Part 2 of 2: Beyond the Basics & What to Do About It

Daniel Lawrence, CEO, bots for that

If you haven’t read part one of this article and you aren’t intimately familiar yet with how the costs of AI actually work, please read that first. This one won’t make much sense without it, I covered what a token actually is, how input and output get charged differently, why a subscription feels unlimited when it isn’t, and why an agentic workflow can quietly rack up cost in a way a single chat never will. Consider that the foundation. This part two builds on top of it.

I said at the end of part one that once you actually understand how the cost works, some of the more comfortable assumptions about AI adoption might start to look less solid.

The bill nobody’s looking at

In most businesses using AI right now, somebody, somewhere, pays for it every month, and almost nobody who’s actually typing the prompts ever knows what any one of them cost.

That’s meant as an observation and not a criticism, but a subscription is a flat fee and an API bill gets posted to an IT expense line. Either way, the person deciding whether to run one more query, connect one more automation, or let an agent take one more pass at something, isn’t the person watching the numbers go up.

To help explain this, there’s a more familiar analogy of this same issue. Smart Meters have been rolling out across British homes for over a decade now, and the point of these is simply visibility. The meter shows what you’re actually using, in something close to real time, instead of an estimated reading every few months.

But the visibility on its own doesn’t change anything, especially if you’re on a flat-rate tariff. However, once that visibility is paired with a tariff that prices usage depending on when we use it, our behaviour definitely changes. Time-of-use tariffs have us all scheduling our washing machines to run at 2am once we know it costs about 50% less, especially with energy costs about to spiral out of control yet again.

AI Tokenomics has both halves of that same mechanism available to us, but most businesses are only using half of it, at best. The “meter” half is straightforward with a per-call or per-project breakdown of what’s actually being spent, down to the team or individual, rather than one number on a monthly bill. The “tariff” half already exists too, and it’s very underused: Anthropic, OpenAI and Google all offer a batch processing option, each at roughly half the standard rate, for exactly the same reason electricity is cheaper overnight. So, work that doesn’t need to happen instantly, in real time, can be processed more cheaply than work that does. The trade-off is the same one you’d accept with the washing machine, in that batch jobs aren’t instant, and typically take up to 24 hours to complete in exchange for the discount. However, there’s plenty of decent use cases for this, many vendors adding AI wrappers to their software are doing exactly this, for instance: bulk classification updates, overnight reconciliation runs, data synchronisation. Basically anything that doesn’t need an answer in the next three seconds, is the AI equivalent of running the washing machine after midnight rather than at 6pm. Most businesses run everything at the equivalent of peak rate, all day, because nobody’s knows there’s an off-peak option.

It’s important to be honest about the limits of this, too, because theirs often a catch with these kind of time-of-use pricing. The risk is that you could also end up paying more overall than you would on a standard flat rate, and the reason is simply due to not shifting enough of your usage to actually benefit from the cheaper rate. So, although meters and price tariffs make a behaviour change possible, they don’t make the savings automatic. Somebody still has to actually notice the incentive and act on it, which is exactly the kind of thing that needs a person with clear ownership of the AI spend.

Cheaper AI won’t save you money. It’ll just mean you do more of it.

I’ve seen this one catch out quite a few, and it sounds backwards.

Token prices have fallen a long way, and they’re still falling. So the intuitive assumption is that AI spend should be falling with it. For a lot of businesses, it isn’t. It’s going up, month on month, even as the price per token keeps dropping.

There’s a well-established economic pattern with technology like this, where the more efficient your AI-driven task becomes it won’t necessarily reduce your total AI token consumption, but in fact can increase it, because efficiency makes processing more viable for more use cases or proportion of transactions than before. So, cheaper doesn’t mean less spend. It means more will get done at the same spend, or higher.

Applied to your business, if generating a first draft, a summary, or an analysis gets cheaper, you don’t do the same amount of work for less money. You do more work, because now it’s affordable to ask for things you’d never have asked for before. That’s not a flaw. In plenty of cases, that’s the entire point of adopting AI in the first place.

There’s one other lesser-known, uncomfortable truth to add to this, and that’s that your bottlenecks of the past were never really the cost of shifting them. It was (or still is) the cost of a qualified human actually checking it, approving it, and standing behind it. That cost doesn’t fall when token prices fall. It doesn’t fall at all. So AI makes the input side of the work almost free, whilst the human judgement on the output side stays exactly as expensive as it always was, and now there’s a great deal more of it arriving for review. This should toll a very familiar sounding bell to any accountants reading this in particular right now.

That’s not an AI tokenomics problem on its own. It’s the mechanism behind something I’ve written about before, and it’s worth repeating here: the Productivity Paradox, where AI is accelerating output generation without accelerating the human judgement and decision-making that has to sit downstream of it. AI Tokenomics is the part of that story you can actually put a number on. Cheaper tokens, more generated work, an unchanged bottleneck at the human review stage, and a firm that’s individually more productive while collectively further behind. That’s not a contradiction. That’s what happens where you add capacity to one part of a system that can make the whole system slower, not faster, if the part you added capacity to wasn’t the only constraint – I’ve seen this phenomena play out several times over the past decade with process automation – the chain is only as strong as the next weakest link.

The cost of a long chain isn’t just the sum of its steps

In part one, I flagged that agentic workflows, the kind that loop, retry, and hand work between multiple steps or sub-agents, can produce a genuine diseconomy of scale, where the cost per unit of output rises rather than falls as the system gets more complex. It’s not just that longer chains cost more though, it’s that they can lose value at each and every hop, whilst you’re still being charged in full for every hop in the chain.

With each hop in a multi-agent AI workflow, where one agent’s output becomes the input of another, losing a bit of precision, context, or intent along the way is normal, whilst still consuming the full token cost of each step. When you string together enough hops, you can effectively end up paying for five or six full-price steps to deliver the value of only two, and there’ll be nothing in the charges that highlights this has happened. Nothing will very visibly break, it’ll just quietly get more expensive to achieve less, one hop at a time.

The practical implications of this aren’t to avoid agentic workflows altogether, but to be aware that the number of hops in an agentic chain are design decisions with costs attached, and shouldn’t be treated as one whole. The fewest hops that reliably get the job done doesn’t just result in a cleaner build, it’s usually the cheaper one too.

Drowning in manual, repetitive work? Tell us the task and we’ll show you what to automate.

There’s one more point to remember when making your AI agent design decisions, and it’s the one I see most often and easily overlooked by anyone reaching for GenAI as the default tool for every step – not every hop in a chain needs to be a generative AI model call at all. If a step is genuinely deterministic, pulling a fixed field from a structured document, moving data between two systems, applying rules that rarely change, then that’s exactly the kind of task established automation methods already do reliably, at a fraction of the cost, and without the value-loss-per-hop problem above, because a deterministic script doesn’t lose precision the way a model call can. The question worth asking at every hop in an agentic chain isn’t just “can GenAI do this?” It’s “does this step actually need judgement, language or reasoning to get done, or am I reaching for the most expensive, least predictable tool, because GenAI is the interesting and fashionable new option?”. The workflows that hold up best, cost-wise, tend to be the ones that treat GenAI as one tool among several in the chain, reserved for the hops that genuinely need it, rather than the default setting applied to all of them.

You’re budgeting for the wrong kind of risk

Most organisations that have gotten as far as putting any cost controls around AI at all are watching an average: a monthly figure, a per-user estimate, a trend line on a P&L. That’s reasonable risk management for the kind of cost that getting started and exploring with AI usually is, but it’s not for the kind of cost it becomes when quickly moving into scaled deployments.

A single poorly configured agent, or a loop nobody catches quickly enough, isn’t an average-cost event, it’s a disproportionately expensive outcome when it happens. And I’ve seen it happen in just the last week, where a firm built and released an agent to reconcile bank transaction for all 400+ of it’s clients overnight – it did the job well enough, but the token usage exploded, along with the cost, and a £20 plan went up to £250 overnight.

Most AI cost governance right now is focused on a hard spend cap that forces the process to pause, and receive an alert, or a human sign-off before continuing. This is ok as a rough budgeting tool during the exploration, but it doesn’t prepare you for when one week doesn’t look like any of the others in the early phases.

The false economy hiding inside using the cheaper model

I flagged this tension at the end of part one and it’s worth spelling out again here, because it’s the one most likely to bite a business that wants to take AI cost seriously.

Right-sizing, is the process of using a smaller, cheaper model for simple tasks rather than defaulting to the most capable one, and it’s generally good advice. It’s also very easy to fall into the trap where it stops being good advice, because cheaper and reliable aren’t the same outcome when it comes to AI.

I’ve written about this same failure before in a different article. The most reported AI failures aren’t about the capability of the models, but about deployment literacy. A generalist AI tool asked to do specialist work, configured by someone who didn’t fully understand what they were configuring. In these cases, chasing the cheapest available token price is the same kind of mistake. Neither one is a cost problem dressed up as a technical one. Both are a judgement problem, and no amount of clever pricing analysis fixes a judgement problem.

And this connects straight back to accountability, which is the thread running through nearly everything I write about AI in professional contexts. If cost per token becomes the metric a business actually steers itself by, rather than a metric it merely watches, you’ve unknowingly handed the decision about “good enough” to whichever number happens to be visible on the billing statement, rather than to whoever’s actually accountable for the outcome. That’s the same structural gap as every other accountability problem in this space right now: the professional duty of care either sits with a named, accountable person, or it doesn’t. Optimising for the cheapest defensible token spend, rather than the most appropriate one, is a silent way of letting it not.

What to actually do about this

All of the above is only useful once it turns into something you actually do on a Monday morning. That looks different depending on the size of your business, so here’s some practical steps.

If you’re a small or medium-sized business:

Find out whether you’re actually on a subscription or paying per token. If it’s just people chatting with an AI tool day to day, you’re on a flat fee, then most of this article is background knowledge rather than a list of action items for you. The moment you connect AI to an automation, an integration, or a workflow tool, you’ve almost certainly moved to per-token billing, whether anyone consciously decided that or not. First step, go and check.
Give your AI spend a named owner, even if it’s just you. A shared plan and a monthly surprise isn’t a budget. Someone should be actually looking at what’s being spent, on what and why, on a regular basis, not just when the bill looks unusually high.
Put a hard spending cap on anything connected via an API. You almost certainly don’t have a finance team watching AI spend in real time, which makes a hard limit the closest thing you’ve got to an insurance policy. Set one, low enough that you’d actually notice before it became a problem.
Ask whether any of your recurring AI tasks can wait. If you’re running a weekly summary, a batch of categorisations or reconciliations, or anything that doesn’t need an answer in the next few seconds, check whether the tool or platform you’re using supports batch processing. If it does, that’s a straightforward saving sitting there unused.
Before you build, or pay someone to build, an agentic workflow, ask what’s actually deterministic in it. If part of the process is “always do X when Y happens”, then there’s a cheaper, more reliable and simpler automation than a chain of AI calls. Don’t pay GenAI prices for a job a much simpler tool already does. This includes any which are embedded in a core software system.

If you’re at a larger business, running this at scale:

Push cost visibility past the finance reports. A company-wide monthly total tells you spend is high. It doesn’t tell you which team, workflow, or habit is driving it. Genuine chargeback, to the department or project actually generating the cost, is what turns a report into something people can act on.
Give every agentic use case a named business owner, not just a technical one. Someone needs to be accountable for what it costs and what it does, and that needs to be a person and an outcome, not just an IT spend category.
Create mechanisms to control overspend risk. Build in hard caps, alerts, and an escalation path that can pause a process automatically, rather than relying on a monthly spend review that will only tell you about the bad week roughly a month too late.
Audit your agentic workflows for hop count and real-time necessity, on a schedule, not just at launch. Every step should have to justify why it’s a GenAI call rather than a deterministic script or existing integration, and why it needs to run live rather than as a cheaper batch job. Workflows can drift, where hops get added faster than they get removed unless someone’s actually checking.
Keep quality and cost review separate, and owned by different considerations even if the same person handles both. The moment “cheapest” and “good enough” collapse into a single decision made on price alone, nobody’s actually accountable for the outcome any more.

Where this leaves us

Six essential truths to sit alongside the five straightforward ones from part one:

The bill may be invisible to the people best placed to control it, and visibility alone won’t fix that. You need the meter and the off-peak rate, a per-use breakdown paired with a genuine cost incentive, like batch processing, or nobody has a reason to change or improve anything.
Falling prices won’t shrink your AI spend. They’ll likely grow the amount of work you ask AI to do, while the human judgement bottleneck downstream stays exactly as expensive and exactly as slow as it always was.
Long agentic chains lose value at each hop. Silently, while still being billed for every hop in full, so the number of steps in a workflow is a cost decision, whether or not anyone’s treating it as one.
Not every hop needs to be GenAI. Deterministic steps are usually cheaper and more reliable done an established way, with GenAI reserved for the parts of the chain that genuinely need judgement or language.
Average-cost monitoring won’t catch the risk that actually catches you out. That risk needs a different kind of control than a monthly trend line.
The cheapest model is a false economy the moment it becomes the goal rather than a consideration. If the token price is doing the deciding, nobody’s actually accountable for the outcome anymore.

Put those five together with the mechanics from part one, and the real point of this whole two-part piece comes into focus. AI Tokenomics is never really about the price of a token. It’s about how AI gets costed and how it gets used, who notices the spend, what gets built, how far a workflow is allowed to sprawl, and who ends up holding the judgement call at the end of it.

Understanding the mechanics in part one lets you ask the right question. Addressing the uncomfortable points outlined in this one is what hopefully helps you find the right answers for your AI program.

What are AI “Tokenomics”? | Part 2