Build vs Buy: The high bar for building your own AI agent

Build vs buy: The high bar for building your own AI agent

It’s enticingly easy to build AI-powered software. But if you do it for the wrong reasons – or don’t grasp the trade-offs – you’ll likely waste time and money.

The breakthrough capability of LLMs has flipped SaaS on its head: the biggest prize is in building software that replaces human work, instead of building tools that help humans do that work.

Our focus is on AI-first customer service, and our vision is that in a short number of years, the vast majority of customer service interactions will be with AI, not humans. This is a multi-hundred-billion-dollar opportunity and success means enabling all businesses to be leaner and more nimble, while also providing world-class customer service.

Every company can benefit from this shift by redeploying talent and capital to innovation and growth. But we are still early in the adoption curve, and great outcomes are not guaranteed – they hinge on the choices you make. Two stand out:

  1. Whether to build or buy an AI agent.
  2. If buying, the AI agent you choose. (Hint: Fin.)

I have a simple mental model for ROI (return on investment): What percentage of CS volume can I automate? And how much does it cost relative to human costs?

A further important nuance is that as you increase the automation rate, you are doing harder and more costly work so the ROI gains accelerate.

I respect anyone who wants to build their own AI agent, but here are the pitfalls you must overcome to match – let alone surpass – a “buy” approach.

The fundamental pitfall: Falling short of “best-in-class”

It’s easy to automate low-hanging fruit. The real question is how high you can push your automation rate – and how much you’re willing to invest to keep improving it.

Assuming an average cost of $6.60 per human-assisted support interaction (ref Deloitte), each AI-resolved conversation can save you roughly $5-$6. If you only reach a 40% resolution rate, which is not easy, but an off-the-shelf solution can get you to 50% or higher, the difference quickly means you are falling short on ROI.

To illustrate this, I’ve worked out savings showing that even if your self-built solution costs you $0 (which is obviously unrealistic), it will still provide worse ROI than a solution with higher resolution rates.

This is the fundamental pitfall: building or picking a tool that falls short of best-in-class will leave money on the table. Your top priority should be to achieve and sustain best-in-class performance.

Underestimating the total cost of ownership

Launching your own AI agent is a forever project, not a one-time build. Spinning up a prototype can be cheap and easy, but a production-grade system demands constant attention from a talented team. Best-in-class is a constantly moving target, and the game is to match or beat it. This means ongoing feature improvements, model upgrades, prompt tuning, as well as the non-trivial, unglamorous work around availability, data privacy, and security.

“There is a lot of software to build, and if you under-invest you’ll quickly become a bottleneck”

Once you get going, you’ll realize that most of the work isn’t in the bits of the system you can see; it’s in all the parts around it that you use to train and control it, how you integrate with other systems, and how you define policies and processes it can reliably follow.

There is a lot of software to build, and if you under-invest you’ll quickly become a bottleneck, holding your company back from higher resolution rates, and once again miss out on higher ROI.

To make the previous example more realistic, I’ve factored in a conservative $50,000/month for an engineering team to own this. To just break even on ROI, you’ll need to match the performance of best-in-class tools you can buy and you’ll need massive volume (approximately one million conversations per month) to offset the ongoing costs of your team.

This makes me think that most people pursuing this path either haven’t figured out the economics or are doing it for other reasons without thinking about the price tag or opportunity cost.

As you drive higher and higher resolution rates, your AI agent will be solving more difficult and time consuming cases, delivering increasingly higher ROI. However, you should anticipate higher engineering costs as the work gets much more difficult too.

Deflection is easy, delight is hard – speed of improvement matters

Automation rate alone isn’t the whole story. If you care about customer experience, not all deflections are equal. Quality matters.

When one of our customers compared Fin to a competitor, they saw that while both handled simple queries at a similar rate, Fin achieved 15 percentage points higher on CSAT.

“The winners in all AI categories will be those who improve the fastest”

This is no accident – it’s the result of hundreds of experiments run over many months against significant volume by dozens of specialized machine learning PhDs and engineers, enabling us to consistently improve both automation rate and customer satisfaction.

You’ll struggle to ever match this with a small investment of a single team and longer experimentation cycles due to lower volume.

The winners in all AI categories will be those who improve the fastest. High scale leads to faster feedback loops on experiments, leading to faster improvements in performance, leading to stronger demand and usage, further compounding the advantage. It’s incredibly hard to compete against this.

While you’re struggling to keep up with high quality support automation, the best-in-class AI agents will be leaning into customer success, generating real value for your business.

Don’t just take my word for it

If you have a strong AI team, massive support volume, and very specialized needs, it might make sense to build your own agent. But even Anthropic – one of the leading AI labs – uses our agent Fin because they recognize the constant iteration required to stay safe, accurate, and deeply integrated with support workflows.

“If you’re debating whether to build your own AI solution or buy one, as a fast-growing company in a complex space, my advice would be to buy – and specifically, buy Fin. Intercom combines safe, reliable, and cutting-edge AI built by Anthropic with decades of customer support expertise and an AI-first approach, which means you can deploy a world-class solution in days or weeks instead of months, with the confidence that your brand and users are in good hands.” – Isabel Larrow, Product Support Operations at Anthropic

Ultimately, Anthropic decided their engineering capacity was better spent improving their core products.

“Invest in what makes you unique”

This logic applies to companies in general: invest in what makes you unique. This is why we leverage vendors like AWS or PlanetScale – not because we don’t have excellent engineers who could self-host systems of our scale, but because it’s undifferentiated heavy lifting. Leaning on great partners enables us to apply greater focus on our primary mission.

Winning in software often looks like small teams generating tens or hundreds of millions in revenue. Cost-saving side quests rarely justify the diversion of engineering talent, especially when an off-the-shelf solution can get you there faster.


I get the appeal of building your own system – it’s fun, it’s a great learning experience, and there’s something special about shipping your own AI code. But my advice? Channel that energy into your own product, not a non-strategic side quest.

If you’re on the fence, I’m always happy to chat about how we approached building our AI agent and the lessons we learned along the way – whether you end up building your own or not.

Fin2 AI Agent blog CTA - Horizontal