This past week, both a subscriber and a Pricing Director at a legal/tax compliance SaaS asked me a version of this same question.
It’s an interesting and timely question with an embedded assumption we should unpack explicitly before answering.
The assumption:
- LLM-backed software compounds in quality with each use.
- Providing a more consistent, more reliable, higher-quality outcome.
- Thus it becomes exponentially more valuable to the customer.
Two problems with this assumption:
First, product improvement isn’t linear or exponential. Partially because the vendor’s definition of improved quality and the customer’s definition have always been two different things. This is why bug reports and backlogs exist. Additionally, all third-party dependencies (like external models), create a supply chain risk; a model running on someone else’s infrastructure, subject to someone else’s pricing decisions, business strategy, and uptime struggles.
These dependencies threaten customers’ value realization. Long term the dependency will be addressed, whether becoming a harness to the LLM our customers’ have committed to, maintaining the LLM on our own infra, or understanding the customer value enough to transition to a cheaper, more deterministic, solution.
Second, and more fundamentally: As Bob Moesta so concisely put it, “Demand is Solution Angnostic”. They hired the software to help them with a job, and they don’t particularly care about the details. At best, customers are indifferent to the underlying technology and most don’t want AI getting in the way. In the same way they don’t value which frontend framework or database was used. At worst they’re actively disappointed and frustrated by the intrusion.
The compliance director doesn’t want an nondeterminstic summarization of audit findings; she wants increased confidence in the audit. The field inspector doesn’t want AI-generated inspection notes; he wants confidence nothing was missed.
A Customer Time to Value Timeline
This leads to a more useful framing: Customer-perceived value peaks at the moment of switch, then decays as the product becomes normality.
Two curves drive this:
- Perceived value declines as familiarity increases. The initial 10x improvement becomes the new baseline.
- Switching cost rises as the product embeds into workflows, data, and habits.
Day 1: The product is a clear step change from the prior solution. Time saved and friction removed are obvious if not fully realized.
Day 60: The product is the new normal. Expectations rise. The prior solution is forgotten.
Day 180: The product is embedded. It feels less differentiated, but harder to replace. Newer users only know this solution.
Will the software get significantly better with use – independent of the users’ increased familiarity? Maybe. AI or no AI, this is the SaaS product manager’s job.
Any change to the product post-purchase, whether self-improvement, bug fixes, or new features, are all about retention and holding off the decay curve.
So how do you price something customers will eventually fully normalized and metabolized?
Anchor to Day 1 customer economic value, not to hypothetical future improvement.
If the product delivers a 10x improvement at the point of switch, price against the marginal difference. The fact the product is early, imperfect, or still evolving is secondary to the realized economic impact. If the value were not real, the customer would not have switched.
But pricing high is not sufficient on its own. It increases scrutiny. It raises the burden of proof. It slows decisions in complex buying environments.
High pricing only works when paired with fast onboarding and white-glove de-risking.
This is not just an operational choice. It protects the perceived 10x before normalization sets in.
Low pricing with self-serve onboarding does the opposite. It signals replaceability, reduces commitment, and accelerates normalization before value is fully realized.
Returning to the principle of pricing the Customer’s Most Valuable Noun, there are only 3 inter-related customer improvements we can price against:
– Speed (marginally less time taken to produce a noun of any quality)
– Volume (marginally more nouns produced of any quality)
– Quality (all nouns produced with marginally fewer defects)
LLMs are speed play, not a quality or volume play. Treating LLMs as inherently improving quality or volume overstates their reliability. However, just like same-day shipping, dramatically shortening time-to-value is worth a premium.