motivation measure. In a 530-student survey, engagement with instructors and content went with less digital distraction [+M] (associational, small phi). Course structure and facilitation are separate: in an accelerated online MBA, course organization was rated highly while the instructor's facilitation behaviours fell below 4 on 5 [~M] (32 respondents), so a clear template does not stand in for a person who responds.
• Taking a course online, and belonging. For the starting point, an instrumental-variables study of college courses found that taking a course online instead of in person reduced student success and progress [-M]; mentor support is added on top of a format that, on average, costs learners something. In two randomized trials, brief belonging interventions improved college outcomes [+M]: in the larger (26,911 students, 22 institutions, delivered online in under 30 minutes before college) first-year full-time completion rose, most for students from groups whose progress had historically been lower (a gap reflecting circumstances and treatment, not a cause), and only where the college gave those students real opportunities to belong. Tutors' part in this is argued, not tested, in the community-of-inquiry literature: a sense of belonging can help promote and consolidate online learning [+W]. No effect size is recorded for either trial, and neither is a self-paced course.
• Responding to an alert. In one mathematics classroom with AI-generated emotion alerts (qualitative, single lessons), repeated alerts led the teacher back to a student, and her engagement reorganized only when the teacher asked what she was trying to do rather than explaining again [+W]; brief encouragement after an alert steadied another student only briefly [~W]; and the teacher chose not to intervene when the class-level display showed negative alerts declining [+W]. A theoretical guide proposes that a brief emotional response should validate the feeling, reframe the appraisal and connect the struggle to a larger goal [+W]. These shape the probe table's "signal fires again" row; none was tested asynchronously.
• What competency-based programmes look like without a tested mentor model. In a Mathematica evaluation of three community colleges' online competency-based IT programmes, 35% of participants completed a programme [~M], participants took on average under two terms to complete a first certification course and about four to complete a certificate or degree [+M], and gatekeeper-course completion was slightly lower than for comparison students, with credential differences varying by college [~M]. These are baselines for reading a programme's own numbers, descriptive or weakly compared, and say nothing about which support produced them.

Do not rank these results against one another: comparators (no coaching, matched non-participants, no messages, the same course before the email), outcomes (retention, grades, completion, activity visits, satisfaction) and populations differ. A precise abstention names the missing comparison, outcome or observation and the next way to resolve it.

Fitting the design to a situation

Ask for four facts when the brief omits them: what is the mentor's caseload and response time? (outreach without capacity behind it is the configuration that failed), is success completion, pace, or mastery, and how is each measured?, which work is judged by a person?, and who, by the programme's records, uses on-demand help now?

If the brief says…	Then change…	Basis
Learners new to online or self-paced study, or in their first term	Put the planning contact and a personal reply in the first week, before any content, and the first scheduled check-in before the first milestone; consider a brief belonging message in the first contact	Orientation-week attrition online is twice that of on-campus first weeks [~M]; after instruction starts, online and on-campus drop rates are about the same [~M]; Planning prompts raised MOOC completion [+M]; brief belonging interventions improved first-year college outcomes [+M] (college entry, not self-paced courses)
Experienced adult learners in work, choosing their own pace	Keep the plan-and-obstacles activity, and let the learner choose check-in frequency after the first milestone; keep triggered outreach	Implementation intentions worked when obstacles were work or family, not lack of time [+S]; fading schedule is untested
Secondary-school or younger learners	Route alerts to a teacher or family as well as the learner, and act on grades and attendance signals	Ninth-grade early-warning systems on grades and attendance were followed by improvements [+W] (a research agenda's summary); parent text alerts about missing homework raised performance [+M]
The goal is completing a credential, not one course	Plan coaching across terms and measure persistence each term; do not expect completion to follow from one term's support	Coaching raised retention at 12 months [+M]; no detectable effect on degree completion at 24 months [~M]
The goal is mastery judged on performance tasks (cases, labs, simulations)	Keep a person on feedback for graded performance work and on debriefing; use automated help for practice	Instructor feedback raised lab scores more than AI feedback [+M]; Feedback, analytics and structured reflection strengthen learning in simulations [+W] (a review's summary)
Online, self-paced, asynchronous, no class meetings	Use a consistent course layout with goals, due dates and feedback turnaround stated, and schedule personal messages rather than waiting for learners to ask; do not count the layout as facilitation	Taking a course online reduced college success compared with in person [-M]; A standard template heightened teaching presence in accelerated online courses [+W]; organization rated high while facilitation rated lower [~M]; Personal messages had the largest association with satisfaction [+M]
Rolling enrolment with no deadlines	Set learner-chosen milestones in the planning contact and trigger outreach on missed milestones, not on calendar dates; give a recommended order and pace rather than an open menu	untested proposal; dropout is not concentrated at one lesson in most online courses [~M]; reviews of behavioural economics argue that abundant choice without structure may add to dropout [+W] and that self-control problems explain much dropout [+W] (arguments about college and adolescents, not tests in self-paced courses)
Competency-based programme, pace set by the learner, credential spanning several terms	Plan mentor contact across terms; record time to each competency and completion against the programme's own comparison group, and take the gatekeeper courses as the place to concentrate triggered outreach	descriptive baselines only: 35% of consortium participants completed [~M]; gatekeeper completion slightly lower than comparison students [~M]; about two terms to a first certification course, four to a credential [+M]; the mentor model itself is untested
Little learner time each week	Keep contacts short and specific: name the one next activity; avoid plans phrased only as times	Time-focused plans forecast lower completion [~M]
Mentor caseload is large or mentor hours are limited	Use automated outreach for reminders and routing, reserve mentor time for replies and for learners with repeated failure; set an alert budget from what mentors can answer and fit the threshold to it (a consecutive-failures rule is a reasonable first trigger); in the request queue, answer learners not yet helped before repeat requesters; give mentors time and training to read the data	GEAR UP messaging did not raise persistence; broad caseloads may have limited it [~W]; alert threshold changes reshape advisor workload [~W]; a consecutive-failures rule concentrated alerts within a fixed budget [+W]; serving new students first raised the share helped [+W] and did not change requests resolved or wait time [~W] (simulation of office hours); workload and skill limit teachers' data use [-W]
No mentor at all; automated or AI support only	Use chatbot outreach that links to whatever human help exists (tutoring, office hours) and expect course-specific, short-lived effects; let the AI prompt planning, monitoring and reflection, and judge it on learning, not on how much learners like it	Chatbot outreach raised A/B grades [+M]; no spillover or next-term effects [~M]; agentic SRL coaching raised strategy use in one K-12 trial [+W]; high recommendation, no attitudinal effect [~M]
An AI tutor answers content questions	Require an attempt before hints, include regular unaided checks (a short closed-book test), hand off to a person on repeated answer-seeking without progress, and track whether use of human help falls	Unrestricted GPT-4 access harmed performance after withdrawal [-M]; help-seeking shifted toward answer-seeking within sessions [~M]; students valued weekly closed-book tests beside a self-paced AI block [+W]; office-hours and forum use fell as LLMs spread [~M]; weak effort regulation went with AI dependency [~M]; hand-off rules untested
High stakes (licensure, certification)	Keep a person on judgments that certify competence, and audit any risk model's misses by subgroup before using it to decide who gets contact	Standard prediction models had unequal true-positive rates across subgroups [-M]
Health professions programme	Add career or programme coaching that connects learners to the programme's next steps, and evaluate it against a comparison group	Career coaches appeared to expand exposure to and completion of health sciences programmes [+M] (associational)
Teacher education with clinical or field assignments	Give probing feedback that asks for evidence from the learner's own students, with a required revise-and-resubmit, and expect uneven progress	Probing feedback with revise-and-resubmit moved one candidate, not another [~W]; instructor access to work-in-progress enabled coaching [+W]
Business programme (for example an accelerated online MBA)	Make contributing peers visible and give feedback that is professionally relevant; learners evaluate peers for what they offer	Peer monitoring went beyond social presence in MBA learners [+W]; learner presence showed as intentions, metacognition and peer monitoring [+W]
Technology or computing courses with labs	Put instructor or mentor scaffolds and probes in online labs, and keep human office hours alongside AI help	Online labs seemed more effective to students when instructors were present with scaffolds and probes [+W]; students recommended office hours over LLMs [+M]
Concern about equity of who uses help	Make the first contact proactive for everyone, track uptake and outcomes by subgroup, and do not target outreach by group membership or add group variables to a risk model; check who answers any satisfaction survey before reading it	Only about 10% attended optional drop-in tutorials [~M]; chatbot effects differed by gender in one course [~M]; voluntary chatbot use was bimodal [~M]; survey response differed across groups [~M]; demographic predictors give mixed gains [~W]; alert systems serve marginalized students only with governance and human oversight [~W]

The evidence by question

• Does proactive coaching raise persistence and completion? InsideTrack coaching: retention up at 12 months [+M]; no detectable effect on completion at 24 months [~M]; Success Boston coaching: progression up [+M]; Success Boston: GPA unchanged [~M]; integrated advising [+M]; coaching hours and effect size, for teacher coaching [~M]
• Does automated outreach help, and how? Chatbot outreach raised A/B grades [+M]; partly through tutoring attendance [+M]; assignment effects in one course only [~M]; no attitudinal effect [~M]; text messaging did not raise persistence [-M]
• Does a brief start-of-course belonging message help? Belonging interventions improved college outcomes [+M]; belonging and online learning [+W]; drop rates similar once instruction starts [~M]
• What should a person do when an alert fires or repeats? returning with questions after repeated alerts [+W]; encouragement alone [~W]; deliberate non-intervention [+W]; micro-intervention functions [+W]
• What do competency-based programmes achieve as a baseline? 35% completion [~M]; time to completion [+M]; gatekeeper completion [~M]
• Do planning prompts and reminders help self-paced learners finish? Planning prompts raised MOOC completion [+M]; implementation intentions [+S]
• Does supporting self-regulation online improve learning? SRL interventions online and blended [+M]; metacognitive prompts [+S]; SRL and achievement [+S]
• Can early-alert models be trusted to decide who gets contact? Cross-semester prediction accuracy [+M]; Course Signals confounding [-M]; subgroup bias in true-positive rates [-M]; a consecutive-failures rule within a budget [+W]; demographic predictors [~W]; governance and oversight [~W]; staff data use [-W]
• When should a person rather than AI give feedback? Instructor vs AI feedback on lab scores [+M]; AI tutoring systems less effective than human tutoring [~M]
• Who uses on-demand help? Help seeking and MOOC certification [~M]; drop-in tutorial attendance [~M]; tutoring-centre visits by student group [~M]; help seekers and passing after a lecture [~M]; human help-seeking as LLMs arrived [~M]; help seeking and AI dependency [~M]
• How should a mentor's queue be served? new students first raised the share helped [+W]; no change in resolved requests or wait [~W]

Further evidence, not yet read against this model

Claims read against the model on 2026-10-19 that do not change it; each keeps the marker its recorded evidence allows. The other 31 claims once listed here are now used above. New links the linker adds below are still unread.

• Neurodivergent and neurotypical students both prefer some instructor involvement [+W] — about collaborative active learning in in-person computing classes (comfort ratings, small survey); consistent with keeping a person involved, but it does not bear on contact or outreach in self-paced courses
• Checkpoints, definitions and staged pacing improved clarity for novice programmers using an AI tutor [+W] — about the design of an AI tutor's dialogue, which this page leaves to AI Tutoring
• Reflection interventions in higher education: medium average effect, varying with duration and peer interaction [+M] — bears on reflection prompts a mentor might use in a check-in, but tests reflection activities, not mentor contact; it does not change the default

Illustrative design instance and observation record

Illustration only, invented, not a tested programme. A self-paced online certificate in health informatics for working adults, with rolling enrolment, a mentor with a caseload of 150, case- and simulation-based performance tasks, and an AI tutor for practice questions. In week 1 each learner writes a plan with two expected obstacles and if–then responses, and their mentor replies within three working days. Milestones are set by the learner (one module every two weeks on average). Outreach fires after ten days without activity or two failed attempts at one performance task: the first message is automated and names the next activity and the mentor's booking link; a second unanswered message goes to the mentor. Performance tasks are scored and given feedback by the mentor with one resubmission. One learner says she wants to finish quickly for a promotion; the programme wants mastery of each competency. Agree a pace that keeps resubmission time, and check at the end that completion, scores on an independent final case and help uptake were recorded separately and by subgroup.

Record: learner plan and named obstacles → contacts scheduled and actually made, with dates → signals fired, thresholds and who answered → message content and channel → learner's next action and time to it → help used (mentor, AI tutor, tutoring) → candidate interpretations and confidence basis → milestone progress and assessment results at the stated horizon. Preserve the learner's stated purpose and any change in it. An agent may summarize this record but must not fabricate missing inputs or assign numerical risk probabilities without a calibrated, subgroup-audited model.

Elements and limits

Coaching, check-ins, goal setting, feedback and learning analytics feedback.

This pattern is scoped to support around self-paced online courses and programmes: planning contact, triggered outreach, feedback on work, and the division between people and automated help. Course design itself, peer-learning formats, tutoring as instruction and the design of an AI tutor's dialogue need their own configurations and evidence. The causal evidence recorded in the wiki comes from campus and online college courses and MOOCs, mostly in the United States, measured on persistence, grades and completion over one or two years; it does not test contact frequency or dose for learners (the one dose finding is about coaching teachers), hand-off rules, a belonging message inside a self-paced course, what a mentor should do when an alert repeats (that evidence is one qualitative classroom study), or any mentor model inside a competency-based programme, and the diagnostic branches above remain to be tested with actual learners.

Claims

• Guiding support stimulated targeted interactions: teachers posted more messages to groups (14 vs 4) and performed more dashboard interventions than under mirroring [+W]
• Proactive tutoring increases learners' use of lexical help tools compared with reactive tutoring [~W]
• Proactive tutoring increases learners' use of metacognitive help tools compared with reactive tutoring [+W]
• Learners heed proactive promptings more systematically when a human tutor is accessible than when it is not [+W]
• Learners use navigational help less when a human tutor is accessible, contradicting the expected direction [~M]
• Support demand across participants is concentrated: the most active decile accounts for 34.6% of queries and 42.9% of participants raised exactly one query [~W]
• People value emotional and esteem support more from close humans than from AI systems, with no significant difference for informational support [~W]
• Teachers envisioned chatbots as extending teacher capacity and availability by providing additional access to support and redistributing planning workload [+W]

Related Patterns

• Online Course Design
• Self-Regulated Learning
• AI Tutoring
• Practice with Simulated People
• Competency-Based Learning
• Formative Assessment
• Cognitive Apprenticeship
• Mastery Learning
• Blended Learning

Key Sources

• What Works Clearinghouse, Institute of Education Sciences, U.S. Department of Education. (2019, November). Supporting Postsecondary Success intervention report: InsideTrack© Coaching. https://ies.ed.gov/ncee/wwc/InterventionReport/696 (reports Bettinger & Baker, 2014)
• What Works Clearinghouse, Institute of Education Sciences, U.S. Department of Education. (2020). Success Boston Coaching (WWC 2021-001). https://ies.ed.gov/ncee/wwc/InterventionReport/708
• Karp, M., Ackerson, S., Cheng, I., Cocatre-Zilgien, E., Costelloe, S., Freeman, B., Lemire, S., Linderman, D., McFarlane, B., Moulton, S., O'Shea, J., Porowski, A., & Richburg-Hayes, L. (2021). Effective advising for postsecondary students: A practice guide for educators (WWC 2022003). Washington, DC: National Center for Education Evaluation and Regional Assistance, Institute of Education Sciences, U.S. Department of Education. https://ies.ed.gov/ncee/wwc/PracticeGuide/28
• Meyer, Katharine, Lindsay C. Page, Catherine Mata, Eric N. Smith, B. Tyler Walsh, C. Lindsey Fifield, Michelle Tyson, Amy Eremionkhale, Michael Evans, Shelby Frost, and Eye Eoun Jung. (2026). Let's Chat: Leveraging Chatbot Outreach for Improved Course Performance (NBER Working Paper No. 35397). National Bureau of Economic Research. doi:10.3386/w35397. Also released as EdWorkingPaper 22-564, Annenberg Institute at Brown University: https://edworkingpapers.com/ai22-564
• Alina Martinez, Tamara Linkow, Hannah Miller, Amanda Parsad, Cristofer Price. (2021). Study of College Transition Messaging in GEAR UP: Impacts on Enrolling and Staying in College. Washington, DC: U.S. Department of Education, Institute of Education Sciences, National Center for Education Evaluation and Regional Assistance. https://www.mathematica.org/publications/study-of-college-transition-messaging-in-gear-up-impacts-on-enrolling-and-staying-in-college
• Wei Dai, Jionghao Lin, Flora Ji-Yoon Jin, Yi-Shan Tsai, Namrata Srivastava, Pierre Le Bodic, Dragan Gašević, and Guanliang Chen. (2025). Learning Analytics for Early Identification of At-Risk Students and Feedback Intervention. Journal of Learning Analytics, 12(3), 102–125. https://doi.org/10.18608/jla.2025.8735
• Weidlich, J., Gašević, D., Drachsler, H. (2022). Causal Inference and Bias in Learning Analytics: A Primer on Pitfalls Using Directed Acyclic Graphs. Journal of Learning Analytics, 9(3), 183–199. https://doi.org/10.18608/jla.2022.7577
• Raymond A. Opoku, Bo Pei and Wanli Xing. (2025). Unveiling Accuracy-Fairness Trade-Offs: Investigating Machine Learning Models in Student Performance Prediction. Journal of Learning Analytics 12(2). https://doi.org/10.18608/jla.2025.8543
• Yeomans, Michael and Reich, Justin. 2017. "Planning Prompts Increase and Forecast Course Completion in Massive Open Online Courses." Association for Computing Machinery (ACM). https://doi.org/10.1145/3027385.3027416
• Gollwitzer, P. M., & Sheeran, P. (2006). Implementation Intentions and Goal Achievement: A Meta‐analysis of Effects and Processes. Advances in Experimental Social Psychology, 38, 69–119. doi:10.1016/S0065-2601(06)38002-1
• Kizilcec, R. F., & Cohen, G. L. (2017). Eight-minute self-regulation intervention raises educational attainment at scale in individualist but not collectivist cultures. Proceedings of the National Academy of Sciences, 114(17), 4348–4353. doi:10.1073/pnas.1611898114
• Kizilcec, R. F., Pérez-Sanagustín, M., & Maldonado, J. J. (2017). Self-regulated learning strategies predict learner behavior and goal attainment in Massive Open Online Courses. Computers & Education, 104, 18–33. https://doi.org/10.1016/j.compedu.2016.10.001
• Guntur, M., & Purnomo, Y. W. (2024). A Meta-Analysis of Self-Regulated Learning Interventions Studies on Learning Outcomes in Online and Blended Environments. Online Learning Journal, 28(3). https://doi.org/10.24059/olj.v28i3.4025
• Guo, L. (2022). Using metacognitive prompts to enhance self‐regulated learning and learning outcomes: A meta‐analysis of experimental studies in computer‐based learning environments. Journal of Computer Assisted Learning, 38(3), 811–832. doi:10.1111/jcal.12650
• Wong, J., Baars, M., Davis, D., Van Der Zee, T., Houben, G.-J., & Paas, F. (2019). Supporting Self-Regulated Learning in Online Learning Environments and MOOCs: A Systematic Review. International Journal of Human–Computer Interaction, 35(4–5), 356–373. doi:10.1080/10447318.2018.1543084
• Er, Erkan; Akçapınar, Gökhan; Bayazıt, Alper; Noroozi, Omid; Banihashem, Seyyed Kazem. (2024). Assessing student perceptions and use of instructor versus AI-generated feedback. British Journal of Educational Technology. https://doi.org/10.1111/bjet.13558
• Wisniewski, B., Zierer, K., & Hattie, J. (2020). The power of feedback revisited: A meta-analysis of educational feedback research. Frontiers in Psychology, 10, 3087. doi:10.3389/fpsyg.2019.03087
• Means, B., and Neisler, J., with Langer Research Associates. (2020). Suddenly Online: A National Survey of Undergraduates During the COVID-19 Pandemic. San Mateo, CA: Digital Promise. https://digitalpromise.dspacedirect.org/items/32aee7c4-5363-4d50-bc9c-a338e7dfac7e doi:10.51388/20.500.12265/98
• Honig, C.A. & Salmon, D. (2021). Learner presence matters: A learner-centered exploration into the community of inquiry framework. Online Learning, 25(2), 95-119. https://doi.org/10.24059/olj.v25i2.2237
• Sadaf, A., Kim, S. Y., & Olesova, L. (2022). Relationship between metacognition and online Community of Inquiry in an online case-based course. Online Learning, 26(4), 79-93. https://doi.org/10.24059/olj.v26i4.3474
• Jia Frydenberg. (2007). Persistence in University Continuing Education Online Classes. International Review of Research in Open and Distance Learning, 8(3).
• Kraft, M.A., Blazar, D., Hogan, D. (2018). The effect of teacher coaching on instruction and achievement: A meta-analysis of the causal evidence. Review of Educational Research, 88(4), 547-588. https://doi.org/10.3102/0034654318759268
• Bettinger, E. P., Fox, L., Loeb, S., & Taylor, E. S. (2017). Virtual Classrooms: How Online College Courses Affect Student Success. American Economic Review, 107(9). doi:10.1257/aer.20151193
• Walton, G. M., Murphy, M. C., Logel, C., Yeager, D. S., et al. (2023). Where and with whom does a brief social-belonging intervention promote progress in college? Science, 380(6644), 499–505. doi:10.1126/science.ade4420
• Person, A. E., Thomas, J., Bruch, J., Johann, A., & Maestas, N. (2016). Outcomes of Competency-Based Education in Community Colleges: Summative Findings from the Evaluation of a TAACCCT Grant. Oakland, CA: Mathematica Policy Research. https://www.mathematica.org/publications/outcomes-of-competencybased-education-in-community-colleges-summative-findings-from-the-evaluation
• Swidan, O. (2026). Responding to AI-generated emotional alerts: teachers' intervention and students' engagement in the mathematics classroom. ZDM–Mathematics Education. https://doi.org/10.1007/s11858-026-01810-7
• Gao, Z., de Oliveira, G. S., Babalola, D., Lynch, C., & Heckman, S. (2024). Who should I help next? Simulation of office hours queue scheduling strategy in a CS2 course. Proceedings of the 17th International Conference on Educational Data Mining. https://doi.org/10.5281/zenodo.12729866

Page metadata

Field	Value
type	pattern
id	mentor-support-in-self-paced-online-learning
title	Mentor Support in Self-Paced Online Learning
description	A reusable policy for supporting learners in asynchronous, self-paced online courses and programmes through mentors, coaches or instructors who are available on demand and who reach out proactively: what to set up at the start, which progress signals trigger contact, what a contact does, how automated outreach and AI help divide the work with people, and how to check who the support actually reaches.
status	review
grain_size	course
canonical	True
generated	by claude/unspecified on 2026-10-19
trust	unverified — no one has confirmed this page's content

Frontmatter is the page's machine-readable record — see the Schema & Guide.