AI spend has become real money, but few can currently connect spend to returns
Written by Martin Hulbert, Chief Technology Officer, Ignite Technology
“We are spending real money on AI. What are we actually getting back for it, and how do we prove that to the board?”
It’s a question more leadership teams are asking themselves right now, and the honest answer for most of them is that they cannot answer it with a number.
The gap between spend and proof
AI spend has become real money. API bills, model usage, agent orchestration, the infrastructure to support it. It shows up on the P&L now, not as a pilot budget line but as ongoing operating cost. Very few organisations can currently connect that spend to a return in a way that would satisfy a finance director, let alone a board.
“The real issue is measurement.”
Ask most technology leaders what a specific AI initiative cost, and they can usually point to the model bill. Ask them what it would have cost to do that same piece of work with people, and the answer gets vague fast. Ask them whether it cost less or more than the last comparable initiative, and there usually isn’t an answer at all.
The real issue is measurement. Most organisations do not yet have the instrumentation to answer the question, and that gap sits underneath every “prove it to the board” conversation happening in Exco rooms right now.
Why is this landing on the board’s desk now?
Three shifts are pushing this question up the agenda at the same time.
AI spend is moving from fixed licence costs to variable, usage-based billing. That behaves the way cloud cost did a decade ago, unpredictable and hard to forecast until an organisation builds the discipline to track it properly.
Boards are already asking finance and operations leaders to justify AI investment in ROI terms, often well ahead of the organisation having the instrumentation in place to answer that question with evidence.
“AI spend risks becoming exactly the kind of unaccountable
cost centre that CFOs spent the last decade fixing.”
Without a way to tie usage back to outcomes, AI spend risks becoming exactly the kind of unaccountable cost centre that CFOs spent the last decade fixing in cloud and SaaS spend.
None of that requires knowing the technical detail of how a model works. It is a governance and accountability problem, and any CFO or COO would recognise it immediately once it is put to them in those terms rather than in vendor language.
Why people costs and token costs sit apart
Two different costs, tracked in two different places. Finance tracks people costs, headcount, salary, overhead, usually allocated by team or cost centre. Technology tracks token spend, if it tracks it at all, usually as a line on a cloud bill, sometimes broken down by model, rarely broken down by task or outcome.
Nobody is putting those two numbers next to each other at the level that would actually answer that opening question, which is the individual piece of work.
What did it cost, end to end, to have a person do this task versus have AI do it? A calculation done once for a pilot business case six months ago and never revisited will not answer that. It takes a number that stays current as model pricing changes, as usage patterns shift, and as headcount costs change.
Without that number, every claim about AI paying for itself is an assertion rather than a proof point. That is exactly the gap the opening question is pointing at.
What does “Tokenomics” actually mean?
Tokenomics is the discipline of tracking what an organisation’s AI and LLM usage costs against the business value that usage produces, measured at the level of unit cost rather than a single monthly total.
The word is worth being precise about, because it carries baggage. Most people who come across it will assume it refers to cryptocurrency token economics. It doesn’t, not in this context.
Here, a token is a billable unit of AI usage, and tokenomics is about connecting that unit of usage to what the organisation got back for it, time saved, cost avoided, revenue enabled, in the same way FinOps connected cloud usage to business value a decade ago.
“A token is a billable unit of AI usage, and tokenomics is about
connecting that unit of usage to what the organisation got back for it.”
The term made it into the mainstream in August 2026, when Broadcom joined the Tokenomics Foundation as a founding member and launched a product built around exactly this problem. That is worth paying attention to on its own terms. When a vendor at Broadcom’s scale backs a standard for connecting AI usage to business value, it confirms this has moved from a niche engineering concern to a board-level one.
I’d rather not need the word, but there isn’t a better one yet for what this actually requires. In practice, it means tracking token consumption the same way organisations already track headcount cost. Unit cost, not just totals. Cost per transaction, per case, per document, per decision, not just a monthly total on an invoice.
And an understanding of how that cost moves with volume, with model choice, with how a prompt or workflow is designed, and with failure and retry rates, because all of those change the number.
Get that in place and the opening question stops being a difficult one to answer. It becomes a number you can pull up on request.
Why does this need to sit in a Portfolio Management tool?
A technology monitoring dashboard cannot answer this. A dashboard showing token spend over time tells you what happened. It does not tell you whether a task should move from a person to AI, whether a task should move back because retry costs have made it the more expensive option, or whether this project cost more or less than the last comparable one.
Answering that opening question needs that decision sitting next to the rest of the work portfolio, with the same visibility and governance already applied to people and projects.
That is a PPM or SPM function. Work needs to be tracked at the task or workflow level, not just team or system level, so cost attaches to a unit of output. People cost and token cost need to sit in the same model, so they can be compared directly rather than reported separately to different audiences.
“People cost and token cost need to sit in the same model,
so they can be compared directly.”
Every project needs its actual costs captured on exit, not just planned at the start, so it becomes a data point the next initiative gets measured against. That comparison needs reviewing on a regular cadence, because pricing and usage move faster than annual planning cycles. And the output has to feed back into resourcing decisions, or it stays a report nobody acts on.
Done properly, this gives the board exactly what they are asking for. Every AI initiative gets checked against its actual unit economics and against how the last comparable piece of work performed. Some will hold up. Some will not, and those are the ones where the savings are immediate, either by moving the work back or redesigning how the AI does it.
How do you measure ROI from Agentic AI?
Agentic AI changes the shape of this problem rather than solving it. An agent does not make one model call and stop. It chains several together, retries when a step fails, and keeps running unattended long after a person would have paused to check the output. Token cost tracking has to follow that whole chain, not just the number of times someone opened a chat window.
An agent’s real cost is the cost of every step it takes, not the cost of the single response a person eventually sees.
The tokenomics discipline described above still applies. What agentic AI adds is a layer of AI agent cost management sitting on top of it, because the unit being costed is no longer a single response. It is a sequence of actions, some of which fail, retry, or call other agents in turn.
How do we manage the cost of running AI agents across the organisation?
The same discipline that already applies to people cost and token cost extends here. Measure at the level of the individual agent, not just an organisation-wide AI total, because two agents doing similar work can have very different cost profiles once retries and failure rates are accounted for. Track those retries and failures specifically. They are often where an agent’s true running cost hides, well above what the headline token price suggests.
Every agent needs a named owner for its running cost, in the same way every project already has a budget owner.
That ownership question matters as much as the measurement itself. An AI ownership and financial accountability framework gives every agent a person who can answer for what it costs and why, rather than leaving agent spend as an unowned line inside a larger AI budget. The same portfolio model already holding people cost and token cost side by side needs to hold agent cost the same way, broken out agent by agent rather than folded into a single total.
My Final Thought
Answering “what are we getting back for our AI spend, and can we prove it” takes a model that holds people cost and token cost side by side, a consistent way of capturing project cost on exit, and a PPM or SPM setup built to carry both over time.
If your organisation cannot currently answer that question with a number, Ignite’s team can help you build the model that does. Speak to one of our experts about bringing token cost and people cost together in a way that gives your board the proof they are asking for.
Frequently Asked Questions for Managing AI Spend
The same discipline used for any other AI initiative extends here, pairing token cost against business value at the level of an individual task or decision. Agentic AI adds one complication. The unit being measured is a chain of actions rather than a single response, so cost has to be tracked across every step an agent takes, including retries and failures, not just the number of times it’s called.
An agent’s token usage covers a chain of steps, sometimes including calls to other agents, rather than one request and one response. Tracking needs to follow that whole chain and separate out retry and failure costs, which are often where an agent’s real running cost hides, well above the headline token price.
Every agent needs a named owner responsible for its running cost, the same way every project already has a budget owner. Without that ownership, agent spend tends to sit as an unowned line inside a larger AI budget rather than something anyone can be asked to account for.
Measure at the level of the individual agent rather than a single organisation wide AI total, since two agents doing similar work can have very different cost profiles once retries and failures are factored in. Then hold that number in the same portfolio model already tracking people cost and token cost, broken out agent by agent.
Tokenomics is the discipline of tracking what an organisation’s AI and LLM usage costs against the business value that usage produces, measured at unit cost rather than a single monthly total. In this context, a token is a billable unit of AI usage, not a reference to cryptocurrency.
