How startups are capturing more value with AI
October 2026
Deepthi Madhava
I recently attended Madrona’s IA40 (intelligent applications) summit where the theme was around harnessing the power of AI. A key question at the summit was where does value accrue in software as models keep getting smarter?
One answer from the stage was clear: AI systems are more powerful than models.
One panel put it well: models are smart, but “corporate stupid” until they are grounded in the context of how an enterprise actually works. Other panels made a broader point: value can accrue at many layers of the technology stack, depending on what a company uniquely owns.
My point of view on the value question is that, while too early to form a definitive answer, I believe there is value to be delivered and captured at all levels of the tech stack. Whether you provide bits or chips, there are still opportunities to build defensible moats and thus value can accrue at nearly every layer of the tech stack. Specific to software, I agree with the summit’s takeaways that the greatest value sits in harnessing models to deliver outcomes enterprises and users actually want.
So, what is a harness?
I like Anthropic CTO Rahul Patil’s simple framing: the harness is the gap between what models can do and what we want them to do. That harness can include context, memory, integrations, workflows, security, observability, permissions, human judgment, and anything else required to turn model intelligence into a reliable and usable outcomes.
Models will keep getting smarter. But as they do, our ambitions rise with them. We ask them to solve harder problems, operate with more autonomy, and deliver increasingly complex outcomes.
In AI phrasing, the gap moves. It doesn’t disappear.
I think the best harness companies are going after highly ambitious problems - problems where the desired outcome is still far ahead of what the model alone can reliably deliver. If the problem is modest, models will catch up quickly and the harness risks becoming just a feature.
The best companies aren’t building for “least common denominator” model optionality, if that means missing new capabilities. They ride each model leap toward a bigger outcome. They don’t hedge against model progress; they benefit from it.
Ultimately, value will be measured in outcomes. Are you delivering the outcome more accurately, faster, and cheaper?
If you sell outcomes, your pricing and business model need to follow.
Three examples from the summit stood out for me:
The end of per-user pricing. Seat pricing alone can produce negative gross margins for AI products. Some form of consumption or outcomes-based pricing becomes necessary.
Sell the work, not the tool. Crosby AI sells legal work at a fixed price per task. Its margin is driven by how little human time it takes to complete each task.
Go after services budgets, not just software budgets. Gradial, an agentic marketing operations solution, targets the $313 billion spent annually by Fortune 500 firms on outsourced marketing. Its pricing is based partly on measurable marketing results and is not simply a CMS software line item expense.
This leads to another conclusion I have been considering: an outcome-focused AI business will often have a services component. And that may be a feature, not a flaw.
Services here could mean onboarding, implementation, exception handling, or a human in the loop where judgment still matters.
The important distinction from a traditional services business is that the underlying technology makes the work increasingly scalable. Each task should require less human effort over time. Models improve. The harness improves. More exceptions get automated. Margins expand.
Skeptic might counter that “services” are less scalable and mean a lower valuation multiple.
I’m not sure that framework works anymore. With the opportunity gap between AI model capability and AI adoption in enterprise and SMBs, we are already seeing this paradigm shifting.
When technology continually shrinks the human share of the work, the services component can actually strengthen the moat. It keeps the company close to the customer, exposes the hardest edge cases, and generates the context and engagement data needed to make the next outcome better and cheaper to deliver.
So, for every AI software company I meet, I increasingly find myself asking two questions:
Is the ambition big enough that the model alone cannot deliver the desired outcome?
Does the business become more scalable as it chases that gap?
If the answer to both is yes, that gap may be exactly where the value accrues.