Towards effective entrepreneurship: what makes a startup high-impact?


This post owes a great deal to prior work and thought by Spencer Green­berg, Eric Gast­friend, and Peter Hartree.

This post is a sum­mary of the ob­ject-level thought on what makes a startup high im­pact which we de­vel­oped while work­ing on the Good Tech­nol­ogy Pro­ject.

A lot of this ma­te­rial is more-or-less ob­vi­ous ap­pli­ca­tions of EA thought to startup the­ory. Nonethe­less, it man­aged to be sur­pris­ing and use­ful to peo­ple, so per­haps it is less ob­vi­ous than it seems. I’ve con­densed the pre­sen­ta­tion given the in­tended au­di­ence of this post—there is a lot more to say on many of these points. This ma­te­rial might have even­tu­ally de­vel­oped into a “guide” to effec­tive en­trepreneur­ship.

In ad­di­tion, some of the ma­te­rial re­lates to how to man­age a startup in later stages. We never re­ally got a chance to try that out, so it is es­pe­cially spec­u­la­tive.

What makes a startup high im­pact?

We’re in­ter­ested in star­tups be­cause we think that they might be a mechanism by which we can have a large pos­i­tive im­pact on the world. But what are the qual­ities that we should look for in a startup?

Im­pact model

Be­fore we get started on as­sess­ing how good a com­pany is, we should try and get clear on how that com­pany benefits the world, to make an im­pact model for the com­pany.

The first big con­sid­er­a­tion is who the com­pany helps. Usu­ally there will be one group in par­tic­u­lar that you are ex­pect­ing to benefit. A good way of figur­ing out who these are is to con­sider the var­i­ous groups of “af­fectees” for your com­pany.


Cus­tomers are the most ob­vi­ous peo­ple who benefit (or suffer) from the ex­is­tence of a com­pany. They pay a cost in money and time, and they gain your product in re­turn.

For ex­am­ple, Mesh Power’s1 pri­mary benefi­ciary group is its cus­tomers (in­so­far as you think that pro­mot­ing clean en­ergy over burn­ing kerosene might have en­vi­ron­men­tal benefits, Mesh Power may also have some ex­ter­nal­ity benefits, see be­low).

While we can usu­ally as­sume that peo­ple will buy things that ac­tu­ally im­prove their lives, this isn’t uni­ver­sally true. Ci­garettes and ad­dic­tive drugs or games are ex­am­ples of this.2

Third parties

The op­er­a­tion of your com­pany will also af­fect peo­ple who are not part of the trans­ac­tion, or even in­volved at all. Th­ese effects are called ex­ter­nal­ities. Often these are pos­i­tive, in the case of in­no­va­tion and eco­nomic growth, but they can also be nega­tive, such as pol­lu­tion, de­vel­op­ing dan­ger­ous new tech­nolo­gies, or caus­ing tech­nolog­i­cal un­em­ploy­ment.

For ex­am­ple, de­spite be­ing a car com­pany, ar­guably Tesla’s pri­mary benefi­ciary group are third par­ties, be­cause ac­cel­er­at­ing the progress of elec­tric cars and stor­age will help to ame­lio­rate cli­mate change.

An im­por­tant class of ex­ter­nal­ity is benefits pro­duced by your cus­tomers, which will of­ten hap­pen if you’re sel­l­ing to busi­nesses or in­sti­tu­tions. For ex­am­ple, dis­ease out­break mon­i­tor­ing sys­tems may be sold to gov­ern­ments, but the benefi­cia­ries are the peo­ple who don’t get ill be­cause of the gov­ern­ment’s’ im­proved pre­ven­ta­tive ac­tion.


A third cat­e­gory of benefi­cia­ries is your em­ploy­ees. They will gain pay and satis­fac­tion from work­ing for you, but will also spend their time. In bad cases they could ex­pe­rience phys­i­cal or psy­cholog­i­cal harm be­cause of the job.

For ex­am­ple, one of M-PESA’s benefi­ciary groups are among its em­ploy­ees, since it needs lots of places for cus­tomers to buy and sell mo­bile money, and this pro­vides ad­di­tional in­come for a lot of rel­a­tively poor shop own­ers.

Im­pact mechanism

The next thing to do is to work out how you think your startup will ac­tu­ally af­fect your tar­get group of benefi­cia­ries. This is likely to be very un­cer­tain, es­pe­cially if you ex­pect to cre­ate an im­pact through ex­ter­nal­ities. How­ever, it’s bet­ter to ex­plic­itly write down what you’re un­cer­tain about nonethe­less.

For ex­am­ple, here’s one mechanism by de­vel­op­ing a bet­ter test for drug-re­sis­tant TB might im­prove wellbe­ing:

  • De­crease cost of TB test

  • In­crease availa­bil­ity of test in low-re­source areas

  • Ac­cu­rately dis­t­in­guish more cases of drug-re­sis­tant TB from nor­mal TB

  • Give more drug-re­sis­tant TB suffer­ers the cor­rect drugs

  • Cure more peo­ple of drug-re­sis­tant TB than otherwise

  • Fewer peo­ple go through the lengthy suffer­ing of drug-re­sis­tant TB

  • In­crease wellbeing

There may well be sev­eral such mechanisms, of course!

Once you have an ex­plicit im­pact mechanism, that gives you a two use­ful things: a set of hy­pothe­ses about how your im­pact oc­curs, which you can test; and a set of stages in the mechanism which you can mea­sure.

Most of these won’t be things you can test or mea­sure now, but it’s worth think­ing from time to time whether you might be able to mea­sure more of them. For ex­am­ple, in early de­vel­op­ment you might fo­cus on mea­sur­ing the cost of the test, but as you roll out you might also be able to mea­sure im­prove­ments in availa­bil­ity.

Assess­ing the im­pact model

We can ap­ply our usual INT heuris­tics in this case, al­though we can pick out some par­tic­u­lar con­sid­er­a­tions for the do­main. Th­ese can work both for pick­ing out a broad prob­lem area, and for di­rectly pick­ing out fac­tors rele­vant to a par­tic­u­lar im­pact model.


As ever, we care about both how many peo­ple we help and how much we help them.

We should think about max­i­mum scale here: if you could even­tu­ally sell your product to ev­ery­one on Earth, that’s bet­ter than if you’re limited to just one na­tional mar­ket. If we think about our pos­si­ble benefi­ciary groups, third par­ties tend to be the biggest group, fol­lowed by your cus­tomers, and then your em­ploy­ees.

Similarly, a life-sav­ing product is much bet­ter for each per­son than some­thing that merely saves them some money.


There are a cou­ple of big things that af­fect tractabil­ity.

The first is ob­vi­ous: the prob­lem may be hard. Or the prob­lem may be easy, but mak­ing it prof­itable may be hard. And we’re pri­mar­ily think­ing about busi­nesses here, so if you can’t make it prof­itable, you can’t do it.

Se­condly, you might not want to do it. Run­ning a busi­ness is hard work, and you face pres­sure not only to drop out, but to cave in on is­sues where your in­vestors or ad­vi­sors may not be al­igned with what you want. If your benefi­ciary group is your cus­tomers, then your profit goals and your im­pact goals are rel­a­tively al­igned, so this may be eas­ier.

In other cases this is less likely. For ex­am­ple, Uber (may be) benefit­ing its 1.5 mil­lion drivers. But they are not in­cen­tivised to em­ploy these peo­ple, be­cause do­ing so costs them, so as soon as they can au­to­mate them away, they will.

Fi­nally, you might not be able to figure out what to do. Even if you can iden­tify the prob­lem, you may not be able to figure out a plau­si­ble mechanism to ac­tu­ally have an im­pact on it, or your mechanism might fail to work.

Tractabil­ity is­sues re­sult in two big failure modes:

  • The busi­ness fails entirely

  • The busi­ness suc­ceeds, but it has a low or nega­tive impact


As­sum­ing that you start a busi­ness that solves a real prob­lem, we can as­sume that some­one would have solved it even­tu­ally. That means that the effect you have is the differ­ence be­tween those two, which will look like get­ting X ex­tra years of the solu­tion. We can call this your time ad­van­tage.

Gen­er­ally, the big­ger the time ad­van­tage the bet­ter. If the prob­lem is big enough, then even a short time ad­van­tage may not be a prob­lem—get­ting a malaria vac­cine a year ear­lier would be huge!

But gen­er­ally big­ger is bet­ter. There are a few ways you might have a big time ad­van­tage:

Firstly, the tech­nol­ogy you use has ex­isted for a while but hasn’t been ap­plied to the prob­lem that you are ap­ply­ing it to. That sug­gests that it would con­tinue to be un­solved in that way for a long time if you don’t do it.

Coun­ter­in­tu­itively, this sug­gests that you should stay away from new tech­nolo­gies: it is very likely that some­one will try “ma­chine learn­ing for X” rel­a­tively soon, so it is un­likely to be ne­glected.

Another com­mon case is that the prob­lem re­quires an un­usual com­bi­na­tion of skills, knowl­edge, or in­cli­na­tions. For ex­am­ple, you might know about both fi­nan­cial ser­vices and the de­vel­op­ing world, while also be­ing al­tru­is­tic. Com­bi­na­tions of traits are cor­re­spond­ingly rarer—if you have at least one mod­er­ately rare skill, then it is likely that you also have one very rare com­bi­na­tion of skills. It may be a long time be­fore this com­bi­na­tion comes along again, and so if there are prob­lems that re­quire it, they may go un­solved un­til then.3

This sug­gests that you should look es­pe­cially hard for prob­lems that only you (or you and your friend with the other un­usual skills) can solve, be­cause that is likely to give you a big time ad­van­tage.

Fi­nally, the in­cen­tives to solve the prob­lem may be lack­ing (e.g. the cus­tomers are poor). This is a tough case, be­cause those in­cen­tives will also be lack­ing for you. So you need a good story about how you are go­ing to keep your im­pact on track. Many benefits to third par­ties have this form. Often if the ex­ter­nal­ity is in­no­va­tion then a strong founder can en­sure that most of the benefit is pro­duced be­fore they are phased out. For ex­am­ple, Tesla has cho­sen to give away their patents for free, which might not have hap­pened with a less al­tru­is­tic CEO.

  1. Sadly, it looks like they’ve gone bust since I last checked, but they’re still a good ex­am­ple of the prin­ci­ple.

  2. Spencer Green­berg’s pod­cast dis­cusses some of the ways star­tups can un­ex­pect­edly cause harm.

  3. Peter Thiel talks about “se­crets” which are un­usual be­liefs that you have which make you think that a prob­lem is sol­u­ble, even though the gen­eral be­lief may be that it is not. Th­ese are an­other thing that can make you un­usual.