Three Archetypes of AI Application Startups
On copilots, autopilots and service-as-software
This is a weekly newsletter about the business of the technology industry. To receive Tanay’s Newsletter in your inbox, subscribe here for free:
Hi friends,
AI is moving so quickly that it feels like we are very frequently seeing “weeks where decades happen” as Lenin put it. Despite the changing ground and ever improving models, over the past year it feels like we’ve settled into a few common archetype of AI applications, although startups are still moving between them as the technology improves or they realize one of them is the right approach to go-to-market in the market they’re going after.
The three broad archetypes are:
AI copilots, where AI functionality helps assist workers perform some core work or operational related tasks
AI colleagues, where AI agents autonomously handle some subset of tasks for human workers, with human workers “managing” them
AI native services, where companies provide an end-to-end service to other companies using a mix of AI agents and humans on the backend
Let’s go deeper into them.
1. AI Copilots
AI copilots act as supercharged assistants, designed to enhance productivity by supporting users in various tasks. These copilots help people get up to speed faster, execute core tasks quicker and better, and even assist spending less time on mundane tasks.
ChatGPT, Claude, Glean, Microsoft Copilot among others essentially serve as general-purpose copilots for many different kinds of knowledge workers, but more and more products are emerging which are more specialized copilots for certain roles and functions.
Examples include:
Github Copilot: The most well-known and most successful product in this vein is Github Copilot which is an AI pair programmer that suggest lines/blocks of code and is integrated into developers’ workflows. Github Copilot has 1.8 million paid subscribers and generates over >$100M in revenue.
Casetext and Harvey which are copilots for legal professionals which assist with core legal workflows such as research, analysis and document creation.
Copilots can help people be more productive on the job and assist them to do their core work tasks better and faster. Separately, some copilots focus or assist on the admin/more mundane aspects of the job so that workers have more time for the more interesting work they do. For example Billables helps lawyers and other service professionals with timekeeping, Anterior helps clinicians with admin tasks and a number of different products help with various forms of note-taking and recording entries in appropriate systems of records.
Every incumbent has also been rushing to add in copilots into their products, such as Salesforce, Adobe, Microsoft, and Google, but there will still be opportunities for startups, particularly in specific verticals or functional roles to go much deeper into the specific workflows performed by that role.
Additionally, some companies are also building copilots for their employees internally. For example, Lilli is an internal copilot built for consultants within McKinsey, to help them synthesize transcripts, research their knowledge library and generate artifacts.
2. AI Colleagues
AI colleagues go a step further by acting not just as assistants, but as proactive, autonomous agents that can take on more complex, decision-making tasks themselves. These AI agentic systems are designed to work alongside human teams, seamlessly integrating into the workflow and often acting autonomously to manage specific tasks.
Some people often think that these AI colleagues may map 1:1 with what human colleagues in that role did, which isn’t always the case. Instead, they will fully automate/autonomously carry out some portion of the work, such that a human employee can offload certain types of tasks to it.
The idea of “selling the work” with AI has been en vogue, and both AI colleagues and AI-native services (covered next) are different encapsulations of that. In the AI colleague case, the “human work” is done by the customer, while in AI-native services, the “human work” is done by the startup selling the services itself.
Examples include:
Devin, which is billed as the world’s first AI software engineer. Companies can “hire” Devin, the same way they may a human engineer. Now, a human engineer, typically an IC, may have access to a few Devin’s that they manage (the way they may manage an intern), allowing the human engineer to produce more output when factoring the output of the Devin’s in. We’re also seeing more narrow AI colleagues in engineering such as specifically for QA/testing or for SRE.
11x, AiSDR and others offer autonomous AI SDRs. These AI SDRs are given an ICP, targeting criteria, and autonomously identify leads, research accounts, and personalize emails to them, and book meetings, to then be handled by a human AE. Again here, while they aim to be fully autonomous, today they may be used by an SDR to perform the output of 5 of them as an example, or allow for a sales-team to be restructured such that every AE has their own Ai SDR which helps it develop pipeline that they can manage (And sends emails in the AE’s voice/tone).
Sierra, Decagon and others offer autonomous support agents that can resolve a reasonably high fraction of support queries (50-70%) across email/chat/phone autonomously. In these cases, these function as a colleague to other agents, except they can handle only a subset of queries but at extremely high volumes. Therefore, they may complete 65% of the queries but only roughly the easiest 65%, meaning that the remaining human agents now handle the more difficult queries only, that still may take 50-60% of the time (which they may be assisted by an AI copilot for). In that world, human customer support teams may need to be upskilled such that most of them can handle the complex tickets to resolve.
Slang which is an AI-powered receptionist that can handle inbound calls into restaurants and stores, scheduling appointments and answer questions for them.
AI colleagues are particularly valuable (and work well today) in roles where a high volume of decisions are required and the processes are generally quite repeatable and well-understood (or the cost of a mistake is generally low). They enhance team capabilities by taking over a large volume of these workflows/decisions, allowing human colleagues to deal with the more complicated work.
Over time, we may also see some of the AI copilots shift towards “doing the work” and become AI colleagues for some tasks, while continuing to assist and be a copilot for others.
3. AI-Native Services
Sometimes, customers still want a service handled for them end-to-end, and don’t want to do the human part of the work that is required in addition to the AI to get the job done end to end. That’s where AI-native services businesses come in.
AI-native services startups are essentially “full-stack businesses” that are solving a problem end-to-end for the customer. Many past managed marketplaces such as Uber and Doordash were full-stack businesses delivering outcomes to the end-customer, managing all the complexity and interplay between tech and labour themselves.
Now with AI, many other service markets, particularly more knowledge work oriented ones can be transformed and 50-95% automated when the workflows are understood and the human workflows can be altered to fit in with what AI can do. But some companies may still want to just buy the “service” rather than buying AI and then using it internally get some of these automation benefits.
These AI-native startups typically compete with old-school non-tech incumbents that may be traditional service firms, agencies and outsourcing firms and aim to provide the service at equal or better quality but with much higher automation and therefore lower costs.
We’ve seen many flavors of tech-enabled services businesses over the years, where the amount that the “tech enabled” might not have been very large. AI just allows for actually seeing a potentially meaningful difference between a typical services firm and a truly AI-native one, when done correctly.
As AI advances, the markets where this is feasible will expand, but today, many of the AI-native services firms are going after verticals that are document/code heavy with large amounts of paperwork such as legal services (immigration, patents), financial services (tax, accounting), engineering implementation/integration work.
Examples include:
Alma (immigration) and Marble (family law) which are tackling various legal services sectors, competing directly with other law firms
Pilot (bookkeeping / CFO) and Gelt (taxes) which are going after various accounting markets, competing directly with other accounting firms
Isoform which is provides custom integrations as a service, competing directly with implementation consultants.
A couple of points on this approach in closing:
While thanks to the benefits of AI, these startups may be more scalable or have better margins than their service counterparts, ultimately these are service firms and likely will be valued in similar ways to them on EBITDA multiples.
Sometimes, these companies may use acquisitions as a GTM strategy when taking this approach. These acquisitions may be after building out the tech or in some cases, the company may start by buying an incumbent first and then building the AI to automate and improve margins.
I’ll go deeper on these points in a future post!
We’re excited about all three archetypes at Wing and have been actively investing across each of them. If you’re building something in any of the areas, feel free to reach out at Tanay (at) wing.vc
thanks for the mention!
This is a very insightful piece, thanks Tanay!