Tokenomics: The True Cost Lands Only After the Pilot
There’s a conversation I’m having with firms much more often recently, and it goes something like this…
A firm runs a pilot. It goes well. They roll it out to a wider group of users as a new AI tool. The results continue to look genuinely good, and everyone’s pleased. Then the tool is scaled up and given to more people start using it. It gets embedded into day-to-day process. Usage climbs from a few dozen tokens a week to hundreds or even thousands a day.
And then the invoice arrives or more common, the charge leaves the bank. Someone realises the cost has gone from £100 to £500 for the month. Next, someone does the calculation of what this actually cost per client, and how this has impacted your margin.
This isn’t a hypothetical scenario. It’s becoming one of the most predictable patterns I see in businesses moving from pilot to production with their AI tools, and it’s catching out a lot of otherwise sophisticated, well-run firms and businesses. Pilots can be very cheap almost by accident, but scale definitely isn’t if not managed intentionally. The economics of AI don’t move gradually as usage grows, they can shift suddenly, and most aren’t doing the work to understand why before they’re already exposed.
Why the sums can change without you noticing
A pilot, by its nature, is small, with just a few users, a defined use case, and a manageable number of interactions and data. The cost is easy to justify and usually amounts to a line item nobody notices.
Production is a different animal entirely. More users means more queries. More queries means more AI model calls. Agents that work autonomously might make several calls per task rather than one, retrieving data, checking it, reasoning over it, calling other systems, before producing an output. Multiply that by hundreds of tasks a day, and the token consumption, and therefore the cost, can grow in a way that has very little relationship to how the pilot behaved or was intended.
Nobody’s being careless, it’s just that the pilot likely never unearthed the how the full production system operates. And the question you don’t know the answer to until you’ve tested at scale is what’ll actual cost per interaction and in total? Those costs might not scale linearly with usage, they may compound. Some AI agents might need to be genuinely more expensive because they’re doing more valuable work, but you need to avoid building the ones which are more expensive because they weren’t optimised when they were being built.
The majority find out this out only at the last minute when the costs they haven’t planned for get charged.
Cheap isn’t always cheap, and expensive isn’t always wasteful
There’s a bit of a trap hiding inside the AI tokenomics quandary, that’s important to explore, because it can lead you towards pursuing the wrong outcome.
The natural instinct, based on what we’ve explained above, is that once costs start climbing, is to scale back and minimise your AI spend. So, you’d typically cut back the rollout and usage, restrict access, and maybe even revert back to manual or alternative process to control cost. Whilst that might be the right decision, the right question to answer first is what value is being created in relation to what it costs – what is the ROI you’re getting and does it meet your expectations or targets.
The reason is that a more expensive AI solution could actually be incredible value if it produces a significantly better outcome, or frees up genuinely skilled time for higher-value work. A low-value interaction, repeated a million times, could still be wasteful if it’s not actually producing anything valuable. Cost is unfortunately often used solely for making decisions on automation, whereas the true value is often significantly higher, albeit a little trickier to calculate.
One thing I would say is that “cheaper” is rarely the right frame for AI at all. One topic of conversation that I’m having more often now is to correct the assumption that using AI in your firm should mean cheaper or discounted outcomes – on the basis that if your firm is using AI to do work, the fee should fall. That might be the possible direction of travel being used by some leading software vendors of late (naming no names), but that’s potentially very backwards. Good AI gives skilled teams more capacity for the judgement-heavy work that actually justifies the fee. The value goes up, not down. Treating AI purely as a cost-cutting lever, rather than a capacity and quality lever, is how we end up optimising for the wrong outcome entirely.
This needs a finance conversation, not an IT one
Here’s where AI Tokenomics starts to overlap with much of what we covered in the first two parts of this series. Governance is about knowing what your AI can access. Literacy is about knowing what your people can do with it. Tokenomics is about knowing what it actually costs to run, and crucially, getting that understanding in place before scale forces the question, not after.
That means that the finance decision-maker needs to involved far earlier than they are. There are some very fundamental architecture decisions that need to be scoped and costed, including which model is used (they carry different token rates), how an agent is built, how many input and output calls a task genuinely requires. These factors all have direct cost implications that will compound rapidly at scale. A business that treats those as purely technical decisions without financial input, runs the risk of eventually finding next month’s AI charges have skyrocketed.
This is exactly the same discipline I’ve watched mature in enterprise automation delivery over the past few years. The best-run automation programs always had a clear investment appraisal model with cost per transaction / action and a clear line of sight on what happens to that cost as volume scales up. AI makes this discipline more urgent, because the cost variables are less familiar and the scaling curve can be steeper.
Drowning in manual, repetitive work? Tell us the task and we’ll show you what to automate.
Forecast before you need to
None of this means AI Tokenomics are entirely unpredictable, or that businesses should be nervous about scaling. It means the forecasting has to happen deliberately, and early, rather than being discovered by accident once the invoice lands.
That means understanding your actual cost drivers: which agents or tools are token-hungry and why, whether that hunger correlates with the value being delivered, and what happens to the model when usage doubles or triples. It means having a real vendor conversation about pricing at scale, not just at pilot volume. And it means building the same kind of cost forecasting discipline into AI that any mature business already applies to cloud infrastructure, headcount planning, or any other cost that scales with growth.
The businesses that get this right won’t be the ones spending the least on AI. They’ll be the ones who can explain, with confidence, exactly why they’re spending what they’re spending, and what it’s buying them.
Practical starting points
In each part of this series, I’ll leave you with some practical next steps and some links to helpful tools to get you started, so here’s your AI Tokenomics imperatives.
Get a real cost-per-interaction baseline, not a pilot estimate. Understand what a single interaction with each AI tool or agent actually costs today, and don’t assume the pilot number holds once usage scales.
Model what happens at 10x volume, before you’re at 10x volume. Cost doesn’t always scale linearly. Find out whether yours does while it’s still a modelling exercise, not a live problem.
Separate expensive-but-valuable from expensive-and-wasteful. Not every high-cost agent is a problem, and not every cheap one is efficient. Judge each against the outcome it produces, not the invoice line alone. Remember, not every automation needs AI.
Bring finance into architecture decisions early. Model selection and agent design have direct cost consequences. That’s a conversation finance should be part of before the build, not after the bill. If you’re hiring “AI specialists” to take the lead, be cautious as this one can really get out of hand quick.
Stop treating “AI-enabled” as code for “cheaper”. Position clearly the value that AI creates, not just a discount. Cheaper is rarely better, and it trains clients and stakeholders to expect the wrong thing from good work.
Know where you stand
Most businesses can’t tell what automation and AI outcomes are costing or returning to them, or their clients.
My AI Tokenomics Assessment gives you a baseline to determine that: a structured framework for how to costs are generated, driven and controlled.
Contact us to take the AI Tokenomics assessment tool for a test drive.
Next in this series: Part 5, AI Lifecycle & Agent Structure, and why agents don’t retire themselves.
Daniel Lawrence is the CEO and co-founder of bots for that, creators of the Automation Operating System (AOS) and Agent Console. He has spent more than a decade deploying enterprise automation and AI in regulated industries including accounting and professional services. AI: The Serious Phase, is a six-part series exploring what AI-native operations look like for mid-tier and large UK accounting firms.
