GenAI Is Becoming More Cost-Effective. That’s Not the Hard Part.
There’s a growing narrative that developing generative AI-based solutions is becoming “cheap” and accessible to all. And on the surface, that’s true. Model costs are dropping fast, tooling is improving, and you no longer need a room full of PhDs to build something useful.
At Audere, we’ve been building AI-powered health solutions since before GenAI entered the mainstream. We started with computer vision models that interpret photos of rapid diagnostic tests, and over time we reduced the cost of building those models by more than 10x while dramatically lowering the technical skill required to create them.
What we’re seeing now with GenAI follows the same trajectory: lower costs, broader access, faster experimentation. That part is inevitable.
What’s not inevitable is where the real costs actually sit once you move past demos and into real-world, scaled deployments.
The “Cheap AI” Myth Breaks at Scale
To be clear, we don’t build or host our own GenAI models. We rely on inference APIs from commercial providers like OpenAI, Anthropic, and Together. The space is evolving too fast to lock ourselves into a single model, and we regularly evaluate new options to ensure that we’re using what best fits our use cases…most cost effectively. We also have not reached the point where hosting our own models makes economic sense.
Like most teams innovating with nascent technology, our early GenAI experiments did not worry about cost. Our first consumer-facing prototype – focused on HIV vulnerability and PrEP eligibility – was about determining: does this actually work and will people engage with it?
But once you move from proof-of-concept to field studies and then to national-scale deployment, especially in global health, cost stops being an afterthought…it becomes impossible to ignore.
For GenAI solutions like ours—powering personal AI health and wellness companions—the cost of “base conversations” has fallen nearly 10x since the early days. That’s real progress. We’re even preparing to deploy an updated model that will cut those base conversation costs by another 63%.
But here’s the part that matters: in a system built to scale responsibly, basic conversations aren’t the primary cost driver.
Where Our “GenAI Costs” Actually Come From
In production, based on the use cases that we look to support, our consumer-facing GenAI spend clusters around four main areas:
Base conversations
This is relatively inexpensive. Maintaining conversational context over time and enabling predictable, natural, locally relevant dialogue is something that is expected at this point…it’s table stakes.
End User Engagement
This is where cost - and value - start to rise. Keeping users engaged by extracting structured data from conversations to drive conversational flows tailored to user needs, supporting multiple languages (especially low-resource ones which often require fine-tuning models), delivering the right content at the right moment, and “nudging” patients toward desired health behaviours, all require repeated, intentional GenAI use. Without this, many AI solutions quietly fail due to lack of perceived value leading to low sustained use.
Continuity of care
If your system never hands off to a human—or does so without context—it’s not connected health, it’s just another data silo. Generating summaries, surfacing risks, and integrating into clinical workflows adds cost, but it’s foundational.
Safety
This is the most expensive category, and it should be. Periodic evaluations and real-time monitoring for clinical appropriateness, local relevance, empathy, and harm is not optional in consumer-facing health AI. These guardrails, with clearly defined escalation pathways to human support, are what bring predictability, reliability, and accountability to the overall GenAI system.
In one of our more fully-featured deployments, engagement and safety together account for over 75% of GenAI costs. Base conversation handling is a minority of spend. Continuity of care, while extremely important, barely registers by comparison.
That cost distribution isn’t a failure of optimization—we are continually optimizing, leveraging the latest techniques and efficient technologies. It is a reflection of priorities. It is not a mistake that in that previously mentioned deployment, we are seeing 55% of connected users meaningfully engaging with the solution, with over 25% re-engaging with the system after 30 days.
You Can Cut Costs. Just Don’t Cut Value.
Our average GenAI costs per user are already low by most standards, and we are always looking to reduce them further. It would be easy to cut costs by:
Reducing real-time safety checks
Stripping out engagement features
Skipping structured data extraction
Plenty of systems in the global health space do exactly that. They look affordable on paper and struggle—or fail—in practice.
The harder—and more meaningful—work is designing systems that evaluate and use GenAI intentionally: spending where it improves trust, outcomes, and long-term use, and ruthlessly optimizing everywhere else.
That’s where we’ve focused our effort: improving efficiency, continuously re-evaluating models, and reducing cost through engineering investments without compromising things - quality, engagement, and safety - that actually matter.
What are We Willing to Pay for?
GenAI will keep getting cheaper. That’s a given.
The real question is: what are we willing to pay for?
Engagement? Safety? Trust? Continuity of care? Outcomes?
Ultimately, any consumer-facing, health-focused GenAI solution has to deliver a return on investment. The strongest ROI does not come from minimizing spend – it comes from intentional spend:
Investment in user value creation and engagement drives utilization and retention
Investment in safety reduces operational, legal, and reputational risk
Investment in continuity of care turns AI interactions into efficient, effective care with measurable outcomes
Those costs show up early in solution development. The returns show up later – in scale, efficiency, and avoided downside.
So, yes, with today’s tools and lowered barriers to entry, it's easy to build something that looks impressive. Building something people can actually rely on–something that delivers real outcomes at scale–is a very different challenge. That’s where the real cost of building safe, trustworthy, and sustainable GenAI solutions lies.
