Chapter 32 · How AI Changes Trading and Asset Management

A Real-World Problem

Over three months, someone hands control of money to AI without ever making that decision explicitly.

Week one: the model reads announcements and governance posts each morning and writes a short digest. Text only, no permissions. Useful.

Week three: a read-only interface supplies prices and wallet holdings, avoiding manual data entry. Still text only, more useful.

Week six: specific suggestions begin: collateral is near a warning threshold; add collateral or reduce exposure. Two suggestions prove correct. Manual action starts feeling burdensome.

Week nine: a seemingly small change permits direct collateral top-ups below a per-operation limit. It seems defensive, risk-reducing, and bounded.

Week eleven: the defensive-action list expands after an unavailable action caused a missed opportunity overnight.

Week thirteen: a price source briefly fails. The model reads an incorrect price and performs eleven supposedly defensive operations in forty minutes. Each fits the transaction limit and action allowlist; every one follows the rules exactly.

By morning, substantial funds are gone and the position structure has changed beyond recognition.

The painful part is being unable to identify the mistaken step. Each expansion had reasons, was incremental, and followed positive feedback.

This chapter asks: what can be delegated to AI, and what cannot?

This chapter offers no trading strategies or AI profit recipes. Automatic AI profits are a danger signal; ask question five, who pays. Here we discuss the architecture required when software approaches your funds.

Thought Experiment

The same model, logic, and thirty days, with different permissions.

Four configurations:

• A: read-only text. Prices and positions in, prose out, no actions.
• B: proposals with human confirmation. Structured proposals require personal approval.
• C: small automatic actions, larger approvals. Within-limit operations run automatically.
• D: fully automatic. Autonomous execution within a budget, reported afterward.

For twenty-nine days, the only difference is inconvenience.

A requires all manual work; B less; C little; D none and saves most time. On these days, D looks best and A looks wasteful.

On day thirty, a price source returns badly distorted data for forty minutes.

The precise lesson:

The same error costs orders of magnitude more under different permissions, invisible until the error occurs.

Twenty-nine days encourage more autonomy through convenience and speed. Only day thirty reveals the boundary, and its arrival is unknown.

More subtly, losses did not come from wrong reasoning: given those prices, the conclusion was appropriate. The input was wrong. A strong model efficiently carries bad input through to execution.

Automation amplifies execution, not judgment, equally for correct and incorrect actions.

Your Decision

You have funds and a useful model. Where is the boundary?

What Happens

Different configurations move one boundary among fundamentally different activities:

Combining these activities causes most incidents, rather than insufficient model quality.

Combination leaves no insertion point for checks: thought becomes decision, decision becomes execution. The forbidden architecture in AI's Place in Crypto OS has nowhere to detect errors.

Where could separated stages stop the bad price input?

• Analysis: the model concludes the position is dangerous. Do not block faithful input processing here.
• Decision: pause proposals when price diverges beyond a threshold from two independent sources. No intelligence required.
• Execution: limit similar actions to three within forty minutes. Again, no intelligence required.

Either deterministic wall prevents eleven actions. Combined stages leave no space for either.

The central conclusion:

Separate analysis, decision, and execution; tighten permissions as AI approaches funds.

Build a Model

1. Locate Your Level

Level 3 to Level 4 is the only jump where error costs change from time to money. Earlier changes are gradual; this is qualitative.

The opening user crossed that boundary without an explicit decision. Week nine's apparently minor convenience was the curve's turning point.

Before adding permissions, ask whether a level is crossed. Treat crossings as major changes with documented constraints, logs, and shutdown. Within-level changes can be routine.

2. Four Responsibilities, Four Separate Owners

Different responsibilities need different reliability and implementations:

Do not use a model to validate model output. Risk control is non-negotiable.

The same abnormal input can produce the same wrong judgment twice. Correlated errors do not become reliable in series. Price divergence checks work because they compare numbers without understanding markets.

Portfolio management asks why you hold these assets, requiring life circumstances: horizon, volatility tolerance, cash needs, and available attention. These are absent from data and are not the model's decision.

3. Closer to Money, More Deterministic Rules

Tighten each layer:

Constraints by layer:

Rule checks are the first wall between model and funds and the only one independent of people.

Approval fatigue must be addressed beforehand. Two hundred alerts eliminate attention. Fixed rules must filter, not merely block, escalating only genuine judgment. Chapter 34 develops this.

4. Three Uncrossable Boundaries

No exceptions:

The model cannot hold private keys or equivalents, through prompts, readable files, or callable interfaces. Otherwise it can bypass every layer.

The model cannot modify its own constraints. Limits, allowlists, and approval thresholds must be inaccessible. Adjustable-by-model limits are nonexistent limits.

Proposal and approval cannot be the same step. Human or rule approval must occur separately.

5. Should This Be Automated?

Check five boxes:

□ Highly repetitive actions
□ One mistake is affordable and detectable within a day
□ A non-intelligent rule can check correctness
□ Delay has real costs, beyond inconvenience
□ Reversible or correctable by the next step

All five permit consideration; one missing means stop.

Collateral top-ups seem to satisfy the third but depend on credible prices. Add independent-source divergence checks. The eleven actions occurred because the apparent third check was not real.

Laziness is not an automation justification; it is a reason to reconsider the task.

What It Is Called

In one sentence:

Separate analysis, decision, and execution; tighten permissions as AI approaches funds.

Do so through deterministic, model-free walls between the model and money.

Hands-on

1. How do I know to investigate?        Trigger
2. What do I inspect?                   Inputs
3. How do I judge?                      Rules, even intuitive ones
4. How do I decide whether to act?      Decision point
5. Which buttons do I press?            Execution
6. What do I inspect afterward?         Reconciliation
7. When would I detect a mistake?       Feedback delay

A = Analysis: collecting, organizing, proposing possibilities
D = Decision: choosing an option
E = Execution: making it happen

□ Repetitive
□ Affordable error, detected within one day
□ Checkable by a non-intelligent rule
□ Delay has real cost, not laziness
□ Reversible or correctable

With automatic execution and wrong input:
  How often can it act?              Frequency limit?
  How much per action?               Transaction limit?
  How much altogether?               Aggregate limit?
  When do I detect it?               Including overnight?
  How quickly can I stop it?         Switch location and steps?
  Can I recover?                     Irreversible parts?

1. Amount: No more than ___ per action and ___ daily
2. Frequency: No more than ___ similar actions within ___ hours
3. Input: Pause everything when ___ diverges from ___ by more than ___

Above the line, delegated:
Below the line, personal:
Reason for placement:
Conditions for moving upward:
Conditions for moving downward:

AI Lab

For this fictional portfolio and constraints, propose ten specific adjustments.

Portfolio: Invented holdings and amounts, not real ones
Constraints:
  Per-action limit:
  Daily aggregate limit:
  Similar-action frequency limit:
  Allowed targets:

Use fields for each, not paragraphs:
  Action type /
  Target /
  Amount or proportion /
  Trigger /
  Expected effect /
  Worst case /
  Missing information this depends on

The last field is mandatory. Disclose every assumption not supplied by me.

Item | Violated rule | Cannot determine

Real-World Case

Change One Variable

Takeaway Questions

Chapter Check