Your Company Kept Your Rules and Threw Away Your Reasons

The research on that chain is old and unambiguous. Gabriel Szulanski studied 122 transfers of best practice inside eight companies, 271 observations, published in the Strategic Management Journal in 1996. The assumption at the time was motivational: turf, incentives, not-invented-here. The data said otherwise. The dominant barriers were knowledge-related: the recipient's lack of absorptive capacity, an arduous relationship between source and recipient, and causal ambiguity, the plain fact that nobody can fully reconstruct why the thing worked. Your people are applying the part that reached them, and the chain behind that part reached nobody.

Chris Argyris and Donald Schon named the gap in Theory in Practice in 1974. Your espoused theory is the account of your own actions you would give if asked, and that account is what ends up in the values document. Your theory-in-use is what actually governed the decision, operating tacitly, standing to action as grammar-in-use stands to speech. Founders write down the espoused version because that version is the only one easy to articulate. Michael Polanyi set the ceiling in The Tacit Dimension in 1966: we can know more than we can tell.

Frederic Bartlett's serial reproduction experiments, published in Remembering in 1932, showed that material passing from person to person loses whatever does not fit the listener's schema, or has it converted into something more familiar. Canoe became boat. Three or four handovers down, "we do not discount, because a discount concedes the positioning claim we sell on" becomes "we do not discount", and the shortened rule gets applied to a partnership where you would have taken the deal in a heartbeat.

Strip the reason off a rule and what remains is a superstition, defended more fiercely than you ever defended the original, because its defenders can only argue from you.

A company inherits its founder's conclusions automatically, because conclusions are the part somebody wrote down. A company inherits the founder's judgement only if somebody sat the founder down and asked about particular decisions, in enough detail that the reasons came out attached to the story.

Organisations keep the shape of their founder's decisions long after losing the judgement that would have updated those decisions. Arthur Stinchcombe described the mechanism in 1965 in "Social Structure and Organizations": conditions present at founding shape a firm's structure, and those structures persist long after the environment that produced them has changed. Christopher Marquis and Andras Tilcsik generalised the mechanism as imprinting in their 2013 review in the Academy of Management Annals, in three parts. A brief sensitive period. Absorption of the surrounding environment during that window. Persistence of the imprint despite everything that changes afterwards.

Structures persist: the rule, the org chart, the sentence on the wall. The live judgement that would have said "that was true when we sold to procurement teams and it is not true now" persists nowhere.

And plan on not being in the room. Noam Wasserman tracked succession across 202 internet firms for Organization Science in 2003 and found what he called the paradox of entrepreneurial success: hitting the milestones, finishing the product and closing each round of outside money raised the chance that the founder was replaced as chief executive. In the Harvard Business Review in February 2008, on 212 American start-ups founded in the late 1990s and early 2000s, he put the outcome flatly: "successful CEO-cum-founders are a very rare breed". Half the founders were no longer chief executive by year three. In year four, 40 per cent were still in the seat. Fewer than a quarter led their companies' initial public offerings. None of that requires a disaster: a chair role, a sabbatical, a slow step back across three years at the end of which the reasoning has quietly left with you.

Ray Dalio's TED talk in April 2017 described algorithmic decision-making and an idea meritocracy at Bridgewater, built on a written principles corpus. What has changed since is that reasoning no longer has to be compressed into rules first. People can ask instead.

A principle is the place where the reasoning has already been thrown away, so principles will never give the reasoning back. Nonaka and Takeuchi's account in The Knowledge-Creating Company (1995) calls the tacit-to-explicit step externalisation, and puts that step in dialogue rather than documentation. You cannot write down your judgement. You can be asked about a Tuesday in 2019 and produce the judgement live, in the answering.

So the material that carries a method is narrow. The decision you reversed, and not the fact of the reversal but the new information and why it outweighed the old. The trade-off you refused, and what you paid instead, because every principle is purchased with something. The hire who looked right and was not, and the signal you now know you missed. The rule you would break, and the conditions that would justify breaking the rule.

Three of those four are failures, and not out of modesty. Failures carry more reasoning per minute than successes do, because a success can be explained by the rule and a failure cannot.

Do This Today, Free, Before You Buy Anything

Take your operating principles. For each line, record two answers on your phone, no more than two minutes each. First: what was the specific decision that made me write this, with a name, a number and a date. Second: what is the case where I would do the opposite, and how would I know I was in that case? Ten lines is about forty minutes of audio. Transcribe it, hand the transcript to your leadership team alongside the original document, and a conclusions document has become a reasoning document. Some lines turn out to carry exceptions large enough that the principle misled on its own: those are the lines your managers have been applying wrongly for years, in good faith, while quoting you accurately. If a blank recorder defeats you, the question sequences in how to interview yourself properly will carry you through the first ten minutes.

Start building your legacy Free build, 50 memories, no card required.

If the company owns the Persona, the Persona is an asset, and assets travel through events the founder does not control. So the founder owns the Persona personally and grants the company a licence for named uses. The failure mode is not exotic: an IP assignment clause in a founder's employment agreement usually catches whatever the founder creates in the course of employment, so build on company time, on a company account, on the company card, and an acquirer's counsel will treat the whole thing as company property. Then run the events. Under company ownership an acquisition carries the Persona over with the other assets, an insolvency puts the Persona in the estate for creditors, and a board or co-founder dispute hands the founder's voice to whoever controls the company. Under a licence the acquirer buys the company and not the person, what sits in the estate is a licence ending on the terms you wrote, and only the founder ever adds to the record.

Congress has arrived at the same risk analysis. The NO FAKES Act (S.4591, 119th Congress) cleared the Senate Judiciary Committee on a unanimous voice vote on 18 June 2026 and was reported and calendared on 24 June 2026. As at 30 July 2026 the bill has not passed the Senate and is not law. The drafters' structure is the part worth copying. The bill would make the right in your own voice and likeness not assignable during your lifetime, barring transfer "including through involuntary transfer, such as by means of bankruptcy, levy, lien enforcement, court order, or other legal process". It would permit licensing only, in writing, signed by you, containing "a reasonably specific description of the intended uses", for a maximum of ten years.

You can adopt that shape today by contract. Personal ownership. A licence, not an assignment. Named uses. A term. Termination on departure and on change of control. No sublicensing. Sole right to add to the record retained by you, permanently. Settle it before anybody has a reason to argue, and open the account in your own name from the start.

California's AB 2602, signed 17 September 2024 and effective 1 January 2025, added section 927 to the Labor Code. A digital replica provision is unenforceable there only when three conditions all hold: the replica stands in for work the individual would otherwise have performed, the provision lacks a reasonably specific description of the intended uses, and the individual was not represented by counsel or by a labour union. Representation is what makes a broad clause stick, so a lawyer at the table does not protect you. What protects you is specificity in the grant.

These statutes govern the use of a voice and a likeness, not the ownership of an AI version trained from them. Rights in the underlying recordings are a separate question, set out in who owns your voice clone. The law here is unsettled, as the note at the foot of this page sets out, and none of this is legal advice.

Tell people plainly, in the interface and at the point of use. Article 50 of the EU AI Act (Regulation (EU) 2024/1689) becomes applicable on 2 August 2026. Article 50(1) requires that providers "ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious". Your employees are natural persons, and an internal founder Persona is squarely such a system. The duty, though, falls on the provider who builds the system rather than the company deploying one internally, so interface disclosure is your vendor's job.

Article 50(4) is the paragraph that catches companies. Deployers of a system that generates or manipulates image, audio or video content constituting a deep fake "shall disclose that the content has been artificially generated or manipulated". Circulate a synthesised founder voice note to staff and your company is the deployer, so the duty is yours. Penalties under Article 99(4) run to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher. Regulation (EU) 2026/1744, the Digital Omnibus on AI, entered into force on 27 July 2026 and gave providers who had already placed generative systems on the market before 2 August 2026 a four-month transitional period on the Article 50(2) marking obligation, running to 2 December 2026, which moves a compliance date without touching the duty underneath. Full detail sits in what Article 50 actually requires from 2 August 2026.

Elsewhere no law may require disclosure at all. Disclose anyway, and say three things in the announcement and again inside the tool. This Persona is a record of how the founder reasoned, built from interviews with the founder. This Persona is not an approval, and citing the Persona in a meeting is not an approval either. Nobody but the founder can change what the Persona says, including the executive team, so the Persona will sometimes contradict current strategy, and when that happens current strategy wins. Leave the third line out and people will be using your founder Persona as a weapon inside a fortnight.

The limit first. While you are alive and the Executor Lock has not been triggered, you are the only person who can talk to your Persona. You can nominate the people who will be able to, but their access opens through the release process, not on a date you pick. So building now buys the record and the transcripts, not a Persona your leadership team questions on Monday.

Building starts with fifty memories, free, no card, and the free build never expires. A guide called Idy runs the interview, follows what you say instead of a script, and returns to threads you left open earlier. The productive founder sessions are decision post-mortems. Answers are stored as memories, and you can read, edit and delete any memory right up to the day the Lock activates. What comes back is only the reasoning you put into words: answer "we hire for judgement" and the memory says "we hire for judgement", while naming the candidate, the week and the call you got wrong brings the reasoning out attached to the story.

Your voice is cloned from the recordings you make in the sessions, from your audio alone, so your Persona answers in your voice rather than a synthetic stand-in. A founder's voice is a public asset others have reason to imitate, and the defences are in how to protect your voice from AI cloning.

The account is yours personally, never the company's, which is the mechanism behind the ownership structure above, and you are the only person who can add a memory or change one. How a model answering only from one person's material behaves is set out in what an AI trained only on you can and cannot do.

You nominate, by name, who will be able to speak with your Persona. They are called Trusted Contacts: one on the free build, two on Legacy, five on Eternal, ten as the hard ceiling. The cost of that design, said out loud: no single sign-on, no seat management, no admin console, no route to hand a department a login. Access runs through the Executor Lock™ release process, where a nominated person files the request, uploads evidence, and a seven day cooling-off window runs before anything opens. You are notified when a request is filed and can revoke the request inside the window. When the Lock activates, your Persona is snapshotted as built that day: personality and foundational memories fixed, and how you reason cannot be altered by anybody, us included. Conversations after the Lock still leave short-term notes, so that your Persona keeps living, but the locked core never moves.

Your Persona answers from things you actually said, and is instructed never to manufacture a specific fact your memories do not hold, because a Persona inventing a plausible founder-sounding position would be worse than a values document. Silence on a subject is no proof of coverage, and the only way to give your Persona a view is to record one. A record you stop adding to freezes in the year you stopped.

The exception pass costs nothing and tells you whether there is a method here worth keeping. If there is, the next step is the same size: fifty memories, free, no card, no expiry, in your own name rather than the company's. Record the reversal, the refused trade-off and the hire you misread, then decide whether you want the licence drawn up at all. The paid plans are on the plans page.

Your company will not miss your decisions. They are on the wall, in the deck, in the mouths of people who joined last spring. What your company will miss is your reasons, and reasons only survive if somebody sits you down and asks while you are still here to answer.

Start building your legacy Free build, 50 memories, no card required.

If the company pays, does the company own the Persona?

Not if the arrangement is structured properly, because payment does not determine ownership. The founder holds the account personally, and the company holds a licence for named uses and pays for it like any other subscription. Expensing is fine. Assigning ownership is the thing to avoid.

What happens to a founder Persona in an acquisition?

Under a licence, the licence terminates on change of control and the acquirer gets nothing unless the founder grants a new one. Under an assignment, the Persona transfers with the other intangible assets. Settle ownership before anything is built, not during diligence.

Can my board or my co-founder change what my Persona says?

No. Only the account holder can add or edit memories, and the product ships no admin screen that writes to a Persona's memories. After the Executor Lock activates, your personality and foundational memories become a snapshot nobody can change, us included.

Do I legally have to tell employees they are talking to an AI?

In the European Union the system has to tell them, from 2 August 2026, under Article 50(1) of Regulation (EU) 2024/1689, which binds the provider who builds the system rather than the company deploying one internally. What lands on you as deployer is Article 50(4), the moment anyone circulates a synthesised founder voice note. Elsewhere no law may require disclosure at all. Disclose anyway, because an undisclosed founder Persona found by staff costs far more than a sentence in an interface.

What does a founder Persona cost?

Building starts free: fifty memories, no card, and the free build does not expire. What the paid plans add is on the plans page, the only place kept up to date. Do the forty minute exception pass first.

What if the company shuts down, or Afterlife.ai does?

Ask every vendor in this category, and treat a promise of permanence as a bad answer. What matters is whether the material is exportable. Your account details, your Persona, every memory and every conversation download as a single JSON file, while audio and photos have to be requested separately. Personal ownership matters here too, since a company winding up has no claim on an account it never held.

• An AI Persona for creators and professionals, the general case across teaching, consulting, craft and company building.

• Capture your expertise before you retire, the same transfer problem from the individual's side.

• The things only you know how to do, on the operational knowledge you have stopped noticing you hold.

• Who controls your identity after you die, on consent, control and the legal patchwork behind an AI version of a person.

• How to interview yourself, the question sequences behind the exception pass.

The management claims come from the primary literature: peer reviewed articles, the authors' own working papers and the original books. The legal claims come from primary instruments only: the reported bill text on GovInfo, the Congress.gov record, the enrolled California statute, and the consolidated EU regulations on EUR-Lex plus the Commission's own guidance. Dates, figures, quotations and statuses were rechecked on 30 July 2026.

What is not settled: who owns an AI Persona of a person. The statutes in this area regulate the use of a voice or a likeness, and we could not find one that answers the ownership question squarely. The NO FAKES Act is a bill and may never pass. California's section 927 was written for performers and has not been tested against a founder's employment agreement. The ownership structure described here is a way to make the unsettled law matter less, not a ruling, and it is not legal advice.

Found something wrong or out of date? Report an error on this page and we will correct the page.

1. Gabriel Szulanski, "Exploring Internal Stickiness: Impediments to the Transfer of Best Practice Within the Firm", Strategic Management Journal 17(S2), 27-43, 1996. Study of 122 best-practice transfers across eight companies, 271 observations. Source for the finding that the major barriers to internal transfer are the recipient's lack of absorptive capacity, causal ambiguity and an arduous relationship, rather than motivational factors. doi.org/10.1002/smj.4250171105

2. Chris Argyris and Donald Schon, Theory in Practice: Increasing Professional Effectiveness, Jossey-Bass, 1974. Source for the distinction between espoused theory ("the theory of action to which he gives allegiance, and which, upon request, he communicates to others") and theory-in-use ("the theory that actually governs his actions"). infed.org

3. Michael Polanyi, The Tacit Dimension, 1966. Source for "we can know more than we can tell". Overview

4. Ikujiro Nonaka and Hirotaka Takeuchi, The Knowledge-Creating Company: How Japanese Companies Create the Dynamics of Innovation, Oxford University Press, 1995. Source for externalisation as the tacit-to-explicit conversion mode. Overview

5. Frederic Bartlett, Remembering: A Study in Experimental and Social Psychology, Cambridge University Press, 1932. Source for serial reproduction and the finding that elements failing to fit a listener's schema are omitted or converted into more familiar forms. Overview

6. Arthur Stinchcombe, "Social Structure and Organizations", in Handbook of Organizations, 1965, and Christopher Marquis and Andras Tilcsik, "Imprinting: Toward a Multilevel Theory", Academy of Management Annals 7(1), 195-245, 2013. Source for organisational imprinting and the persistence of founding-period structures after the environment changes. The authors' three-part definition is quoted from their working-paper text: "brief sensitive periods of transition during which the focal entity exhibits high susceptibility to external influences", "a process whereby the focal entity comes to reflect elements of its environment during a sensitive period", and "the persistence of imprints despite subsequent environmental changes". doi.org/10.5465/19416520.2013.766076 and working paper PDF

7. Noam Wasserman, "Founder-CEO Succession and the Paradox of Entrepreneurial Success", Organization Science 14(2), 149-172, 2003. Succession histories of 202 internet firms. Source for the finding that reaching milestones, completing product development and raising each round of outside financing, raises rather than lowers the chance of founder-CEO succession. doi.org/10.1287/orsc.14.2.149.14995

8. Noam Wasserman, "The Founder's Dilemma", Harvard Business Review 86(2), February 2008. Source for the analysis of 212 American start-ups founded in the late 1990s and early 2000s, for "successful CEO-cum-founders are a very rare breed", and for the tenure figures: 50% of founders no longer CEO by year three, 40% still in the corner office in year four, fewer than 25% leading their companies' IPOs. hbr.org

9. Ray Dalio, "How to build a company where the best ideas win", TED, April 2017. Source for radical transparency and algorithmic decision-making at Bridgewater. ted.com

10. S.4591, NO FAKES Act of 2026 (119th Congress), reported committee text, GovInfo, reported 24 June 2026; ordered reported by the Senate Judiciary Committee on a unanimous voice vote 18 June 2026. Not law as at 30 July 2026. Source for non-assignability during life including involuntary transfer, and for the written, signed, use-specific ten-year licence cap. Bill text and bill page

11. California AB 2602, adding Labor Code section 927. Signed 17 September 2024, effective 1 January 2025. Source for the three cumulative conditions that render a digital replica provision unenforceable, including the representation-by-counsel condition. Bill text

12. Regulation (EU) 2024/1689 (EU AI Act), Article 50, applicable 2 August 2026. Source for the Article 50(1) duty to inform natural persons that they are interacting with an AI system, and the Article 50(4) deployer duty to disclose deep fake content. Article 50, Article 113 application dates, and the European Commission FAQ on Article 50

13. Regulation (EU) 2026/1744 (Digital Omnibus on AI), of 8 July 2026, published in the Official Journal on 24 July 2026 and in force on the third day following publication. Source for the four-month transitional period on the Article 50(2) marking obligation for providers who had already placed their systems on the market before 2 August 2026, running to 2 December 2026. eur-lex.europa.eu

14. Regulation (EU) 2024/1689, Article 99(4). Source for administrative fines of up to EUR 15 000 000 or, for an undertaking, up to 3% of total worldwide annual turnover, whichever is higher, for non-compliance with Article 50. Article 99

Last reviewed: 30 July 2026. This page carries legal and regulatory claims with hard dates and is on a 90-day recheck cycle. The NO FAKES Act status was verified on 30 July 2026 and will be rechecked sooner if the bill reaches a Senate floor vote. Next scheduled review: 28 October 2026.