2 July, 2026
A practical framework for deciding what to automate in your practice, and the time-token breakeven point research on why some judgment work never will be automated.
Most accounting teams are already making this call informally, task by task: does this go to a person, or does it go to a model? The honest answer usually comes down to cost and speed in the moment, not a clear rule for when AI is genuinely the cheaper option at the same quality. And it doesn't answer when AI never will be the cheaper option, no matter how good the models get.
There's now a formal name for that threshold: the time-token breakeven point, the moment a piece of work becomes cheaper to produce with AI tokens than with a person's billable hour. It comes from a recent paper working through the economics of AI versus human labor in regulated accounting specifically. It's worth understanding, not because the term itself matters, but because the framework behind it gives a much clearer way to decide what to automate in your own practice, and what to stop trying to.
This post is that practical version: what kind of work you're actually deciding on when you automate a task, why "AI is cheaper" claims are usually missing something, what tends to get cheaper over time and what never does, and a simple way to apply this to your own team.
Automating a task isn't one decision. It's usually four different decisions bundled into one, and only one of them is actually a live question.
The rule-following part (posting a transaction to a known chart of accounts, applying a fixed depreciation schedule, charging a standard rate of tax) isn't really an "AI question" at all. It's the older question of software versus labor, and most practices have already resolved it one way or another. The research behind this calls it the deterministic substrate. The genuinely contested part is the judgment underneath it: classifying a novel transaction, deciding how a standard applies to a messy fact pattern, estimating something like an expected credit loss. That's the only piece where "should this be a person or a model" is still a real, moving question, what the paper calls the judgment residual.
The other two pieces aren't cost questions at all, and it's worth being clear-eyed about that before comparing anything on price. Some work is valuable specifically because a trusted person did it, the relational residual. A client isn't only paying for the number; they're paying for someone accountable having actually looked at it. And some work the law simply assigns to a licensed person, full stop, regardless of what a machine could technically produce, the regulatory residual. Knowing which of these four you're actually looking at is most of the battle before you ever get to a cost comparison.
An hour and a token don't convert into each other at a stable rate, which is exactly why so many "AI is now cheaper than a person for this" claims fall apart on closer inspection. An hour is a fixed amount of time. A token isn't a fixed amount of work: a more capable model does more per token, but can also burn far more tokens reasoning to the same answer, so more capable doesn't automatically mean cheaper.
The token price adds a second wrinkle. It's currently held below the real cost of running these models by investment competing for market share, not by production cost. That means it can fall further as compute gets cheaper, or jump sharply if that subsidy gets pulled back. This is precisely the gap the time-token breakeven point is built to close: it's defined as the wage at which producing a given piece of judgment work costs the same in tokens as in human time, at a fixed quality bar, and it only means something once you've pinned down which model, what token price, and how much of the task has already been automated.
In other words, the breakeven point isn't one number you can quote. It's a moving line that depends on those three things. That's a useful thing to internalize even if you never touch the formula: any time someone tells you "AI is cheaper than staff for X," the honest response is "cheaper using which model, at what token price, for how automated a version of X."
Work gets cheaper to automate as a firm builds reusable systems around it (a fine-tuned model, a tested workflow, a checked procedure), not simply because the underlying AI model got smarter. The research splits the token cost into two parts: capex tokens, spent once to build the reusable system, and opex tokens, spent every time it actually runs afterward. Each time a piece of judgment gets built into something reusable, it moves from being redone (and repriced) every single time to running near-automatically at a fraction of the cost, and that's really what shifts the economics, more than any one model release does.
How fast that happens for a given task depends heavily on how clearly the standard governing it is written. Accounting judgment mostly operates inside standards (SFRS, IFRS) that already spell out the treatment for most situations a practitioner runs into, which means most of it is close to being turned into a reusable system once someone invests the effort. Compare that to something like strategic or aesthetic judgment, where there's no comparable authoritative standard to build against. The "right answer" stays genuinely open, so there's nothing to encode and nothing shifts.
The practical upshot: the judgment work that feels irreplaceable on your team today is, in most cases, work nobody has built a system around yet, not work that fundamentally can't be. That also means being first to automate a given task isn't a lasting edge. Whatever standard let you build the system will let a competitor build the same one eventually, so any advantage narrows over time unless it's anchored in something else entirely, which is exactly what the next section is about.
Some work stays with a person permanently, and it's worth knowing which parts so you're not chasing automation where it was never going to pay off. Two categories sit outside the cost comparison entirely, no matter how good or cheap AI gets. Recognizing them early saves a lot of wasted automation effort.
The first is anything where the human presence is itself what the client is paying for. If part of the value is that a trusted, accountable person actually engaged with the numbers, swapping in a model doesn't produce a cheaper version of that service. It produces a different one, and the client notices. The second is anything the law specifically assigns to a licensed individual: signing an audit opinion, attesting a tax return, taking on professional liability. That assignment doesn't loosen because a model got cheaper or better. It's simply not a cost decision to begin with.
In regulated accounting, the first category turns out to be fairly small. What a client is really paying for on a set of audited accounts is mostly the assurance those accounts carry, not the relationship that produced them, which means these two limits mostly collapse into one practical bound: the signature. Everything upstream of that signature is, sooner or later, a matter of time and investment. The signature itself isn't.
Before automating any given task, it's worth running it through four quick questions rather than jumping straight to "can a model do this." First: is this actually rule-following, or does it require interpreting a standard against a specific set of facts? If it's the former, this was never really an AI decision. It's a software one, and it's probably worth resolving either way. Second: is the judgment inside it governed by a clear, explicit standard, or is it genuinely open-ended? The clearer the standard, the sooner it's worth building a reusable system around it, because the cost of doing so keeps falling.
Third: is part of what the client is paying for the fact that a specific accountable person looked at this? If so, no cost comparison is going to change that, and trying to force it usually just produces a worse (if cheaper) version of the service. Fourth: does the law require a licensed person to sign or attest to this output? If yes, that part stays human regardless of everything else. Automate what leads up to it, not the signature itself.
Run through those four for any task your team is debating, and you'll usually land on a much clearer answer than "AI is cheaper" or "AI can't do this yet." Most judgment work in accounting will eventually shift toward being system-assisted as the standards behind it get built into reusable tools. This is the shape of what Jaz's AI agents are built to handle: doing the deterministic and increasingly the judgment work, while keeping a licensed, accountable person exactly where the client and the law still need one.
Deciding what to automate in a practice comes down to separating four kinds of work: rule-following (already a software question), interpretive judgment (the only part where AI-versus-human cost is a live question), the value of a trusted person's involvement, and work the law assigns to a licensed individual. Only the second one actually has a moving cost line, the time-token breakeven point, and that line depends on the AI model in use, the token price, and how much has already been automated, not on a single fixed number.
Most judgment work in accounting sits inside clear standards, which means it tends to get automated over time rather than staying permanently human. But the value of a trusted relationship and the requirement for a licensed signature don't move regardless of how cheap or capable AI becomes. Knowing which category a given task falls into is usually more useful than any cost calculation.
What is the time-token breakeven point?
It's the point at which producing a piece of accounting work with AI tokens costs less than producing it with a person's billable hour, at the same quality bar. Below the threshold, human time is cheaper; above it, tokens are.
Is the time-token breakeven point a single number I can look up?
No. It depends on the AI model generation in use, the current token price, and how much a firm has already invested in automating the work. It's a moving line, not a fixed figure.
Why isn't "AI is cheaper than a person" a fair comparison on its own?
An hour is a fixed unit of time, but a token isn't a fixed unit of work. A more capable model can do more per token, or spend far more tokens reasoning to the same answer. The token price is also volatile, currently held below cost by investment rather than the true cost of production.
How do I know if a task in my practice is worth automating?
Ask whether it's rule-following or genuine judgment, whether the judgment is governed by a clear standard, whether part of the value is a specific person's involvement, and whether the law requires a licensed signature. Only rule-following and standard-governed judgment are realistic automation candidates.
What are capex tokens and opex tokens?
Capex tokens are spent once to build a reusable AI system, like a fine-tuned model or a tested workflow. Opex tokens are spent every time that system actually runs. Together they make up the token-side cost that's compared against an hourly wage.
What do 'deterministic substrate,' 'judgment residual,' 'relational residual,' and 'regulatory residual' mean?
They're the research's names for the four kinds of work inside any accounting task: the deterministic substrate (pure rule-following, already largely automated), the judgment residual (case-by-case interpretation, the only part where AI-versus-human cost is a live question), the relational residual (work valuable because a trusted person did it), and the regulatory residual (work the law assigns to a licensed person).
What kind of accounting work will always stay human, regardless of AI cost?
Work where the law requires a licensed, accountable person to sign or attest, like an audit opinion or a tax return, and work where the client is specifically paying for a trusted person's involvement, not just the output.
Will more tasks become cheaper to automate as AI models improve?
Yes, particularly judgment governed by clear, explicit standards (like SFRS or IFRS treatments). Those tend to get built into reusable systems faster than open-ended judgment with no comparable standard to build against.
Does being an early adopter of AI automation give my firm a lasting advantage?
Not on its own. Whatever standard let you automate a task will let competitors automate the same one eventually, so a durable edge usually comes from client relationships or expertise rather than being first to automate.
Does this mean AI will eventually replace accountants?
No. Automation can move deep into interpretive judgment as reusable systems get built around it, but it's bounded by client trust and by legal signing requirements, neither of which is a cost decision.
How does this apply to auditing specifically?
Testing and interpretive judgment inside an audit can shift toward automation over time, but the auditor's opinion, the signature, stays with a licensed person because it's a regulatory requirement, not a cost decision.
Where can I read the full research behind this?
The paper, The Time-Token Breakeven Point: Capital Expenditure and the Limits of Labor-AI Substitution in Regulated Accounting, is available on SSRN.
