There's a sleeper hit lingering in the shadows of agency work. Websites may be your bread and butter, but when we dug into our data, we found the work that outlives them. (And we do not gatekeep.)
Let's look at some numbers first:
- Agency net margin fell to 13% in 2025, down from 14%
- The average agency offers 6.6 services, exactly what it offered in 2022
- 70% of agencies changed which services they offer
- But agencies are dropping web design, UX/UI and mobile builds while adding branding, content and SEO. AI implementation is the fastest riser and still sits at 17%
Figures from Promethean Research's 2025 and 2026 Digital Agency Industry Reports.
Same number of services, but different services, less money. The shelf is being restocked every year and the margin keeps slipping. What almost nobody is adding is software.
So we went looking for what survives. Bolt.new agency users have built 501,479 projects. We classified them, then tracked which ones were still running a week later.
The answer surprised us.
Internal tools, portals, and CRMs may only account for 4.4% of what agencies build, but they clear our production bar more often than any other category: published, backed by a database, and still running seven days later.
That means two things.
- Internal apps give you a pipeline of maintained work you can keep billing for.
- If our data holds true across agencies at scale, not a lot of people are doing this yet.
Bolt platform data, September 2026. n = 149,687 self-identified agency users, 501,479 projects. Market figures sourced inline.
Why margins are under pressure
Two numbers explain most of it.
The first is scope creep. 57% of US agencies lose between $1,000 and $5,000 a month to work they never bill for, another 30% lose more than that, and 78% say they rarely or never charge for out-of-scope requests (Ignition, 2025). That's revision rounds, mostly, and revision rounds are a project-work disease.
The second is size. Net margin runs backwards to headcount: studios under ten people averaged 19% in 2025, while agencies over fifty averaged 8%, a spread Promethean labels a directional benchmark (Promethean Research, 2026). Promethean attributes that to the management layers that arrive with growth, and its analysis never mentions AI, which is worth saying because it's tempting to read it the other way.
Put those together and the shape of the problem is clear. Small is already profitable. What erodes it is unbillable rework on project engagements.
What agencies build
| Category | Share of projects | Published | Connects a database | Production-grade |
|---|---|---|---|---|
| Client and business websites | 12.9% | 14.9% | 15.6% | 1.48% |
| AI and automation tools | 8.8% | 12.7% | 28.6% | 1.78% |
| SaaS and app MVPs | 6.9% | 12.1% | 32.9% | 1.76% |
| Marketing and growth pages | 6.2% | 14.3% | 12.2% | 1.47% |
| Internal tools, portals, CRMs | 4.4% | 12.2% | 33.5% | 2.43% |
| Dashboards and analytics | 3.8% | 10.1% | 22.4% | 1.70% |
| E-commerce and payments | 2.6% | 11.7% | 26.2% | 1.34% |
| Design systems and prototypes | 2.0% | 8.1% | 14.9% | 0.70% |
Behind those categories sit the descriptions people typed when they opened a project. Company websites for named brands. A daily ad idea generator for a direct-to-consumer client. A car pre-order portal. A maintenance schedule generator. This is billable work, described the way someone describes a job they took.
Not all of it reads that clearly. A quarter of projects resisted classification, and another 17.5% are throwaways and forks. Nearly half of what agencies build is unlabeled or disposable. That is what a tool people open to try something looks like, and we counted it anyway.
Which of it survives
Most of it doesn't. 81% of agency projects are built and finished inside a single day. 11.6% get published. 21.3% connect a database. And 1.5% clear all three bars at once.
Then the interesting part. Internal tools reach production at 2.43%, against 1.48% for client websites, and a third of them connect a database. Fewer projects, more of them alive.
Client websites are the mirror image. They get published more than almost anything, because work for a client has to go live. But fewer connect a database, and fewer are still running a week later.
The pattern is consistent enough to plan around. Work built for a client gets published. Work built to run a business gets maintained.
Why the difference matters
A published website is a delivery. A maintained internal tool is a relationship.
The website is finished when the client signs off, and the next revision round is a negotiation about whether it's in scope. That's exactly the work the margin data says agencies are giving away.
The internal tool is never finished. The client adds a field, a user, a report. Somebody maintains it. That somebody bills monthly instead of arguing about scope, and the work compounds instead of resetting with each project.
This is also the direction the market is moving. AI "is creating new demand for strategy, implementation, governance, workflow redesign, automation, and agent development" (Promethean Research, 2026). All of those are maintained systems, not delivered assets. And the agencies that narrowed their offer to do fewer things well grew 13%. They averaged 30% net margins against 10% for agencies that added services, a gap Promethean also calls directional.
That is the opportunity in a small number. 4.4% is not much of what agencies build, but it is the part that keeps paying. Taking on more of it is a capacity question rather than a hiring one (how small agencies add capacity without hiring).
Should agencies stop building websites?
No, and the data wouldn't support it if we said so. Websites are the largest category of agency work on the platform and the most likely to be published. They're how most agencies get paid.
There's also no credible survey showing agencies broadly replacing campaigns with applications. Service counts are flat. The agencies growing fastest are the ones that narrowed, not the ones that added a software practice.
The argument is narrower than that. The website is the engagement. The internal tool is what turns one engagement into a standing one.
How we counted
The workplace field is self-reported at signup. This measures what people call themselves when they start building, not audited company registrations. Categories come from a two-pass classification of project descriptions, which is why a quarter stayed unclassified. Production-grade means published, connected to a database, and alive seven days later.
What to do with this
Look at your last five client projects and ask which one the client would notice going down on a Tuesday.
That's the one to build next, and the one to charge for monthly.
How agencies build that work in-house, without a developer on staff: Bolt for agencies.
