Thoughts on AI rollups

I recently researched the idea of AI rollups and thought I’d share a writeup of my notes to discuss the idea with others.

The main idea of AI-enabled rollups is straightforward: acquire services companies and supplement them with software that automates their processes, helping them serve more customers with fewer humans and increase margins in the process. Or with another lens: acquire companies as the way to distribute your software product.

Without any surprise, generative AI is the technology that makes AI rollups possible now. LLMs are very good at what most knowledge workers do: transform unstructured data into a polished document and handle human communication. This makes them uniquely fit for automating service jobs.

Let’s illustrate this with a service job I have some familiarity with as a customer: personal tax accountants. Each year, I hire my tax accountant to prepare my tax return and we follow the same process:

I receive an email that notifies me about my upcoming tax return and asks me to prepare my payslips, bank statements, and various other documents.
I send my documents via email or upload them in a shared drive.
I receive a list of questions, typically about the documents or to make sure I didn’t forget anything.
My accountant and their team consolidate all the information in a spreadsheet, and then use that spreadsheet to populate the self-assessment template.
I receive my self-assessment for confirmation. Upon approval, my accountant sends the tax return to HMRC.

My setup is fairly simple. Accountants have to deal with more complicated situations where they have to use judgement to handle undocumented situations or work with counterparts in other tax jurisdictions. But I'd be surprised if we couldn’t automate 90% of UK tax returns with LLMs and specialised workflows.

It’s not just personal tax accounting. GDPVal, a benchmark from OpenAI, shows that today’s leading models are almost at parity with experienced professionals on hundreds of knowledge work tasks across 9 industries and 44 occupations (e.g. drafting a legal brief, preparing an engineering blueprint, or handling a customer support conversation).

AI rollups are a great way to distribute technology

You can think of AI rollups as another route to distribute technology. For example, if you want to build software that automates personal tax accounting, there are two obvious options: you could sell software to personal tax accountants that helps them be more productive, or you could sell your personal tax accounting services to individuals and compete with tax accounting firms.

AI rollups offer a third avenue: acquire personal tax accounting firms and supplement their operations with your software. This approach has several distribution advantages over the other two:

There’s no market risk. You know the market exists. The business you acquire already has customers. Your task as a technologist is to automate existing processes to continue operating and growing the business with fewer humans, not discovering a new market and finding early adopters for your product.
Faster time to market. Since you are building software that helps accelerate work that is currently being done, you can put your software in the hands of users much sooner - even if your first version does a small part of their job.
Increased adoption. Making people adopt software that changes the way they work is often faced with resistance. At my previous company, after we signed a deal and the customer would commit to using our product, it would take months until the entire company would fully adopt our product, even when adopting earlier was in the customer’s best interest. If you own the company, it’s easier for you to redesign the way your employees work around a new piece of software.
You control the entire value chain. If you sell a point SaaS solution to personal accountants, your TAM can be limited to a fraction of the amount your customers charge their customers. If you own the tax accounting firms, you get all the revenue.

When you own the business, you can build automation that would be impossible with SaaS

To automate a process, you need access to data to know what to do and access to tools to actually do it.

If we continue with the example of personal tax accounting, you need data from emails and calendar to know when you last emailed or met with a customer, you need access to previous tax returns to know what to ask them about this year, and you need access to the CRM to know if there’s anything special that happened to them this year.

On the actions side, you need to be able to send emails to customers, schedule appointments, fill the CRM, collect information from PDFs and other documents, prepare the tax return, and send it to HMRC.

When you build a point SaaS solution, getting this data and performing these actions means building integrations with the tools your customers use. And each customer will come with their own unique combination of apps. This introduces three challenges:

Each integration is costly to implement, takes time, and requires ongoing maintenance when a new API is released or when the provider changes some feature.
Your automation is limited by what the API exposes. For example, some CRMs - like Salesforce - track changes to every field so you can easily list what changed for a customer in the past year (e.g. job, marriage, baby) but many - like Pipedrive - do not.
When customers use different tools, the automation potential will be constrained by the lowest common denominator of what tools your customers use or what is available via the APIs of what your customer uses. For example, you can’t rely on call transcripts if only a fraction of your customers generates them.

If you own the company, these constraints disappear. You can standardise the suite of apps your company uses so you only have to build one integration per category. If a vendor doesn’t let you access the data you need or if their API is not flexible enough to perform the actions you want to, you can pick another vendor. And over time, you can build custom tools that are fully integrated in your processes and tailored to your needs.

With complete data access and a unified stack, you can build a system that would be impossible with a point SaaS solution: your system can ask the client for documents via emails, compare this year received documents against last year and automatically email the client to ask for clarifications, continue chasing the customer until you receive all documents, consolidates everything into your tax preparation spreadsheet, generates the draft tax return, and sends it for their review - all without them logging into your software or you or them manually uploading documents anywhere.

A different playbook

AI rollups require a completely different playbook than building a point SaaS solution.

When you sell to enterprises, the standard playbook involves coming up with an idea, finding design partners that help you navigate the idea maze and build the first version of your product, then build a marketing and sales machine to distribute your product to many while you iterate on pricing. You sell to larger and larger customers as you build out features and enterprise requirements, and later use the distribution you’ve built to launch new products that serve the same customers.

With AI rollups, the playbook is completely different:

You buy an existing company with an existing set of customers and fees structure and work on automating the work being done.
You use improved unit economics to undercut competitor and/or spend more on customer acquisition to grow faster
You grow the business without needing to increase headcount thanks to automation
Use cashflow and improved multiples to acquire competitors
You consolidate and repeat

There’s a whole new dimension you need to add where you need to be constantly sourcing deals, and to build systems that supports onboarding new companies constantly.

Path to $100M

What would it take to get to $100M in revenue by building a AI rollup?

I built a small model that lets you adjust company target size, acquisitions per year, valuation, and a few other variables. (You can play with it below or open it in observable to fork it).

It appears, in theory, surprisingly achievable. Assuming 30% EBITDA and 1X valuations and ramp-up-periods for acquisitions, here are three paths to $100M in 7 years:

Acquire 12 small companies per year (after a 2 years ramp up period) at $1M each, grow at 30% Y/Y
Acquire 2 mid-size companies per year at $5M each, grow at 20% Y/Y
Acquire 1 company per year at $10M, grow at 10% Y/Y

What’s striking to me is how much more likely you are to reach that revenue with this strategy (in theory) compared to scaling a new product from 0 customers."

Challenges

While AI rollups have undeniable advantages in building and distributing software automation, they come with their unique set of challenges.

You need to raise far more capital to get started. Services companies like accounting firms in the UK are valued at .7x-2x revenues so unless you buy a very small company, a typical seed round won’t be enough to even get started.
The strategy relies on requiring fewer employees to do the same job. This means you either need fast growth to keep existing staff, or you need to reduce your headcount. In some verticals like property rental, it’s easy thanks to high employee churn, but in low-churn industries like accounting, it creates a difficult transition.
You need to identify and secure acquisition targets, which is a new skill entirely. Most importantly, you need to secure acquisition targets at reasonable multiples. This may get more complex as more companies run the AI rollup playbook.
Client retention and talent retention. The playbook only works if you can maintain your customer base while automating your processes away. To do so, you’ll likely want to keep the employees managing these relationships as long as possible.
Integration complexity, tech transition, and change management. Each acquisition comes with different practices, fee structure, and internal processes and tech stack that you have to consolidate. Even though you own the company, changing how employees work is organisationally difficult as existing staff may resist change for fear of replacement or simply because they prefer their current workflow.
You can’t automate everything. Some businesses have regulatory constraints where licensed professions need to sign off work. In high value services like accounting, there's a lot of value in the personal relationship clients have with their advisor. You can’t automate that either.

These challenges are an advantage to me. I’d rather have execution risk than market risk. It also increases the barrier to entry which can limit the number of competitors.

Where does AI rollup work well

AI rollups don’t work for all types of services. For example, an AI rollup of data consultancies doesn’t make sense. Data consultancies rent labour on customers’ terms and have no repeatable output. Even if there are some commonalities around data needs, each engagement comes with its unique combination of data sources, business peculiarities, and tech stack. Furthermore, consultancies typically charge for time (by hour, day, sprint) and efficiency improvements would cannibalise revenue.

Let’s outline several key features that I think are important to make an AI rollup successful:

Large and fragmented. You don’t want a single player controlling most of the market. They’d be too expensive to acquire and would have the resources to do the automation themselves.
Sells a product that AI can produce. If you sell plumbing services, the service requires human physical labour. You can use AI to facilitate coordination and optimise that physical labour, but your business will ultimately be heavily constrained by human labour. If what you sell are deliverables produced inside the Microsoft Office Suite, you can automate most of the process. If the output is sent via email, it’s probably automatable.
Growing market. Ultimately, your automation will allow you to serve more customers with fewer humans. If the market is growing, you can keep existing employees while increasing revenues and margins. If the market is not growing, you either have to accept lower margins or face the challenges (morale, talent retention, and severance costs) of layoffs.
Not tech savvy. Otherwise they’ll probably go off and build their own solution.
Repetitive workflows with some complexity. The economics only work if you can reuse your automation across many transactions like tax returns or insurance claims. But you don’t want a workflow that’s already commoditized (like payroll processing) as margins would be super thin.
Long term contracts. You want a multi-year customer relationship so you have time to recoup acquisition costs and benefit from margin improvement.

What are some examples of AI rollups?

Below is a list of AI rollups I’ve stumbled upon. I’ll try to keep the list updated as I discover more.

Buena.com offers property management services in Germany. The company builds property management software and acquires property management companies where they replace manual processes with that software. They find tenants, manage contracts, collect payments, manage repairs, and deal with emergencies. Property owners have an app where they can see everything they want about their properties - contracts, bills, inspection results, etc… - and have a dedicated account manager to contact for all matters. They claim that by automating tasks with their software, they increase the gross margins of the business they acquire from 5% to 35%.
Dwelly also offers property management services but in the UK. Unlike Buena which (appears to) runs under its own brand, Dwelly acquires lettings agencies, improves them by integrating its software, and then continues operating them under their own brand (e.g. Lime Property).
Multiplier holdings acquires “high-stakes” service businesses (e.g. tax, accounting, audit) and uses software to let them scale the number of customers they can serve. They have an interesting company set up with a holding company that makes acquisitions, tech pods that build software, and portfolio companies that serve customers.
Long Lake reportedly acquires managers of homeowner associations.
Crete PA acquires small to mid-market accounting firms in the US, empowers them with software and platform capabilities (HR, analytics, expense management…) and operates them under their existing brand.
Swally Wealth is a wealth management company that “buys smaller advisory firms”, “embeds artificial intelligence into the group’s operations”. They use “AI to take on back office tasks such as pulling data for the half a dozen forms that might be needed for any one transaction”
Groupe Archipel acquires accounting firms in France and lets them continue their activities under their existing brand while supplementing with software, and centralised HR and G\&A.
Crescendo builds vertically integrated contact centers powered by AI
Titan acquires Managed Service Providers (MSPs) (i.e. IT) and embed AI at the core of their operations, empowering technicians to work faster, smarter, and more strategically
Eudia is building an AI-Native Law Firm.
Nuvocargo, which provides customs and freight brokerage reportedly acquires firms in Northern America to scale their distribution

–

There are a few things that draw me to AI rollups: it turns building software companies from a work of discovery of what the market wants to an optimisation challenge where you are automating what customers want / have already. It’s also a way to fast track distribution. It’s way more fun to build when you have customers.

I am still wondering whether AI rollups are a temporary arbitrage. If all these jobs will be replaced by AI ultimately, maybe AI-native companies would be better fit for that? The only knowledge work that’d survive is where humans are essential to interact with the physical world, needed licenced humans to sign off the work, or where you need high trust relationships.

I am still early in thinking about AI rollups though. If you are thinking about the idea or have worked at a services firm that you believe is ripe for automation, I’d love to chat!