Illustrative Case Study — Scaling an AI product without giving away the margin
- Client profile: $180M ARR enterprise workflow software company
- Customer base: Approximately 1,200 mid-market and enterprise accounts
- Business model: Subscription software with newly introduced AI capabilities
- Core challenge: AI adoption was accelerating while infrastructure costs and unclear pricing were putting pressure on product economics
- Engagement: 9 weeks from economic diagnosis through architecture, packaging, and controlled customer pilot
- Primary objective: Scale AI adoption while creating an economic model that could remain attractive as usage increased
- Decision at stake: Continue bundling AI broadly, restrict usage, raise prices, redesign the architecture—or combine several of those approaches
- Product and financial data foundation — Product telemetry, customer accounts, subscription data, cloud spend, model usage, support activity, and renewal information were connected so product behavior could be viewed alongside its economics.
- AI unit-economics model — Usage was translated into cost at the customer, workflow, and feature level, giving leadership visibility into which AI experiences created customer value and which were disproportionately expensive to deliver.
- Model-routing architecture — Requests were separated by complexity so simpler work could use lower-cost models while more demanding tasks could still access higher-capability models when the business value justified it.
- Retrieval, caching and context optimization — Repeated information was reused where appropriate and unnecessary model context was reduced to lower inference cost without simply degrading the user experience.
- Pricing and packaging simulation — Bundled AI, usage allowances, premium tiers, overages, and enterprise packages were modeled across different adoption and usage scenarios before changing live customer pricing.
- Customer-value analysis — Product and customer-success teams identified the workflows customers valued most so optimization did not accidentally remove the capabilities driving adoption.
- Performance monitoring — Leadership could track AI usage, cost per workflow, margin, customer adoption, latency, and quality together throughout the pilot.
The problem
The company's AI launch looked successful.
Customers were using the new capabilities. Sales wanted AI included in more deals. Product leadership wanted to accelerate the roadmap.
Finance was seeing a different picture.
Traditional software economics had been relatively predictable. Once the product was built, additional customers could often be served at attractive incremental margins.
AI changed that equation.
Every additional workflow introduced compute and model costs. The customers receiving the most value were often also generating the highest infrastructure expense, while pricing had not meaningfully changed.
Leadership was facing an uncomfortable possibility:
The more successful the AI product became, the worse its economics could become.
What we found
We started by connecting product usage to customer economics.
The headline adoption number was hiding several very different behaviors.
A relatively small group of accounts generated a disproportionate share of AI cost. Some workflows created significant customer value but were running on expensive models even when simpler infrastructure could complete the task.
Other features performed well in demonstrations but saw limited repeat use.
Pricing created a second problem.
Heavy and light users were effectively paying for AI the same way, which meant the company was absorbing growing usage costs without consistently capturing more value.
Engineering, product, finance, sales, and customer success were each seeing a different piece of the problem.
We created one view across all of them.
What we changed
We worked on both sides of the equation at once:
the cost of delivering AI and the value customers were willing to pay for it.
On the product side, requests were classified by complexity.
High-value work that genuinely benefited from more capable models continued to use them. Simpler requests were routed to more economical models. Retrieval and context were redesigned to avoid repeatedly processing unnecessary information, and common outputs were reused where appropriate.
But lower infrastructure cost alone would not solve the business problem.
We also modeled several commercial approaches:
- AI bundled into existing plans
- Defined usage allowances
- Premium AI tiers
- Additional usage pricing
- Enterprise commitments
- Feature-specific packaging
Product leadership provided judgment around what differentiated the platform.
Sales and customer-success teams provided judgment around what customers understood, valued, and would realistically purchase.
Finance established the economics the product needed to meet.
The modeling allowed leadership to see those decisions together instead of optimizing one department at a time.
The decision
Leadership chose a hybrid approach.
Core AI capabilities remained broadly available enough to encourage adoption and demonstrate value.
Higher-cost, higher-value workflows moved behind defined usage allowances and premium packaging.
At the same time, the underlying architecture was redesigned so the company was not paying premium model costs for every task simply because it could.
The changes were introduced to a controlled group of accounts before wider rollout.
The result
The modeled cost per completed AI workflow declined approximately 27%.
Gross margin on the AI-enabled product increased from approximately 61% to 69%, an improvement of eight percentage points.
At projected usage levels, the combination of architecture changes and revised monetization represented approximately $2.6M in annualized margin opportunity.
Customer usage remained strong during the modeled pilot because the work was not approached as a simple cost-cutting exercise.
The company protected the experiences customers valued while changing how those experiences were delivered and monetized.
The original question had been:
“How do we get more customers using AI?”
The better question became:
“How do we create more customer value every time AI usage grows?”
That changed both the product roadmap and the investment decision.
