Token Burn Doesn’t Deliver ROI For The Enterprise. Digital Workers Do, at 15x.

TL;DR

Usage is not value. Prompting at random and spinning up vibe-coded agents produces a bill nobody can forecast and output nobody can price.

$15 back for every $1 of tokens and cloud. We measured the Digital Workers we run in production in life sciences. The return is real, realised value, not a projection.

A Digital Worker is engineered, not prompted. It is scoped to a real workflow and built so every step has a known job and a known cost.

The economics get better with scale, not worse. Infrastructure is largely fixed, so every additional run makes the return bigger.

Adoption is the lever, not model price. The gap between a weak return and a great one is whether the business actually uses the worker.

Our Argument

Every AI business case right now is fighting the same fight. Boards want proof. Finance wants a number. Most teams would be happy to show 2x and call it a win.

The first question we get from every buyer is the same: what will the tokens cost? Behind it sits a fear that agentic AI is a meter running in the background, and that the bill grows faster than the value.

That fear is earned. Earlier this year, many enterprises measured AI adoption by token consumption. Some burned through their annual AI budgets in a matter of months. The bill arrived. The return did not.

The cause is how most enterprise AI gets built. Hand people a model, a budget and an instruction to use more of it, and you get vibe coding at enterprise scale: open-ended prompts, agents that loop until something looks right, consumption with no ceiling and output nobody can price. Research on coding agents found token use on the same task can vary thirtyfold, and more tokens do not reliably mean a better answer. Vendors paid by the token have little reason to tell you to use fewer of them.

Digital Workers are built differently. So we measured what that is worthit.

What we found

We studied the Digital Workers we have deployed in production for a global life sciences organisation, running real workflows for real teams. We took every dollar spent on LLM tokens and cloud compute, and set it against the value of the labour they do.

For every $1 of tokens and compute, the business got $15 back.

To be precise about what this measures: it excludes the causaLens subscription and what the customer pays for the Digital Workers themselves. Think of a new hire. This is not their salary. It is their electricity bill. And it includes the tokens burned while we built the workers, not just while they ran.

Tokens are not the expensive part

The token bill is the number everyone asks about, and it is the smallest part of the cost. For most of the Digital Workers we studied, cloud infrastructure cost more than LLM spend. Two of them made no LLM calls at all, because causal models did the work more reliably and for almost nothing.

That changes the economics. Infrastructure is largely fixed once a worker runs at production scale. Double the volume and the cost barely moves. The return does not decay with use. It compounds.

Ask a better question

Stop asking what the tokens cost. Ask what every dollar of compute returns, and what it takes to get the business using the work. Optimising a model's price per million tokens will move your return by a few percent. Getting a Digital Worker into the daily workflow moves it by an order of magnitude.

Most enterprise AI cannot answer that question, because nobody built it to. We build every Digital Worker so the answer is obvious.

At causaLens, that is how we measure every Digital Worker we deploy: by the outcome/ROI, not the meter.