We expect AI to write the first draft of many token models, and it already can: allocation tables, unlock calendars and supply simulations in minutes. What it cannot do is decide what is fair, foresee everyone who will try to game the design, or answer for the result. In one academic test, an AI planner setting incentive rules landed within 10–35% of the textbook optimum, and a simple brute-force search still beat it.
A note on the legal and regulatory points here: they summarise public sources as of November 2026 and should not be relied on. Rules for AI agents are still being written, and some of what is described here is proposed or coming soon rather than settled. Check the current position, or ask a qualified professional, before acting on it.
What can AI actually do with a token model today?
AI can draft, calculate and simulate, which covers a large share of the hours that go into a token model. Tokenomics means the rules for a token: how many exist, who gets them, when they unlock and what holders can do with them. A capable AI tool can turn a plain-English brief into a first allocation table, a monthly unlock schedule and a model of circulating supply.
It can also build simulations: small artificial economies where simulated holders react to the rules. Academic work suggests the idea has substance. In a 2025 study, researchers had an AI "planner" set tax rules while up to 100 AI-simulated workers responded to them.
Tax design is a fair comparison, because like token design it sets rules that self-interested people then react to. In the paper’s detailed results, the AI planner improved on a simple baseline but landed within 10–35% of the textbook optimum rather than on it. In one setting, a plain grid search, which just tries many options, beat it by about 10%. AI is a useful modeller. On this evidence it is not yet the best optimiser.
Where does AI-written tokenomics go wrong?
It goes wrong where token design is hardest: people who set out to game it. An AI-built model optimises for the goal it is given, under the assumptions it is given. Real holders do not follow assumptions. Some farm rewards, some fake activity, and some wait for the first unlock to sell.
On-chain incentive systems show how fast gaming happens. A June 2026 academic study of ERC-8004, an Ethereum standard for agent identity and reputation, looked at the reviews agents had received. On one network, 90.6% of reviewers showed coordinated fake-account behaviour, known as a Sybil attack. Once that feedback was removed, 86.8% of rated agents there had no valid ratings left.
That was a reputation system, not a token, but the lesson carries over: a design that looked reasonable on paper was gamed cheaply in practice. Simulation helps only against the attacks someone thought to include. In a 2026 academic simulation of token buyback rules, a simple threshold rule lost 72.3% of its treasury under a sustained manipulation attack, while a more robust controller lost 8.5%.
The specific numbers come from a simulation, so treat them as illustrative. The durable point is that a design's safety depends on which attackers its designers imagined. Other failures are more mundane, in our judgement: figures that do not reconcile across tabs, confident "benchmarks" with no source, and template allocations copied from project to project.
What should stay human?
The goals, the trade-offs and the signature should stay human; the arithmetic can be shared. Here is how we would split the work:
| Part of the work | AI's role | Human's role |
|---|---|---|
| Purpose: what the token is for | Summarise options and precedents | Decide, and be able to say it in one sentence |
| Allocation and fairness | Draft splits and compare scenarios | Judge what is fair to community, team and backers |
| Vesting and unlocks | Build the calendar and check the sums | Choose cliffs and durations the team will defend in public |
| Stress testing | Run many scenarios quickly | Choose the attackers and the worst cases to test |
| Legal and regulatory fit | Flag questions to ask | Get qualified advice; AI output is not that advice |
| Sign-off and disclosure | Draft the explainer | Publish under a named, accountable team |
The reason is not that humans are better at arithmetic. They are not. It is that tokenomics is a set of promises to people, and only people can make a promise and be held to it.
Judgement also means saying no. An AI tool will happily produce a model for almost any brief, including one that gives insiders too much or unlocks too early. It rarely pushes back on the premise. A good human reviewer starts by asking whether the token needs to exist at all, and what it should not try to do.
Who is accountable when an AI-drafted model fails?
The team that chose to publish it is accountable, whoever or whatever drafted it. That is our view, and it is the simple case. The harder questions arise when an agent acts on its own.
In October 2026 remarks, listed in the CFTC’s index of public statements and remarks, the CFTC chair openly asked who is liable if an agent goes rogue, and whether the developer of an agent that trades for a user would become subject to registration. He framed these as open questions, in his own name rather than the Commission's. Nothing about them is settled.
A token model is different from a trading agent, because a human usually decides to adopt it. That decision is the accountable moment. "The model suggested it" explains where an idea came from. It does not move the responsibility.
What could go wrong?
The main risk is false confidence: an AI-drafted model looks finished long before anyone has thought it through. A spreadsheet to four decimal places looks rigorous, even when every input is a guess. Five failure modes are worth naming:
- False precision. Clean outputs hide weak assumptions about demand, price or behaviour.
- Unimagined attackers. The stress test covers the scenarios in the prompt and nothing else.
- Unsourced benchmarks. An AI tool may state "typical" allocations it cannot trace to anything.
- Leaked secrets. Pasting investor terms or a cap table into a general AI tool may expose them.
- Same tools, same flaws. If many teams use similar tools with similar prompts, designs may converge, and so may their weaknesses. That last one is our inference, not a measured finding.
The trust takeaway
AI will write a lot of tokenomics. The question that matters is who reviewed it and who stands behind it. Use AI for drafts and simulations; keep purpose, fairness, the choice of attackers and the final sign-off with people who can be named.
If you are reading a token model, or building one, ask:
- Which parts were AI-drafted, and who reviewed them?
- Who chose the attack scenarios in the stress test, and what was left out?
- Can every figure be traced to a source or a stated assumption?
- Can the team explain each allocation in plain words?
- Who signs off, and will they answer questions in public?
This is the question Alcyone’s fundraising and tokenomics advisory, part of TrustSwap's strategy and advisory practice, works on the human half of the table above: the judgement calls on a token model's purpose, incentives and trade-offs. Where this is going, in our view, is AI doing more of the drafting and simulation while advisors spend their time on the judgement calls. That is a direction we see, not a product.