AI investing needs a mandate, not a better prompt

On August 25, 2026, Scalable Capital opened its platform to major AI agents. Clients can connect ChatGPT, Claude, Grok, and compatible local agents to their brokerage account. An external AI agent now has a technical route from a natural language conversation to a securities order.

The conversation can now reach the portfolio

Through an MCP server or a local command line interface, the AI can reach portfolio analysis, market data, watchlists, price alerts, savings plans, and securities orders. This is more than another chatbot inside a broker app.

The order is not executed in secret. Scalable requires the investor to approve it before execution. Payments remain outside the agent. Access can be withdrawn, and the normal account permissions continue to apply.

Those controls answer an important question: may this system reach the account and prepare this transaction?

They do not answer a different question: does this transaction belong in this investor's strategy?

That distinction is about to become one of the central questions in personal investing.

Finance is moving from answers to actions

Scalable Capital is the clearest retail investing example in Germany, but it is part of a wider shift.

In May 2025, DKB announced a direct partnership with OpenAI. The bank described plans for AI voice support in banking and agentic workflows for document processes. ING Deutschland has used generative AI in chat and telephone service since 2025. From February 2026, an AI agent began handling defined card service tasks within regulatory limits.

In March 2026, Visa launched its Agentic Ready program in Europe with Commerzbank and DZ Bank among the first participating institutions. The program tests payments initiated by AI agents in a controlled environment, using authentication and tokenisation to keep each transaction attributable to a real person.

The investment industry is moving as well. In August 2026, LAIQON and Amundi launched an active European equity ETF whose portfolio process uses AI. A few days later, Deutsche Bank presented a Financial Research Agent developed with Google Cloud. It brings together company, market, and financial data for structured and traceable research inside a regulated banking environment.

These projects do very different jobs. Together, they show the direction of travel. AI in German finance is leaving the isolated chat window. It is moving closer to research, decisions, transactions, and operational workflows.

A prompt is a request, not a constitution

A prompt captures what someone wants at this moment.

“Find three semiconductor companies with strong margins.”

“Compare my portfolio with a global index.”

“Prepare a savings plan for this ETF.”

None of these instructions explains whether the investor is building retirement capital, preserving a house deposit, or speculating with a small separate account. They do not state how much concentration is acceptable, which cash reserve must remain untouched, or what should happen after a large loss. They rarely explain which earlier decisions should still govern the next one.

A longer prompt does not solve the problem. It merely places more policy inside a conversation that can be edited, omitted, misunderstood, or replaced by the next conversation.

An AI agent is designed to respond to the request in front of it. An investment strategy is designed to remain intact when that request is emotional, inconsistent, or badly timed.

What an investor mandate contributes

An investor mandate is a written set of rules that the investor gives to themselves before a particular trade is under discussion. In institutional investing, the related concept is usually called an Investment Policy Statement.

For an individual investor, a useful mandate can record:

1. The purpose of the portfolio and its time horizon.

2. The strategic allocation and permitted ranges.

3. Limits for a single position, sector, region, or asset class.

4. Liquidity needs and the cash reserve that must remain available.

5. Instruments or forms of leverage that are permitted or excluded.

6. Rules for rebalancing, exceptional decisions, and cooling periods.

7. The process for reviewing and changing the mandate itself.

The mandate is not a prediction engine. It does not know which stock will outperform or where interest rates will be next year. Its job is more modest and more durable. It defines the conditions under which an investment decision belongs to the plan.

This separation matters because a language model can help analyze an opportunity without being allowed to redefine the investor's purpose. The model may change. The broker may change. The mandate should remain with the investor.

Confirmation is necessary, but it is not alignment

Scalable Capital requires investors to approve orders before execution. MCP guidance also recommends keeping a human able to deny tool calls. This is the correct baseline for a system that can take consequential actions.

Yet a confirmation screen proves only that a person clicked. It does not prove that the decision fits their own strategy.

People approve actions they later regret. They approve them under time pressure, after a persuasive explanation, or during a market move that makes inaction feel intolerable. A capable AI can make a weak idea sound orderly. It can produce a polished rationale for an action that conflicts with the investor's enduring plan.

Human approval and strategic alignment therefore solve different problems.

Approval asks, “Do you want to place this order now?”

Alignment asks, “How does this order compare with the rules you chose when no order was competing for your attention?”

Both questions belong in the process.

MCP connects capabilities, not intentions

The Model Context Protocol gives AI applications a standard way to connect with external data and tools. A broker can expose portfolio information and order functions. A research provider can expose company data. A policy service can expose a mandate, constraints, and decision history.

MCP does not decide which of those sources should govern the others. It is connection infrastructure, not an investment philosophy.

That is precisely why it creates an opening for a separate investor controlled policy layer.

The AI can use all three, but none needs to own the entire relationship. Research remains research. The broker remains responsible for the account and execution. The investor's mandate remains portable.

What a guardrail should do

Suppose an investor asks an AI agent to buy another semiconductor company after a strong rise in the sector.

The research may be sound. The company may be profitable. The order may be technically valid. Even so, the proposed purchase could conflict with the investor's own rules.

The mandate service might find that technology already represents 28 percent of the portfolio while the permitted range ends at 25 percent. It might find that the cash reserve has fallen below the stated minimum. It might also find a rule requiring a 48 hour cooling period for purchases outside the strategic allocation.

The right response is not to choose a different stock for the investor. It is to make the conflict visible before the order reaches execution.

That is a guardrail. It applies the investor's own prior instructions to the current decision. It does not substitute the platform's market view for theirs.

Guardrails need memory

A static rule is useful, but investing unfolds over years. The more valuable system remembers how decisions developed.

It should be able to show that the investor raised a concentration limit three times during a rally. It should remember that a similar purchase was described as exceptional six months ago. It should record the reason for a decision, the mandate version in force at the time, and the result that followed.

This is not memory in the sense of an AI retaining fragments from a chat. It is structured decision memory owned by the investor.

The distinction matters. Conversational memory helps an AI agent sound continuous. Decision memory helps an investor remain accountable to a plan.

Where Investboard fits

Investboard connects portfolio context, the investor's own rules and decision memory. Its built-in AI explains available data. Investboard MCP provides a separate authorized connection for supported AI agents to retrieve policy and status, check a stated change and record the investor's stated decision. With portfolio:read, Premium and enabled portfolio sharing, its three portfolio tools also read imported holdings and their values, fund exposures and historical analytics. MCP changes no mandate and executes no order.

The portable investing.md file serves a different purpose: it carries a dated policy snapshot into a conversation. It grants no access and stays unchanged until the investor replaces it. A model's compliance is not guaranteed by the document.

To put this into practice, choose your AI agent and install the four skills. They guide strategy checks, decisions, theses and portfolio analysis with files you provide. The same installer explains the separate OAuth connection when you want supported account data.

The investing second brain preserves the reason alongside the result through supported theses, decisions and later reviews. Thesis and journal retrieval are Premium capabilities. Free active accounts have one MCP connection and the mandate check; Premium adds unlimited connections and supported history access.

The investor remains responsible for every decision. The broker remains responsible for its own execution process and required confirmations. Investboard checks the rules the investor stated, not whether an investment will succeed.

A workflow that preserves the reasoning

1. State the rules. Record the mandate and check the composition date before sharing a policy snapshot.
2. Separate evidence from interpretation. Ask the AI agent to name sources, their dates and missing information. The blank thesis worksheet helps record the thesis, expected observation and a condition that would invalidate it.
3. Check the intended change. With an authorized MCP connection, state every relevant leg and inspect the verdict, coverage and freshness. Missing data remains missing; a check is not an order approval.
4. Record your own decision. State whether you proceeded, abandoned or deferred, and why. Do not let the AI agent invent your reasoning.
5. Review what changed. Revisit the dated thesis and original reasoning at your chosen review date. Keep new evidence distinct from what was known when you decided.

This is an illustrative workflow, not a tested trading result. Its purpose is to keep an investor's rules and earlier reasoning available when conversations, models and markets change.

The rule should outlive the model

AI will make investment research faster. MCP will make financial tools easier to connect. Brokers will continue to expose more capabilities to agents.

None of that gives an AI a stable understanding of what an investor is trying to achieve.

A prompt can tell an agent what you want now. A mandate tells it who it is acting for, which boundaries matter, and which earlier commitments still apply.

Before AI agents become better at trading, investors need a better way to state the rules.