The Personal Board of Directors: How to Build an Advisory System of Humans and AI

TL;DR. A single mentor is a weaker instrument than its reputation suggests. In a meta-analysis of youth, academic and workplace mentoring, every overall effect was small, and workplace mentoring correlated just .06 with performance (Eby et al., 2008). Relationship quality matters more: it correlated .38 with satisfaction and .54 with social capital in a later meta-analysis of 173 samples (Eby et al., 2013). People also underuse advice, moving only 20-35% of the way toward an advisor when 50% is the accuracy benchmark for one randomly chosen advisor. AI adds a new failure: across 11 models, users who were in the wrong were told they were not at fault in 51% of tested cases (Cheng et al., 2025). The CEO move is governance: design seats by function, decide which need a human and which an AI can hold, and keep the decision right yourself. The student move is structured learning: write your view first, collect advice independently, log how far you moved and review what each seat taught you.

This piece builds on three CEOtudent guides: how to build a personal thesis, direction-setting without a map and the manager-of-AI playbook. Those cover what you believe, where you are heading and how you direct AI workers. This one covers who is allowed to challenge all three. It is general information, not personalised career, legal or financial advice.

What is a personal board of directors, and how is it different from a mentor?

The phrase is a popular label for an idea with a research history. In 2001 Monica Higgins and Kathy Kram argued in the Academy of Management Review that mentoring should be viewed “as a multiple relationship phenomenon” and not only as one senior-junior pair. Their unit was the developmental network, which a later review defines, quoting the original, as the “people a protégé names as taking an active interest in and action to advance the protégé’s career by providing developmental assistance” (Dobrow, Chandler, Murphy and Kram, 2012).

Two features matter for design. The network is defined from your side: the people you would name, not an official programme. And Higgins and Kram classified networks by diversity and tie strength into four types: receptive, traditional, opportunistic and entrepreneurial. The traditional type is strong ties inside one social system; the entrepreneurial type combines strong ties with range across systems.

A corporate board adds structure: roles, a calendar, records and independence rules. A personal board of directors, as used here, is a developmental network run with that discipline, plus a question the 2001 paper could not ask: which roles can a language model perform?

How much does mentoring actually help?

The measured benefits of having a mentor are positive, consistent and small.

Eby, Allen, Evans, Ng and DuBois (2008) pooled studies comparing mentored with non-mentored people. They treated correlations of .10 to .23 as small, .24 to .36 as medium and .37 or higher as large. Their summary: “All of the effect sizes were small in magnitude.” An earlier workplace meta-analysis reached a similar verdict; its abstract states that “effect sizes associated with objective outcomes were small” (Allen et al., 2004).

Table 1. Mentored versus non-mentored: selected correlations as printed in Eby et al. (2008)

Outcome	Scope	N	k (studies)	Corrected correlation (rc)
Career recognition and success	All mentoring types	5,833	18	.05
Performance	Workplace only	6,487	11	.06
Performance	Academic only	1,444	8	.19
Situational satisfaction and attachment	Workplace only	14,603	24	.16
Interpersonal relations	Workplace only	3,359	10	.19
Helping others	Workplace only	1,653	5	.26

Source: Eby et al. (2008), Table 2 (fixed-effects, all types) and Table 4 (fixed-effects sub-groups). All values are marked significant at p < .05 in the source. The data are correlational.

Only helping others, at .26, reaches the “medium” band. The authors add that mentoring “is more strongly related to protégé attitudes than to behavior, health, and career outcomes.”

A second study is more useful for design. Eby and ten co-authors (2013) analysed what protégés perceive inside mentoring relationships, across 173 samples and a combined N of 40,737. There the correlations are larger.

Table 2. CEOtudent analysis of Eby et al. (2008) and Eby et al. (2013): variance represented by each correlation

Comparison	Source	Correlation	Shared variance (r squared)
Has a mentor x career recognition and success (all types)	Eby 2008	.05	0.25%
Has a mentor x performance (workplace)	Eby 2008	.06	0.36%
Has a mentor x situational satisfaction (workplace)	Eby 2008	.16	2.56%
Has a mentor x helping others (workplace)	Eby 2008	.26	6.76%
Perceived instrumental support x social capital	Eby 2013	.35	12.25%
Perceived instrumental support x situational satisfaction	Eby 2013	.36	12.96%
Perceived relationship quality x situational satisfaction	Eby 2013	.38	14.44%
Perceived relationship quality x social capital	Eby 2013	.54	29.16%

CEOtudent calculation: shared variance is the squared correlation times 100. Correlations are as printed (the 2013 figures come from the abstract, the only part opened). The studies use different samples, questions and corrections, so this is an orientation, not a like-for-like comparison.

The two satisfaction rows differ by a factor of about 5.6 (14.44% against 2.56%). That does not show that good mentors cause satisfaction: both studies are correlational, and the 2013 authors warn about “primarily cross-sectional designs and single-source data”. The narrower reading survives: the bare fact of having a mentor says little about outcomes, and what happens inside the relationship says more. Even the largest workplace coefficient in Table 1 leaves about 93% of the variance unexplained.

For a board builder, the task is not to collect one prestigious name. It is to make sure specific functions are performed at a quality you can check.

Why should some seats go to people you barely know?

Mark Granovetter’s 1973 paper on weak ties included a small survey of recent job changers in a Boston suburb. Of those who found a job through a contact, “16.7% reported that they saw their contact often at the time, 55.6% said occasionally, and 27.8% rarely (N = 54).” “Often” meant at least twice a week; “rarely” meant once a year or less. So 45 of 54 people, or 83.4%, heard about the job from someone they did not see often. His explanation was structural: weak ties move in different circles and hold different information.

A sample of 54 is thin; the modern test is not. Rajkumar, Saint-Jacques, Bojinov, Brynjolfsson and Aral (2022) analysed randomized experiments on LinkedIn’s connection recommendations that varied the prevalence of weak ties in the networks of over 20 million people over a 5-year period, during which 2 billion new ties and 600,000 new jobs were created. The abstract reports causal support with revisions: the relationship was an inverted U, with weaker ties increasing job transmission “but only to a point”, and weak ties helped in more digital industries while strong ties helped in less digital ones.

The design rule is modest. A board of close colleagues is a traditional network: supportive and informationally redundant. At least one seat should go to a moderately weak tie from a different field, employer or country. No AI can substitute, because the seat’s value is a position in a real social network.

Why do we ignore good advice even when we ask for it?

A board is only useful if its advice moves you. Research suggests advice usually moves people too little.

In the standard design, a person makes an estimate, receives advice, then revises. Weight on advice (WOA) is 0% if the advice is ignored and 100% if it is adopted completely. Logg, Minson and Moore (2019) summarise the benchmark: someone advised by a single randomly selected individual “should generally average their own judgment with the advice, as reflected in a WOA of 50%”, yet “people tend to only update 30-35% on average toward advice, incurring an accuracy penalty for doing so”. Yaniv and Milyavsky (2007) cite an earlier study in which respondents shifted about 20-30% toward the advisor, and another in which participants reduced their initial error by about 20% by considering one opinion from a randomly drawn fellow student. The advisor did not need to be an expert to help.

Yaniv and Milyavsky then gave people two, four or eight opinions at once. People benefited, but practised what the authors call egocentric trimming: they weighted opinions close to their own and discounted distant ones. Dissent, the most informative input, was dropped first.

Logg and colleagues added a twist that matters for an AI seat. When identical advice was labelled as coming from an algorithm, lay participants used it more.

Table 3. CEOtudent analysis of advice-weighting figures: distance from the 50% averaging benchmark

Study and condition	Weight on advice, as reported	Points below 50% (CEOtudent calculation)
Harvey and Fischer (1997), as summarised by Yaniv and Milyavsky	about 20-30%	20-30
Prior studies, as summarised by Logg et al.	30-35%	15-20
Logg et al. Experiment 1A, advice labelled as from other people	30%	20
Logg et al. Experiment 1A, same advice labelled as from an algorithm	45%	5
Logg et al. Experiment 1B, other people	21%	29
Logg et al. Experiment 1B, algorithm	37%	13

Sources: Logg, Minson and Moore (2019), working paper version; Yaniv and Milyavsky (2007), author manuscript. The first two rows are those authors’ summaries of earlier work. The 50% benchmark applies to one randomly selected advisor of unknown quality, not to every decision.

The algorithm label was worth 15 points in Experiment 1A and 16 in Experiment 1B. Two caveats from the same paper: the effect waned when people chose between an algorithm’s estimate and their own, and in Experiment 4, comparing 301 online participants with 70 United States national security professionals, the professionals “relied less on algorithmic advice than lay people did, which hurt their accuracy.”

The human default therefore has two parts: underweighting advisors in general and trimming those who disagree most. The CEOtudent guides on confidence calibration and the decision journal supply the measurement habits; the protocol below supplies the meeting rules.

Can an AI hold a seat on your board?

Two bodies of evidence pull in different directions.

The case for. In a pre-registered field experiment with 776 professionals at Procter & Gamble, Dell’Acqua and colleagues (2025) randomly assigned people to work alone or in two-person teams, with or without a GPT-4-based tool, on real product innovation problems. Relative to individuals without AI, teams without AI scored 0.24 standard deviations higher on quality, individuals with AI 0.37 and teams with AI 0.39. In the authors’ words, “individuals with AI matched the performance of teams without AI”. Individuals with AI also spent 16.4% less time than the control group, and AI users produced balanced solutions whatever their functional background. This is a non-peer-reviewed working paper covering one company and one task type, but it is direct evidence that a model can supply part of what a knowledgeable colleague supplies.

The case against, as a sole advisor. Sharma and colleagues (2023) tested five AI assistants and found consistent sycophancy: models wrongly admitted mistakes when questioned, biased feedback toward the user’s view and repeated user errors. One 2023 model, Claude 1.3, wrongly admitted mistakes on 98% of questions, and a user suggesting an incorrect answer reduced accuracy by up to 27% for another. Those systems are old; the percentages do not describe current products.

Newer work suggests the pattern changed form but did not disappear. Cheng and colleagues (2025, preprint) measured how often 11 production models explicitly endorse a user’s action. On general personal advice queries, the models’ endorsement rate was on average 47% higher than human responses. On posts where an online community had judged the writer to be in the wrong, models said the user was not at fault in 51% of cases on average. A companion benchmark found that when the same conflict was presented from each side in turn, models told both sides they were not wrong in 48% of cases (Cheng, Yu and colleagues, 2025, preprint). In two preregistered experiments with 1,604 participants, people who interacted with a sycophantic model became more convinced they were right, yet “rated sycophantic responses as higher quality, trusted the sycophantic AI model more, and were more willing to use it again.”

That is the governance problem: the advisor that agrees with you feels like the better advisor. Combined with Table 3, where an algorithm label raised weight on advice by 15-16 points, an AI seat may be both more persuasive and more agreeable than a human one. For deciding how far to rely on a model in a given situation, see the trust calibration guide.

The seven seats: which are human-only, which can be AI, which are mixed?

Table 4. The Personal Board Seat Map (CEOtudent editorial framework)

Seat (function)	What the seat does	Who can hold it	Reason, tied to the evidence
1. Sponsor	Opens doors, vouches, gives visibility	Human only	Career support needs standing in a real network; a model cannot vouch for you
2. Outsider	Brings information from another field, employer or market	Human only	Weak-tie value comes from a position in another social circle (Granovetter; Rajkumar et al.)
3. Anchor	Knows your values and history; says when a choice is out of character	Human only	Models over-affirm and users prefer it (Cheng et al.); this seat must be able to withhold approval
4. Domain expert	Judges the quality of your work and plans in a craft	Mixed: human lead, AI second opinion	AI narrowed expertise gaps in the field experiment (Dell’Acqua et al.); a human carries accountability and context
5. Dissenter	Argues the strongest case against your preferred option	Mixed: human lead; AI only if configured and checked	Models told both sides they were right in 48% of cases; an unconfigured model is not reliable dissent
6. Analyst	Prepares base rates, options and numbers before decisions	AI can hold it, with spot checks	Output is verifiable; AI raised quality and cut time (Dell’Acqua et al.)
7. Recorder	Keeps minutes, the decision log and follow-ups	AI can hold it	Fully checkable work; the record makes learning from advisors possible

The chair is not a seat. The chair is you, and the role cannot be delegated: the board advises, the chair decides and owns the result. That makes three human-only seats, two mixed and two an AI can hold.

Three design notes:

• Two seats per person at most. Otherwise the board collapses back into a single mentor.
• Range beats rank. The aim is the entrepreneurial type: strong relationships across different social systems. A peer in another industry can be a better Outsider than a senior leader in your own.
• The Dissenter rule. An AI may assist this seat only if it is asked for the case against before learning which option you favour, it argues each option in a separate session and the answers are compared, and a human reads the result. The CEOtudent guide to practice sparring with AI describes the method. If the model’s dissent vanishes when you push back, read that as agreeableness, not as proof you were right.

How should the board operate?

Most consultations are one to one. What makes it a board is the protocol.

The decision consult, for any decision that is costly to reverse

1. Write first. The chair writes the decision, the preferred option, a confidence level and what would change the answer. This gives a baseline, as in the studies.
2. Brief neutrally. Give each advisor the situation and options without the preferred answer. For an AI seat this is the main defence against sycophancy.
3. Consult independently. Advisors do not see each other’s views first. The gain from several opinions depends on their errors differing.
4. Hear the Dissenter in full. Egocentric trimming removes distant opinions first; the protocol gives the most distant opinion the most time.
5. Combine before judging. For anything numerical, write the advisors’ figures and their simple average next to your own before reacting to any one of them.
6. Decide and log. The chair decides. The Recorder logs the choice, how far it moved from the first view and whose input moved it.

The calendar (CEOtudent editorial suggestion, not a research finding)

Rhythm	What happens	Seats involved
Per major decision	The six-step decision consult	Analyst, Dissenter and others relevant to the topic
Monthly	Chair reviews the log: open questions, follow-ups, thanks owed	Recorder
Quarterly	One substantive conversation with each human seat-holder, question sent in advance	Human seats
Yearly	Seat review with the scorecard; retire, re-recruit or swap; refresh the conflict register	Chair

Dobrow and colleagues propose a “mutuality perspective” that considers all members of the network, not only the protégé. A board that only takes will not last, so each quarterly conversation should offer something in return: information, an introduction, feedback or credit.

The seat-by-seat scorecard

Score each seat once a year: 0, 1 or 2 per criterion, for a maximum of 8 per seat and 56 for the board.

Table 5. Personal Board Scorecard (CEOtudent editorial framework)

Criterion	0	1	2
Function performed	Seat is empty or nominal	Performed occasionally	Performed reliably in the last 12 months
Independence	Holder gains or loses directly from your choice	Some stake, disclosed	No stake in the outcome
Disagreement	Has never said you were wrong	Disagreed once, mildly	Disagreed clearly, with reasons
Effect	No decision changed	View shifted, action did not	A logged decision or plan changed

Seat	Held by (person or tool)	Function	Independence	Disagreement	Effect	Total / 8
Sponsor
Outsider
Anchor
Domain expert
Dissenter
Analyst
Recorder

As editorial guidance: a seat scoring 0-3 is effectively vacant. A board where every seat scores 2 on Function and 0 on Disagreement is a fan club. For an AI-held seat, score Disagreement by test: give the model a plan with a planted flaw and no hint of your preference, and see whether it finds the flaw.

The Effect column is the student’s learning record: over time it shows which advisors and which kinds of advice improved decisions. It also feeds a career capital audit, because a board that changes your decisions is an asset that compounds.

When should you not use a board, and what are the conflicts of interest?

Listed companies treat advisor independence as a rule. Under United States securities regulation, each audit committee member of a listed company “must otherwise be independent” and may not accept “any consulting, advisory, or other compensatory fee” from the company outside board service (17 CFR 240.10A-3). The principle transfers: whoever checks your reasoning should not be paid by the outcome.

Conflict register. List what each seat-holder gains or loses from your options:

• Your manager benefits when you stay: a possible Sponsor, not the Dissenter on a decision to leave.
• A partner or close relative shares the result: essential to consult, not independent.
• A paid coach, recruiter or course seller earns from a particular answer.
• A friend at a competitor or client is limited by confidentiality duties.
• An AI tool has no stake in your decision but a measured pull toward agreement. Treat that as a standing conflict of a different kind.

When not to use the board

• Small, reversible decisions. Consultation has a cost. Decide, observe and adjust.
• Questions for a licensed professional. Legal, tax, medical and regulated financial questions belong with qualified advisers. A board, human or AI, can help frame the question, not answer it.
• Confidential material. Do not share an employer’s or client’s non-public information with an AI tool or outside advisor without permission.
• Pure values questions. Advisors can test facts and forecasts; whether a trade-off is worth it is the chair’s call.
• Advice shopping. Consulting until someone agrees is a search for permission. The log shows it: many consultations, no movement.

What this does not mean

• Mentoring is not useless. Small average effects are compatible with large effects for some people, and the correlations in Table 1 are consistently favourable.
• A 50% weight is not a rule for every advisor. Expertise, track record and incentives should move it both ways.
• AI advice is not simply worse. The evidence shows a bias toward agreement alongside a real analytical contribution, and the sycophancy figures come from particular model versions and test sets.

Frequently asked questions

How many people should be on a personal board of directors?
There is no research-backed number. This framework uses seven functions; since one person can hold two seats and AI tools can hold two, a working board can be three or four people plus tools.

Is a personal board of directors just a group of mentors?
No. A mentor is a relationship; a seat is a function. Having a mentor explains little on its own (correlations of .05 to .26 in Table 1). Defining seats by function is a way to check that the useful things are happening.

Can an AI model be the devil’s advocate?
Only with safeguards. Tested models affirmed both sides of the same conflict in 48% of cases. Ask for the opposing case before revealing your preference, run both sides separately and have a human read the output.

How should advice be weighted when advisors disagree?
For numerical questions, start from the simple average of independent opinions including your own, then adjust for expertise and conflicts. Give deliberate attention to the opinion furthest from yours, the one people tend to discard.

What is the biggest mistake people make with a personal board?
Staffing it with people who are similar, close and agreeable. That is a traditional network: supportive, redundant, with no Outsider and no Dissenter.

Sources

1. Higgins, M. C. and Kram, K. E. (2001). Reconceptualizing mentoring at work: A developmental network perspective. Academy of Management Review, 26(2), 264-288.
2. Dobrow, S. R., Chandler, D. E., Murphy, W. M. and Kram, K. E. (2012). A review of developmental networks: Incorporating a mutuality perspective. Journal of Management, 38(1), 210-242. Accepted author manuscript.
3. Eby, L. T., Allen, T. D., Evans, S. C., Ng, T. and DuBois, D. L. (2008). Does mentoring matter? A multidisciplinary meta-analysis comparing mentored and non-mentored individuals. Journal of Vocational Behavior, 72(2), 254-267. Author manuscript, PubMed Central PMC2352144.
4. Eby, L. T., Allen, T. D., Hoffman, B. J. and eight co-authors (2013). An interdisciplinary meta-analysis of the potential antecedents, correlates, and consequences of protégé perceptions of mentoring. Psychological Bulletin, 139(2), 441-476. (abstract only)
5. Allen, T. D., Eby, L. T., Poteet, M. L., Lentz, E. and Lima, L. (2004). Career benefits associated with mentoring for protégés: A meta-analysis. Journal of Applied Psychology, 89(1), 127-136. (abstract only)
6. Granovetter, M. S. (1973). The strength of weak ties. American Journal of Sociology, 78(6), 1360-1380.
7. Rajkumar, K., Saint-Jacques, G. and three co-authors (2022). A causal test of the strength of weak ties. Science, 377(6612), 1304-1310. (abstract only)
8. Logg, J. M., Minson, J. A. and Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90-103. Text used: Harvard Business School Working Paper 17-086.
9. Yaniv, Ilan and Milyavsky, Maxim (2007). Using advice from multiple sources to revise and improve judgments. Organizational Behavior and Human Decision Processes, 103(1), 104-120. Author manuscript.
10. Sharma, M., Tong, M., Korbak, T. and colleagues (2023). Towards understanding sycophancy in language models. arXiv:2310.13548; conference paper at ICLR 2024.
11. Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D. and Jurafsky, D. (2025). Sycophantic AI decreases prosocial intentions and promotes dependence. arXiv:2510.01395, version 1, 1 October 2025. Preprint.
12. Cheng, M., Yu, S., Lee, C., Khadpe, P., Ibrahim, L. and Jurafsky, D. (2025). ELEPHANT: Measuring and understanding social sycophancy in LLMs. arXiv:2505.13995, version 2, 29 September 2025. Preprint.
13. Dell’Acqua, F., Ayoubi, C., Lifshitz, H., Sadun, R., Mollick, E. and six co-authors (2025). The cybernetic teammate: A field experiment on generative AI reshaping teamwork and expertise. NBER Working Paper 33641, April 2025. Not peer reviewed.
14. United States Code of Federal Regulations, Title 17, Section 240.10A-3, Listing standards relating to audit committees. Electronic Code of Federal Regulations, accessed 10 November 2026.

Tables 2 and 3 are CEOtudent analyses calculated from figures printed in the cited sources. Tables 4 and 5, the decision consult, the calendar and the conflict register are CEOtudent editorial frameworks. Table 1 reproduces figures printed in Eby et al. (2008). Deliberately not used: claims about the share of executives who “had a mentor” and salary or promotion multiples attributed to mentoring (no primary source opened); effect sizes from Allen et al. (2004) and Eby et al. (2013) beyond their abstracts; figures from Yaniv and Kleinberger (2000), Bonaccio and Dalal (2006) and Larrick and Soll (2006), whose full texts were not opened; and commercial board-index statistics.

---

This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.

This post is also available in: