The economics of multiplayer AI

Agentic AI has changed what's possible for companies over the past year, and the capabilities we've shipped at Dust reflect that shift. Skills let teams codify repeatable capabilities, wake-ups enable autonomous scheduled work, and frames, memory, and chaining help agents hold context and coordinate more complex workflows for users.

These extra capabilities require more compute, which is why we're moving to credit-based pricing. It ties what users pay to the value they’re getting, lets users understand their consumption, and gives admins and finance teams clearer visibility and control over usage and cost. It's also a better fit for what we're building: the multiplayer, multi-model system where people and agents work together, improve over time, and operate responsibly at scale.

For providers of AI, both the labs developing the models and the systems builders like Dust, credit-based pricing aligns cost with value. This is not a call for companies to “use less AI.” It’s an opportunity to get ahead of structural changes to the cost of labor, in an age where workforces are made up of humans and agents alike. The companies that come out ahead will be those who turn AI costs into real operating leverage whose impact they can measure.  

Breaking down the cost of AI in 2023 was simple: human-triggered conversations via chatbots consumed tokens per message, while API calls consumed tokens programmatically. Today, digital teams of humans and agents coordinate complex work across many tools and data sources, much like cross-functional teams always have. If units of work within firms are increasingly carried out under the direction of agents and humans interchangeably, the only pricing model that can align cost to value creation is a unified consumption-based one. Credits are simply the unit by which Dust measures that consumption.

Tying pricing to consumption also keeps us accountable because consumption is a direct signal of value. As a founder, I see this as a natural alignment that helps us stay relentlessly focused on what actually creates the most value for the teams using Dust. The early feedback from customers we’ve been building this with tells us we’re headed in the right direction, and we’ll keep listening as we go.

In the 2010s, the cloud flipped infrastructure from fixed capex to variable opex. Companies that built real visibility into who was spending what, and could tie usage to value, turned this shift into leverage. AI consumption is on the same trajectory, and the companies that build the same visibility into AI spend that they built into cloud spend will be the ones who turn it into leverage rather than a line item.

Credit pools give teams a more flexible way to plan for AI usage: instead of tying consumption rigidly to individual seats, companies can allocate credits across the workspace, direct heavier usage to where it creates the most value, and give finance teams clearer visibility into spend. Per-seat allocation forces companies to provision for each person's peak usage individually, which means paying for idle capacity most of the time. Pooling lets the workspace provision for the aggregate instead of the individual, the same way a shared support team is cheaper than dedicating a rep to every possible spike in volume. 

Credit pools are becoming a foundational part of Dust's Enterprise plans, where this workspace-level flexibility and granularity matter most. They'll also be available as an option on self-serve Business plans, and workspace admins will be able to manage workspace credit pools that power programmatic usage and any consumption beyond individual seat allowances.

Pooling credits allows companies to more readily adapt to fluctuating usage, but it’s just one part of optimizing credit consumption. Companies will need to invest in AI systems that get more efficient on their own, the more that those systems are used. We believe that self-improvement against bespoke business goals is the real unlock for compounding productivity, and that this is made possible only in a system that is multiplayer, multi-model, and governable at scale.

Multiplayer means facilitating the bidirectional execution of work by both humans and agents, working together simultaneously. Multi-model gives companies the flexibility to choose LLMs from frontier labs, open-source providers, or bring their own. This isn’t just about cost, reliability, and provider lock-in; it’s also about sovereignty of a company over its intelligence and learning loops. In order to govern all of this at scale, large teams need a control layer for optimizing usage and cost, and controlling access to data and tools – for humans and agents alike. 

Together, these pillars create a system that compounds not just task-level productivity, but company intelligence. In addition to features like agent memory that exist in Dust today, we’ve been working on new capabilities for continuous learning and improvement – features like self-improving skills – that will be made available to customers as part of the new credit-based model. 

More to come on that in the coming days. We can’t wait to continue building the future of work with you.