Enterprise AI projects increasingly reach production when vendors prove value on live workloads, provide ongoing testing and iteration, and expose ROI. Successful deployments start with decomposable workflows that ship quickly and expand, rather than broad transformations with undefined success criteria.

The Agent Graveyard Isn’t Real Anymore

I’ve noticed a big shift in how enterprises are buying AI this year. A LOT more projects are making it into production, but only a specific type.

Last year, MIT published a popular report stating that roughly 95% of generative AI pilots have no measurable financial impact.

It’s safe to say that this is no longer the case. Agents have become more effective as models have improved and buying patterns have matured.

Some interesting stats I read from Menlo Ventures’ survey of 500 US Enterprise buyers: 47% of AI deals go to production, compared to 25% for traditional SaaS. Additionally, 76% of enterprise AI use cases are now purchased rather than built internally, up sharply from a nearly even split in 2024.

In 2024, when the models were still new, everything was about experimentation. Leadership teams had pressure to just try things, so all sorts of pilots were being spun up, both with internal builds and with external partners.

Today, when we have the first conversation, it’s clear that the execs need no convincing that our use case is something they will invest in. It’s more about convincing them our approach is the best fit for them.

So what separates use cases that make it vs those that don’t? Why do companies sometimes decide to build in-house?

Why ‘AI converts better than SaaS’ actually makes sense

Traditional SaaS sold capability. You bought a CRM and then spent nine months and a systems integrator getting your team to use it correctly, and the value showed up eventually in ways your CFO mostly took on faith. The gap between signature and value was long, and a lot of deals died in that gap.

AI sells output. When a deployment works, the value is visible in the same reporting period you signed the contract, often in the same week. AI delivers enough immediate value to short-circuit standard procurement. That’s not buyer enthusiasm. That’s the sales cycle compressing because the proof arrives faster than the process designed to demand it.

In our experience, this is also why buyers in AI never decide off a demo. Every AI demo looks good now, which means demos have become completely useless as a buying signal.. So evaluations have moved into the product itself, and often into production: shadow mode against live traffic, real historical conversations, a slice of actual volume with real customers on the other end. Buyers need to feel it, not watch it. That raises the bar enormously on the vendor side, but it’s exactly why conversion is high. By the time someone reaches a contract, they’ve already seen the thing run in the environment it has to survive in. The evaluation isn’t a preamble to the proof. It is the proof.

Build vs Buy

Part of the shift toward buying is about expertise; internal teams often aren’t as AI-heavy as they need to be, and vendors are living in this problem full time. But honestly, that’s a small part of it.

The real reason is that most of these use cases aren’t worth building from zero in-house, and the reason isn’t the build. Take our category: a competent engineering team can absolutely put together a solid AI customer service agent. That’s not the hard part anymore. The hard part is everything after: the constant iteration as policies and products change, the analytics across millions of conversations to find where the agent is quietly failing, the simulation and regression testing before every change ships, the QA on edge cases nobody anticipated. That work never ends, and it doesn’t scale down.

Putting your engineers permanently on the hook for that is a bad trade. It isn’t their comparative advantage, and in practice it’s the work that gets deprioritized the moment a real product deadline shows up, which is exactly when an unmaintained agent starts degrading in front of customers.

And build vs buy isn’t really black and white with AI agents anymore, because products now let you do far more yourself. Buying no longer means giving up control of behavior; it means buying the iteration layer (the tooling, testing, analytics, etc) while your business teams keep their hands on the wheel and ship changes without filing an engineering ticket. That’s why enterprises are partnering more, and much faster than they did with previous software categories.

So why do some projects still never make it?

That said, I have also observed many reasonable-sounding use cases just struggle to get traction in the market.

1. The ROI was never crisply defined.

Some AI value is genuinely hard to attribute. If an AI agent helps with a workflow, or makes a team more efficient, how does that reflect in the bottom line? If that’s hard to define, it will be very difficult for a leadership team to prioritize the use case, much less commit to a long-term partnership.

The nice part of customer service AI, for example, is that the ROI is very easy to reason. As the AI does a better and better job, it’s both driving very clear operational efficiency and elevating the customer experience.

As it expands into more proactive conversations, that ROI becomes revenue-generating and not just cost-saving, which makes the argument even more powerful.

2. The problem couldn’t be broken into pieces.

This is the underrated one, and the single best predictor of whether a deployment makes it.

If your initiative is “transform our operations,” there’s no version of that you can ship in six weeks, measure, and expand. You get an eighteen-month program, a steering committee, a change in sponsor halfway through, and a quiet cancellation.

All-in-one use cases tend to be very hard because too much has to go right for the business to even see a shred of ROI.

Customer service has an unusually good property here: it decomposes almost perfectly. You can start with just one user journey, prove that out, go live in production, and expand from there.

The reframe

The agent graveyard was never evidence that enterprise AI doesn’t work. It was evidence that in 2024 a lot of organizations attempted indivisible transformations with undefined success criteria — which has always failed, with every technology, in every decade.

What survives won’t be the most sophisticated deployments. It’ll be the ones that started small enough to ship and had a number written down before the vendor was selected.

If yours doesn’t, you’re still in the 2024 cohort.