The crypto glossary at Andreessen Horowitz  gives the following definition of a Decentralized Autonomous Organization, or DAO:
“Decentralized autonomous organizations” or DAOs represent exactly what they are called; they are: (i) decentralized so, the rules cannot be changed by a single individual or centralized party; (ii) autonomous, so they operate based on logic written into a smart contract, without human intervention. They will continue to function for as long as the underlying blockchain continues to function; (iii) organizations or entities that coordinate activity among a distributed community of stakeholders – for example, developers and users on a given blockchain network. DAOs are examples of what is known as ‘on-chain governance’. In traditional corporate governance, for example, companies have bylaws that dictate certain policies, such as how a board is elected. A DAO extends this concept into the digital world by encoding bylaws into smart contracts.” 
The concept is interesting because it is based on the questionable premise that collaboration can be governed by algorithm: rules for decision making in a group can be specified as an algorithm, and whenever the group produces information that triggers the algorithm, the algorithm performs actions it was designed to perform on behalf of the group. For example, an algorithm may be designed so that anyone in the group can propose something to others, and if so, the algorithm will ask everyone to accept or reject the proposal, whereby the algorithm may be defined in such a way that a proposal is accepted if and only there are 50% + 1 votes in favor.
In traditional organizations, such as an academic institution or a corporation, governance mechanisms will be specified as rules over transactions and interactions (what is a business process than rules about who does what when?), and norms (rights, permissions, obligations) distributed across roles (and when a person takes on a role, they thus get responsibilities). As such, and relative to how a DAO functions, traditional governance mechanisms are underspecified; personally, I have never seen governance mechanisms defined to the level of detail where all input and output data types and allowed values are known, functions over data and their sequence available, that is, a traditional governance mechanism is never specified to the same level of detail an implementable algorithm is.
Underspecification does not work for a DAO (and there are many examples [2,3]) since the whole point is that the community needs clear and precise rules for how to accept or reject actions of its members.
Let’s take a simple example. You and I want to pool funds that we will then invest together; we need to agree on how we jointly make investment decisions; for example, we could say that we need consensus on every investment proposal in order to turn it into an actual investment; this can be defined as a simple procedure, where anyone of us can propose an investment (e.g., the identifier of the target investment and the amount of money to invest), then we both vote on it, and if we both voted to invest, the investment is automatically made. Because this governance mechanism is clearly defined, with straightforward inputs, outputs, and process to transform the former to the latter, that governance mechanism can be implemented as a program. The major difference, which makes DAO so interesting, is if we implement the program on top of an immutable and independently verifiable record of transactions/decisions (i.e., the blockchain). So we have a platform to record decisions and transfer funds, which is what blockchain and cryptocurrency enable, and then, we add on top of that programmed rules which filter out investment proposals.
Notice how interesting this is. The governance mechanism is implemented on a verifiable, decentralized, and immutable ledger, and we can program it in such a way that there are requirements for changing the governance mechanism itself. For example, we could say that not only you and I as participants make decisions on what to invest in, and how much, but also on whether and how to change the rule that requires us to have consensus on each investment proposal. Moreover, if the governance mechanism is triggered as soon as we reached consensus, it is autonomous in the sense that the only things you and I need to do, is to vote on each proposal (that you or I made) – that’s all: we do not need to vote, and then sign off on a joint decision, and then go and transfer funds; all that is automatically done as soon as we voted (if we had consensus, the rest happens automatically; if we did not, we need to vote again or make another proposal that will then be voted on). This is the Autonomous for A in DAO. Finally, implementing this governance mechanism allowed you and I to organize ourselves to do something that requires us to collaborate – this is the O for Organization in DAO.
Being able to program governance mechanisms on top of a decentralized infrastructure for transferring tokens that carry value opens up a lot of possibilities.
At the same time, it is important to understand limitations and have a critical attitude towards the hype around DAOs. There are claims that DAOs solve some long-standing governance challenges , such as the principal agent problem  and substantially reduce transaction costs ; those are hard to take seriously.
Well-Defined Governance Mechanisms Only
The governance mechanism needs to be such that it can be specified as an algorithm which can be implemented. This means that
- the governance mechanism needs to have triggers that can be observed without requiring human intervention (e.g., when I click to vote, no one needs to see me clicking, the software will register the click, and thereby the vote);
- the governance mechanism needs to have clear steps, with data as all its inputs and outputs (all parameters of a decision need to be so simple that they can be provided to software, through sensors, manually filled out forms, or by drawing from other and existing data sources);
- the governance mechanism needs to have a specified outcome for any combination of inputs (which implies that we know, at the time of implementing it, exactly which combinations of inputs lead to which outputs, including errors).
This is where it is hard to take DAO for an “organization”. It looks more like a way to streamline well-defined decision processes, where inputs, outputs, and steps between those, are clear and acceptable to all participants. In an organization, such as a firm, there will be many wildly different kinds of decisions to make all the time, the most important ones and the most interesting ones often being decisions made to solve unstructured problems, such as those made during strategic planning, resource allocation, competitive repositioning, product and service innovation, and so on. Only fragments of these complicated decision processes are amenable to representation as implementable algorithms.
This observation isn’t one about details of terminology. It is important not to be naive about the complexity of human organizations. Significant decisions, such as those I mentioned above – on values to promote, products to design, communications to make, and so on – are grounded in complicated, often implicit assumptions in the realm of personal values, beliefs about what is right and wrong, personal preferences, most of which will not be revealed prior to participating in the decision process. We can automate a procedure to allocate resources, for example, in a budgeting process, but this in no way solves problems of imperfect information and knowledge (the fact that different people, and in a DAO, accounts, have different information when making a decision), different values and preferences (which in human organizations usually involve interactions usually called politics and negotiation prior to the decision), and so on. In short, the fact that we can use DAO as a technology to automate fundamentally simple decision rules (e.g., majority voting, reputation-weighted voting, etc.) does not make the social reality any simpler or better in any particular way.
“A DAO is a centerless mesh network of agencies, which is also an agency in itself. There is no single point of control, or failure, in the organization. Instead of central management there is indirect coordination between agents, also known in biology as stigmergy, triggered by incentives and code. A DAO is a self-organizing entity, and at large better resembles an organism rather than an organization.” 
There is some truth to depictions of managers in Dilbert cartoons. The problem, however, that a management layer in an organization solves, is precisely the problem of coordination that DAO is also intended to solve. A simple way to think about the role of managers is that they produce and distribute information needed to coordinate the work of specialists. Management is also not always central, or saying that it is central is simplistic: there are various organizational structures, and even if org charts usually are hierarchical, this does not mean that information in an organization flows only through those same lines that connect positions in a hierarchy.
If so, then how does DAO accomplish this same function? The only way it could, is if these conditions are satisfied:
- all information required for coordination is available to all who may need it for coordination, let’s call this the transparency condition;
- everyone (in the DAO) has resources (time) to access all the information that may be relevant for their coordination with others;
- everyone has the competence to understand how the information that they have access to about the work of others impacts their own work, and vice versa.
In other words, a DAO will only work if everyone who participates in it is both a specialist, an expert whose work contributes to the value of the DAO, and at the same time sufficiently competent to understand the work of everyone else in the DAO, even if this may involve deep and various expertise. Obviously, there is a tension here between one’s role as specialist and generalist. This leads to an observation, a hypothesis really, that a DAO will look like functions in traditional firms: a DAO will group together people who have similar expertise; if so, this is much more limited than traditional organizations, which involve the coordination of multiple functions that need to work together to generate value. So DAO may be more suited to, say, the running of professional associations than to creating anything that resembles organizations that make and sell products and services. It would be great, however, to be proven wrong on this.
Voting and Its Variants
In most that I’ve seen so far (which is by no means exhaustive), DAOs implement variants of voting on a proposal. Some of these mechanisms are designed to increase one’s stake when voting, so as to make it costly or beneficial, depending on the implications of the vote on the community that constitutes the DAO. In other cases, votes are weighted by a measure of reputation, whereby one’s reputation is a function of an interpretation of that account’s past actions/transactions.
DAOs are truly interesting as a relatively cheap technology to coordinate when voting is good enough.
There is a problem with voting that DAO cannot solve: it is efficient, in that it requires little information and knowledge to cast a vote, but that also means that the decision can be an aggregate of incompetence and misinformation. There are no prerequisites when voting to know something in order to demonstrate specific competence, which then gives the right to vote. I’m not going to dig too deep into this, because it can be read as a criticism of democracy (which I have no intention of making); it is simply important to understand that there are limits to voting, and that even if it is a widespread mechanism for collective decision-making, this does not make it useful for any kind of situation.
Closed Systems Only
Let’s first clarify terminology: a closed system is one in which all parameters are known, and no new parameters can matter, and therefore, any new information is always information about new values of known parameters; an open system is one where new parameters can come up, and so new information can be information that there are new parameters, and about new values of new or known parameters.
I observed above that only well-defined governance mechanisms can be implemented in DAO; consider the subtleties of diplomacy, which have to remain out of scope for DAOs. Secondly, only mature governance mechanisms can be candidates for specification and implementation. Third, in decision-making which requires variable inputs (e.g., cases where the decision requires exploratory data analysis for example) will be out of scope for DAOs. More generally, any unstructured collective decision process cannot be implemented.
This implies that to design a DAO, you need to design the governance mechanism first, specify it as an algorithm, verify it, and then implement. This level of rigor is valuable, since it forces thinking through how decisions should be made, by who, why, before even starting to make any decisions.
Because the governance mechanism needs to be well defined, it will work only for a specific set of inputs. When new inputs, or unsupported values of inputs are encountered, it will fail. We can design it to fail gracefully of course, but the more important general observation is that a DAO works best for closed systems, not open ones: as long as it has no automatic means to adapt to unanticipated changes to inputs, it will fail where less well defined, but social and non-implementable governance mechanisms succeed. In social governance mechanisms, those we usually encounter in traditional organizations, unexpected inputs lead to discussion and negotiation, not failure to act.
Oddly enough, we could say that all DAOs are in fact Closed Decentralized Autonomous Organizations, not closed to new members, but closed to decision parameters that were overlooked at design time, and therefore either ignored, or lead to failure at run time.
Digital-Physical Interface and Identity
DAO are native digital so to speak. This begs the question of how it interfaces with the “real world”, the physical reality. One part of this problem is the general problem of ascertaining that a decision translates in the right action of the software, via hardware, on the world. That problem is certainly not specific to DAO, but to any kind of software which leads to changes in the physical environment (e.g., activates robots in a manufacturing plant). The other part of the problem is identity of people: if the DAO implements a governance mechanism whose possible outcome is that a specific person needs to do something in the physical environment, how do we ascertain that the person who does that is in fact the person identified in the DAO, i.e., the usually problem of tying identity in a system to a specific individual in the world. On the Internet, nobody knows you’re a dog. This makes DAO interesting for and “in” the metaverse, but far from being able to effect verifiable change in the physical world in the way that people do through traditional organizations.
Principal-Agent Problem and Transaction Costs
The principal-agent problem in economics concerns the alignment of interests between owners of assets (principals) and those given the right to manage these assets (agents). Transaction costs, on the other hand, are about boundaries of firms, that is, about which parameters influence (and how) the decision to perform some transactions inside a firm (between its employees) and others across the boundaries of the firm (between employees and third parties).
In light of the observations made above, I do not see the relationship between what DAOs are, and their ability to change anything regarding the principal-agent problem or transaction costs. I must be missing something here.
Implication of Limitations
The limitations above do not make DAOs uninteresting. When the governance mechanism to implement is clear, when decision inputs and outputs are or can be moved into software (i.e., are data), and when decision problems it applies to are repetitive, then DAO is a relevant means for implementing these governance mechanisms, especially in environments where decisions have an effect on data, not necessarily directly on the physical environment.
- The Crypto Glossary, Andreessen Horowitz, at https://a16z.com/2019/11/08/crypto-glossary/
- Aragon, 15 Ways the World is being Transformed by DAOs, at https://blog.aragon.org/15-ways-the-world-is-being-transformed-by-daos/
- CoinMarketCap, Top DAO Tokens by Market Cap, at https://coinmarketcap.com/view/dao/
- DAOstack, An Operating System for Collective Intelligence White PaperV1.1 April 22nd, 2018 at https://daostack.io/wp/DAOstack-White-Paper-en.pdf
- Jensen, Michael C., and William H. Meckling. “Theory of the firm: Managerial behavior, agency costs and ownership structure.” Journal of financial economics 3.4 (1976): 305-360. https://doi.org/10.1016/0304-405X(76)90026-X
- Williamson, Oliver E. “Comparative economic organization: The analysis of discrete structural alternatives.” Administrative science quarterly (1991): 269-296. https://doi.org/10.2307/2393356