Vertical AI’s Moment of Truth

If 2023 and 2024 were the years of speculative AI pilots, 2025 is looking more like a stress test. Across industries, vertical AI startups (those offering domain-specific copilots for finance, law, healthcare, and beyond) are bumping into their first real market constraints. It’s no longer enough to dazzle with demos and shout “copilot” at the top of your lungs. Buyers want proof. And they’re asking tougher questions: Is this worth it? Can we do it ourselves? What exactly are we paying for?

The premise, to be clear, still has legs. Embedding a large language model into a sector-specific workflow can definitely automate drudgery and unearth insights in ways that look like magic—at least on a good, non-hallucinatory day. Companies like Hebbia (for finance) and Harvey (for law) rode this wave early, raising large rounds and landing marquee clients. But now, they’re entering a different phase. It’s not adolescence, but something closer to market audit: where bold ideas meet budget reality.

Sticker Shock: When SaaS Starts to Look Like Sun Valley

Start with Hebbia. Their Matrix product, pitched as “the AI platform for finance,” is purportedly priced somewhere in Bloomberg Terminal territory: $20–25k per user, per year. That’s a material commitment, and one that implies serious ROI. Hebbia positions itself as a mission-critical tool for analysts and researchers: search across filings, extract key data from hundreds of PDFs, and do in seconds what used to take hours. And to be fair, some customers say it delivers exactly that.

Over in the legal sector, Harvey has taken a slightly more restrained approach. It reportedly sells for around $1K–$1.2K per user per year and has been adopted by major law firms like Allen & Overy. The pitch is similar: automate research, drafting, and contract review—safely, securely, and without the hallucinations of general-purpose models. The early headlines were glowing, with time savings in the 50–80% range and partners declaring it a “game-changer.” But in private, many midsize firms have started quietly asking if it’s overkill.

The ROI Is Real…Until It Isn’t

Some buyers are thrilled. Reports abound of lawyers and analysts saving hours per week, shaving days off review cycles, and reducing errors. In industries where time is literally money, those gains add up fast. At top-tier firms billing $300–$500/hour, even modest productivity bumps can justify a five-figure annual tool.

But not all feedback is glowing. A growing number of users are asking why these tools should cost so much when general-purpose models like GPT-4—now faster, cheaper, and more capable—can do much of the same with a bit of smart prompting. One private equity user summed it up bluntly: “Didn’t add value relative to baseline GPT-4.” That’s not exactly the testimonial you put on your homepage.

Commoditization, Platform Encroachment, and the Race to the Middle

Vertical AI’s core problem is simple: the floor is rising. Foundation model costs are dropping fast, while capabilities are improving (see my last post). GPT-4 can now handle long context documents. Anthropic and open-source challengers are close behind. Tools once seen as specialized are now starting to feel, well, standard. I think this will get even more challenging with the arrival of GPT-5.

And then there’s the platform threat. Microsoft is baking AI into Office. Bloomberg has its own finance-tuned model. Thomson Reuters acquired Casetext. When your customer’s existing vendor starts offering “AI features” for free (or bundled), your standalone product starts looking more like an optional add-on…and a pricey one at that.

What Comes Next: Adapt or Get Absorbed

The path forward requires sharper differentiation. If the model is no longer your moat, then the product experience has to be. That could mean tighter integrations, proprietary data access, stronger support, better UX, or simply more pricing flexibility. We’re seeing some companies running hard at consolidating 3rd party data providers to extend the platform’s value and moat. That makes a lot of sense.

As you’ll see later this year, Drift is taking a more restrained approach on pricing. Expect to see vertical AI companies adjust with usage-based plans, lower-cost tiers, and more modular offerings. Some will double down on specialization. Others will look for strategic partnerships or acquisitions. Casetext didn’t beat Westlaw, it joined it. Don’t be surprised if more follow.

Bottom Line

Vertical AI isn’t going away. But the gold rush phase could be ending. The companies that succeed from here will be the ones that adapt—not by shouting louder, but by proving real, sustained value. I expect prices to drop, and not to stay at the same levels we saw at launch.

The days of vague claims and wide-eyed pilots are fading. What comes next is execution, refinement, and a clear answer to the only question that matters: why should I keep paying for this?