The matching algorithm is one of the most heavily marketed components of enterprise mentorship software. It is also only one part of what determines whether a mentoring programme works. Research supports the value of mentoring overall, but the evidence is much less definitive about how much an automated matching system adds to the quality or longevity of a particular relationship.
A multidisciplinary meta-analysis by Eby and colleagues, published in the Journal of Vocational Behavior in 2008, found that mentored individuals generally had more favourable outcomes than non-mentored individuals across several areas, including behavioural, attitudinal, career, motivational and relational outcomes. The study also found that the overall effects were modest, which is important context when evaluating claims about any single component of a mentoring programme.
That does not make matching technology unimportant. It means that buyers should distinguish between what the matching engine does and what the wider programme does. A matching system can organise information, apply criteria, reduce administrative work and help coordinators manage large participant pools. It cannot independently create trust, commitment or a productive mentoring relationship.
What Vendors Mean When They Say AI Matching
There is no single industry classification that divides AI matching into four universally accepted categories. For the purposes of evaluating enterprise mentorship software, it is more useful to treat the following four mechanisms as a practical framework for understanding what a vendor actually means when it uses the term “AI matching.”
Rules-based matching applies predetermined filters and constraints. For example, a programme might restrict a mentee to mentors within a particular function, seniority range or geographic region. No machine-learning model is necessarily involved.
Weighted scoring assigns different levels of importance to criteria such as function, skills, goals, interests, language or location, then produces a compatibility score or ranked list. This approach can be relatively transparent because administrators can inspect the criteria and weights.
Machine-learned matching uses historical programme data to identify patterns in previous matching outcomes. Depending on the system, the data might include accepted invitations, participant feedback or other programme signals. This approach generally requires enough relevant historical data to train or calibrate a model, which may be a limitation for a new programme.
Language-model matching can analyse free-text information such as biographies, goals, skills and participant statements to identify possible connections. Because language models can produce complex outputs from unstructured information, buyers should ask vendors how the recommendations are generated, what information is used and what level of administrative oversight is available.
The useful procurement question is therefore not simply whether a platform uses AI. It is which mechanism it uses, what information feeds it, how the output is evaluated, and whether a programme administrator can inspect or change the result.
Four Claims About AI Matching That Need More Care
The Myth That a Better Model Automatically Produces Better Relationships
Research from outside workplace mentoring illustrates why predicting relationship quality is difficult. In a well-known study of romantic attraction, Samantha Joel, Paul Eastwick and Eli Finkel used machine-learning methods to examine whether pre-interaction characteristics could predict attraction and pair-specific outcomes. The models were able to explain some variation in individual-level attraction, but predicting the unique dynamics between two particular people was considerably more difficult.
Professional mentoring is different from romantic attraction, so the findings cannot be transferred directly to enterprise mentoring. The broader methodological point is still relevant: information about two individuals before they meet may reveal useful characteristics, but it does not necessarily provide enough information to predict how those two people will interact.
Mentoring research reaches a similarly cautious conclusion about matching. A 2024 study of 565 science doctoral students found that several commonly assumed matching factors, including mentor rank, mentee capital and the way the relationship was initiated, were not associated with higher-quality relationships. The researchers also found no support for the idea that gender or racial similarity alone produced better mentoring quality. Shared attitudes, beliefs and values, along with culturally aware mentoring, showed stronger associations in that study.
The practical implication is not that an algorithm cannot improve matching. It is that buyers should be cautious about treating a compatibility score as a reliable prediction of relationship quality.
The Myth That Demographic Matching Is Always the Most Inclusive Choice
Demographic matching can matter, particularly when participants value having a mentor who shares an important aspect of their identity or understands a particular experience. At the same time, research does not support treating demographic similarity as a universal requirement for an effective mentoring relationship.
The National Academies’ guidance on mentorship notes that research on same-race and same-gender matching is mixed. It distinguishes surface-level similarities such as race, gender and age from deeper similarities such as goals, interests, values and attitudes. The guidance also notes evidence suggesting that identity similarity can be particularly valuable for some underrepresented participants, especially for psychosocial support.
More recent research also points in different directions depending on the population and context. A 2024 study of doctoral students found that gender and racial or ethnic similarity were not meaningfully associated with perceived mentoring quality, while shared attitudes, beliefs and values were associated with stronger outcomes. A 2025 study of 76 academic mentor-mentee pairs, meanwhile, found that one measure of demographic and professional similarity was positively related to mentee satisfaction, although it did not find the expected relationship between skills-based complementary fit and learning or research outcomes.
For an enterprise programme, demographic characteristics can therefore be treated as one potential matching consideration rather than a universal priority. The appropriate weighting may depend on the programme’s goals, participant preferences, representation objectives and the type of support being offered.
The Myth That Automation Removes Bias
Automation does not automatically remove bias from a matching process. It can move decisions about bias into the data, criteria, weights and constraints used by the system.
A rules-based system can reflect the assumptions built into its filters. A weighted model can amplify whichever criteria administrators choose to prioritise. A machine-learning system can also learn patterns from historical programme data that reflect previous participation or selection practices.
That makes governance relevant even when the matching system is marketed as objective. The NIST AI Risk Management Framework, released in 2023, is a voluntary framework designed to help organisations incorporate trustworthiness considerations into the design, development, use and evaluation of AI systems. It provides a useful reference point for thinking about documentation, risk management and ongoing evaluation, although it was not created specifically for mentoring software.
For a matching platform, that can translate into practical questions: What criteria are being used? Who controls the weights? Can administrators override a recommendation? Can participants request a different match? What information is retained? And how can the organisation review whether the system is producing patterns it did not intend?
The Myth That Matching Is a One-Time Event
Match acceptance is easy to measure, but it does not capture everything that happens after participants are paired.
Research on mentoring consistently treats relationship quality and the development of the relationship as important parts of mentoring effectiveness. Studies also indicate that relationships can vary over time rather than following a fixed path after the initial match.
That does not establish a universal point at which enterprise mentoring programmes fail, nor does it show that weeks three to eight are always the critical period. A more defensible approach is to monitor the relationship throughout the programme.
Useful signals can include whether meetings are taking place at the expected cadence, whether participants remain engaged, whether goals are being discussed and whether either participant requests a change. A straightforward re-matching process can give programme administrators a way to respond when a pairing is not working without assuming that every unsuccessful match represents a problem with the original algorithm.
What Actually Determines Match Quality
The evidence does not support reducing match quality to one or two variables. Several factors are worth considering when designing a matching process.
Input quality and goal specificity. A matching system can only use the information participants provide. Clear goals, relevant skills, interests and expectations give the system more useful information than generic statements. This does not guarantee a successful relationship, but it can make the matching criteria more meaningful.
Availability and practical constraints. Time zones, working patterns and expected meeting frequency can be useful constraints because a technically compatible pair still needs a practical way to meet. Research on formal mentoring has also identified factors such as time availability as relevant considerations in matching processes.
Mentor supply and workload. A matching algorithm cannot create additional mentors. If a programme has a shortage of mentors with particular skills or experiences, the available pool may limit the combinations that can realistically be offered. Programme administrators may therefore need to consider mentor capacity alongside matching criteria.
Participant agency. Giving participants some involvement in the selection process is another option. Research on mentoring matching has suggested that both mentor and mentee input can matter to the perceived quality of a match, although the appropriate balance between self-selection and administrator-led matching will depend on the programme.
These are programme-design considerations rather than guaranteed predictors of success. Their importance can vary by population, programme objective and mentoring format.
Configuring the Match Is the Work the Model Cannot Do
The practical buying question is often less about whether a platform has AI and more about how much control the programme administrator retains.
A matching engine can process participant information at a scale that would be difficult to manage manually. It can also apply the same criteria consistently across a large cohort. What it cannot know from structured profile data is everything happening outside the system, such as a mentor’s changing workload, an upcoming leave period or an interpersonal concern that a participant has not entered into their profile.
This makes human review an important design option. Administrators may want to define criteria, review proposed matches and allow participants to request changes rather than treating the algorithm’s output as final.
MentorCity provides one example of this model. Its published materials describe matching criteria that administrators can configure around factors such as seniority, role, interests and skills. The platform also describes options for self-matching, administrator-led matching and hybrid approaches, including a configuration in which mentees can rank several mentor preferences before a mentor accepts the relationship.
MentorCity also publishes programme results and satisfaction figures, but these should be understood as company-reported results rather than independent evidence that its matching approach will produce the same outcomes elsewhere. MentorCity reports a 95% satisfaction rate across mentor and mentee matches and cites an AO Foundation programme in which 81% of participants maintained their mentoring relationship through the end of the cohort, while 81% reported professional growth. The same case study reports that 85% rated the platform’s usability as good or very good.
Those figures are useful as examples of reported programme outcomes, but they do not establish that the matching technology caused those results. Programme design, participant selection, mentor recruitment, support and other factors may also have contributed.
Where AI Does Earn Its Place
Some of the more practical applications of AI sit around the mentoring relationship rather than making the final decision about who should be paired.
Session assistance can reduce administrative work by helping participants capture notes or summaries. Programme reporting can help administrators identify engagement patterns without waiting for a quarterly survey. These uses still require appropriate privacy, security and data-governance controls, particularly when mentoring conversations contain sensitive professional information.
AI-assisted communication can also help programme teams prepare reminders, check-ins or suggested discussion prompts. The value depends on how the organisation uses the output. A generated message can reduce drafting time, but it does not replace the programme manager’s responsibility to understand why a participant may have stopped engaging.
The strongest use cases are therefore not necessarily the ones that make the boldest prediction. They are often the ones that reduce administrative work while keeping the human relationship and programme decisions visible.
Questions Worth Putting to a Vendor
Which of the matching mechanisms described in this article does the platform actually use, and can you demonstrate the matching process rather than showing only the final pairing?
Can administrators set, change and weight matching criteria themselves? Can different cohorts use different criteria? Can a human review or override a proposed match before participants receive it?
What participant data feeds the matching process? How does the platform handle HRIS, LMS and SSO data when people change roles, leave the organisation or update their skills?
What reporting is available after matching? Can administrators see engagement, meeting frequency, participant feedback and requests for re-matching throughout the programme?
If the platform uses machine learning or language models, what historical data is used, how is it evaluated, and what controls exist for reviewing unexpected or potentially biased outcomes?
The Short Version
Treat AI matching as one part of programme design rather than assuming that a more sophisticated model will automatically produce better mentoring relationships.
Use the four mechanisms in this article as a practical way to interrogate vendor claims: rules-based matching, weighted scoring, machine-learned matching and language-model matching. They are useful categories for evaluating what a platform actually does, not a formal industry classification.
Consider goals, skills, interests, values, availability and demographic factors according to the objectives and participants of the programme. Research does not support one universal matching hierarchy, and different forms of similarity may matter in different settings.
Give administrators and participants an appropriate degree of control over the final pairing, particularly where the available data cannot capture important contextual information.
Monitor the relationship after the initial match rather than treating acceptance as the end of the process. Meeting frequency, participant feedback, goal progress and re-matching requests can provide a broader view of programme health.
For anyone running a procurement process, the practical implication is straightforward: spend less time asking whether a vendor has “AI matching” and more time finding out what the system actually does, what data it uses and how much control your programme retains.
Frequently Asked Questions
Is AI Matching Better Than Letting People Choose Their Own Mentor?
There is no universal answer. The appropriate approach depends on the size of the participant pool, the programme objectives, the available mentor supply and how much choice participants are expected to have. A small programme may be manageable through a directory and participant-led selection, while a larger programme may benefit from automated shortlisting or administrator-assisted matching. A hybrid approach is another option, with software narrowing the available choices while participants or programme administrators make the final decision.
How Long Does It Take to Implement Matching in an Enterprise Mentoring Program?
Implementation time varies considerably depending on the number of participants, data quality, integrations, programme design and approval process. A simple pilot with clean participant data may require less preparation than a large enterprise programme connected to HRIS, LMS and SSO systems. Rather than relying on a universal implementation benchmark, ask the vendor to provide a timeline based on your participant volume, integrations and matching workflow.
What Data Does an AI Matching Tool Need From Our HRIS or LMS?
Potential inputs include role, function, location, tenure, skills, learning history and other organisational information. However, HRIS and LMS data may not capture the participant’s actual mentoring goals, preferred areas of development, interests or availability. Those details generally need to be collected through the mentoring platform itself.
It is also worth distinguishing between completed training and demonstrated capability. A course completion record may indicate exposure to a subject, but it does not necessarily establish that a participant has mastered the skill.
How Should We Measure Whether Matching Is Working?
Use several measures rather than one matching metric. Relationship continuation can be tracked at different points during the programme, but the appropriate checkpoints depend on programme length. Meeting cadence can show whether the relationship is active, while participant feedback can indicate whether the mentoring relationship is useful to both sides. Goal progress can provide another outcome measure when the programme has clearly defined objectives.
Match acceptance can still be useful as an operational metric, but it should not be treated as a complete measure of match quality. A participant accepting a pairing tells you that the match was accepted, not necessarily that the relationship will remain productive.
Feature image by LinkedIn Sales Navigator from Pexels
