The Changing Intersection of Software & Services
An open letter to our team, and to anyone running a services business, on what AI is doing to the line between agencies and software. Article one of many to come.
TL;DR
- AI has made execution cheap, so the value for service providers like agencies is shifting toward judgment: knowing what's worth building and where to point the tools.
- The agency-to-product path that produced products like Mailchimp is still open, but it's changing shape: products and services built together instead of a single product outgrowing the agency.
- Agencies have a head start on AI-native services because they sit inside client problems, but cheap building means choosing deliberately what to invest in.
In the early 2000s, there was a small web design agency in Atlanta called the Rocket Science Group. You've probably never heard of them, but they're one of my agency heroes.
I'll admit I can't name a project or client, but they did something almost impossible: they went from agency to product. Founders Ben Chestnut and Dan Kurzius built an email marketing tool for their small-business clients, and it became Mailchimp, used by countless businesses. In 2021, Intuit bought it for $12 billion.
When Edgar Allan started in 2013, one of our ambitions was to build digital products, first for our clients and eventually for ourselves, on our own dime and without outside capital. Our goal has never been a billion-dollar exit. I love building things the way a songwriter loves writing songs. For me, building is how you bring something new into the world.
We've been building things since the beginning, and not all of it was software - from our office, card games to a D2C bed-in-a-box company called Loblolly. Then a series of products - most notably - Slater, now our most widely used product, and WES, which over the years has quietly paid the bills.
Building products is hard, and we're getting better at it. But over the past year I've watched the line between software and services start to move, and it's made me question whether the model that made Mailchimp possible still makes sense for us.
The path from agency to product is changing
The Rocket Science Group had something that's hard to imagine now: time. Mailchimp started in 2001 and grew alongside the agency until it became the business. Ben and Dan had years to find customers and build a company around the product.
I don't think we get that. Four years ago, ChatGPT wasn't public. Today someone who has never written code can describe an app and have working software by lunch. In February, software stocks sold off hard because people feared AI makes software cheap to build. The good news is that while code is cheap, building something that works for people, and now for agents, is alive and well. Still, I can't picture spending a decade on a standalone product and expecting the software, or the market, to look anything like where we started.
Time cuts both ways. We've been building products for years, and one of the hardest parts has always been putting together a full product team: research, strategy, UX, design, and engineering. Today our product team is about three people, and I think they can get more done than a 30- or even 300-person team could when Mailchimp was getting started.
The cost of execution is going down. The value of knowing where to point these tools is going up. As execution gets cheap, outputs start to look the same. The companies that win will know what they stand for and who they're for. That's true for our clients and for us, and it changes what the services in our stack are. Less of the work is producing things. More of it is the judgment that decides what's worth producing.
But I don't think the path from agency to product is closed; I think it's changing shape and quickly.
The clearest version of that argument comes from Julien Bek at Sequoia. In Services: The New Software, he points out that for every dollar companies spend on software, they spend six on services. His bet is that the next big companies will sell outcomes, using AI and human expertise together to deliver them. He also thinks they'll charge differently, billing for results instead of hours.
Y Combinator is making the same bet about agencies. In its spring Requests for Startups, partner Aaron Epstein wrote that AI-native agencies will “look more like software companies, with software margins.” YC expects new startups to get there first. I think agencies like ours have a head start. We sit next to clients while they're trying to solve problems. We see where their systems break, often in places no outsider would find. And now we can build the fix in a fraction of the time and cost it used to take.
So our goal now is products and services that make each other better. We already do a version of this. Our AI site search ships inside client engagements, and CSP and our AEO and SEO work keep us inside client sites month after month, which is where we spot the next problem worth solving.
Not everyone buys the services-as-software idea. Nathan Benaich and Nikola Mrkšić argue in Fortune that the best services firms use AI to make their people better, and that the smarter bet is to own the software. Both camps are partly right. We want to own the software and the service. It's also early, and nobody knows yet where this settles.
Getting there changes two things: how people grow here, and how we decide what to build.
Managing products is just another day
Traditionally, you could grow your career in one of two directions: as an individual contributor or as a people manager. Now there's a third option that can be mixed in - managing products.
We used to document our processes in PowerPoints and on whiteboards. Now we can turn those processes into products that cut the effort inside a given block of work, like the discovery synthesis a strategist runs on every project. The product doesn't have to be special. It should be everyday, and it doesn't have to live forever. The question is how does it make us better?
So it's less of an either/or. The more interesting question is how a role becomes a multiplier. How do we set a direction and execute so one person can have an outsized impact, with the same set of resources, whether that's people, tokens, or the constraints we live with day to day?
The multiplier is already showing up in the data. Emergence Capital, which coined the term "AI-native services," reports that across its portfolio, practitioners who live in AI tools deliver three to ten times the output of their peers on comparable work. They also warn that a firm's culture is the weighted average of who does the work and how they spend their day. We can't bolt AI onto a traditional agency and call it something new.
This asks a lot of our managers. People who've spent their careers developing teams and delivering client work now also have to decide what's worth investing in, and when to stop.
Not every idea needs to become a company. Most will make us better at our jobs. I want room for all of it.
We can build anything but not everything
A few times recently, someone has shown me what they're building, and it turns out someone else here is building nearly the same thing. I love that people are building, but it tells me collaboration is and will be a huge consideration.
Software used to come with a built-in brake. You needed time, a budget, and perhaps someone's sign-off before you started.
I don't want to put the brake back on with an approval process. Being able to try something without asking is what makes this moment exciting. But trying something and investing in it are different decisions, and right now most companies (ours included) aren't making the second one intentionally.
The companies that get good at this will move fast without five people solving the same problem alone. I'd like us to be one of them.
Where we go from here
Net: I'm more excited about building products than ever.
I used to picture one product that grew into a company and eventually left the agency. Now I picture products and services built together, each tied to a problem one of our clients or our own team has, with a few of them growing into businesses of their own.
So what does this mean for Edgar Allan right now? We're still figuring it out. We're exploring new ways of building with Slater. As we do more work around optimization and systems, Slater is a natural surface for custom applications. Some will stick around for years. Others will be ephemeral, built for a single discovery session, for synthesizing a stack of documents, or for deciding what to keep and what to throw away in a migration.
Our products will keep coming out of client work, and now we need to find ways to pick what we build, buy, or mix.
FAQs
What does “services as software” mean?
It’s the idea, argued by Sequoia’s Julien Bek, that the next big companies will sell outcomes rather than tools, delivering them with AI and human expertise together. Companies spend about six dollars on services for every dollar on software, so the opportunity is in services, not seats. Firms built this way may also bill for results instead of hours. Edgar Allan’s view is that agencies can own both the software and the service.
Can a digital agency still become a product company?
Yes, but probably not the way Mailchimp did. The Rocket Science Group had years to grow a product alongside its agency until the product became the business. Today markets and software change too fast to spend a decade on a single standalone product. The more realistic model is a portfolio of products that grow out of client work, a few of which become businesses of their own.
Why would an agency have an advantage over AI-native startups?
Agencies see client problems up close, including where systems break in places no outsider would find. Y Combinator expects new startups to build AI-native agencies first, but an existing agency already has the client relationships and the context. With AI cutting the cost and time of building, the agency that spots the problem can now build the fix too.
How is AI changing career paths at agencies?
Alongside the individual-contributor and people-manager tracks, there’s now a third option that can be combined with either: managing products. A strategist might turn a repeatable process, like discovery synthesis, into a tool that cuts the effort on every project. Emergence Capital reports that practitioners who work deeply in AI tools deliver three to ten times the output of their peers on comparable work.
How should an agency decide which internal tools to build?
Keep experimenting cheaply, and make investing a deliberate decision. When anyone can build anything, teams end up solving the same problem separately unless there’s a way to see what others are building. The aim isn’t an approval process. It’s choosing on purpose what to build, buy, or combine, and knowing when to stop.