Agentic AI and employee share plans: who decides when the AI trades?

It has considered the share price, your tax position, the rest of your portfolio and the concentration of your wealth in one company. Economically, the recommendation makes sense.

There is a problem. You have inside information. A commercially sensible recommendation does not mean that the trade can lawfully proceed.

Robinhood, the US retail investment platform, says that more than 150,000 customers have opened agentic trading accounts since May. Its forthcoming Loops feature will allow standing strategies to operate in the background and execute trades when specified conditions are met.

It is easy to imagine the same technology reaching employee equity. An agent could monitor option expiry dates, tax, share prices, dealing restrictions and concentration risk, then translate a broad financial objective into transactions.

But automated dealing is not new. Before MAR, the FCA’s Model Code expressly contemplated trading plans based on fixed terms, discretion given to an independent third party or a written formula, algorithm or computer program.

The current regime is different, but that history is useful. The novelty is not the use of an algorithm because we already see the same basic idea in employee share plans. PDMRs may acquire shares during closed periods under standing instructions established beforehand. The transaction happens later, but the relevant decision was made earlier.

The law and market practice are therefore familiar with automated execution. Agentic AI becomes more difficult when the machine is not simply implementing a settled instruction, but continues to exercise investment judgement under a broad mandate.

The agent monitors a £1 million portfolio, including £250,000 of employer shares. The executive transfers £200,000 of cash into an account outside the portfolio being monitored. The shares have not moved in value, but they now exceed 25% of the measured portfolio, so the agent sells.

The executive has not changed the trading instruction. They have changed an input that determines whether it executes. That does not itself mean that insider dealing has occurred. The cash might have been moved for an unrelated reason.

Now change the facts. The executive learns that next week’s results will be materially worse than the market expects and deliberately moves the cash because they know that doing so will trigger a sale.

Nobody typed “sell”. The algorithm followed its existing rule. But the executive may still have used inside information by manipulating the inputs that caused the trade.

Even this is not uniquely agentic. A conventional formula-based trading plan could create the same problem. The more difficult case is a mandate such as:

Manage my exposure to my employer prudently, taking account of my tax position, liquidity needs and the rest of my portfolio.

The employee does not therefore need personally to press the sell button. The question is whether inside information has influenced the transaction.

UK MAR does recognise that some transactions arise from obligations created before a person obtained inside information. In certain circumstances, completing such a transaction will not, merely because the person now has inside information, amount to using that information.

But this is narrower than showing that the employee had previously intended to sell. The transaction must generally discharge a genuine obligation arising from an earlier order or agreement.

A standing instruction to an AI may not meet that test. Telling an agent “sell if my exposure gets too high” may leave the employee free to change or withdraw the instruction. An order already lodged with a broker may be more firmly committed. If genuine investment discretion has been transferred to an independent person, the later decision may no longer be the employee’s decision at all.

The answer therefore depends on the legal and practical structure. Relevant questions include:

• Has the employee assumed a binding obligation?
• Can the instruction be changed or withdrawn?
• Who decides whether, when and how much to trade?
• What influence does the employee retain?
• What information can reach the agent or other decision-maker?

There is also a timing issue. Preventing an agent from creating a future order is not necessarily the same as cancelling an order already lodged with a broker. The critical control point may therefore arise before execution: when the agent is established, its mandate is changed, new data is made available or the employee gains fresh influence over its decisions.

Share plans already involve substantial automation. Awards can vest without a fresh investment decision. Standing arrangements can fund tax liabilities. Employee saving plans can generate regular purchases without a new decision each month.

The closed-period rules also recognise the importance of advance planning and retained discretion. They contain specific, conditional permissions for certain employee-plan transactions. Particular option exercises and related sales may be permitted where conditions including advance notification, an irrevocable decision and issuer approval are satisfied. Savings arrangements can also qualify where participation was established beforehand and the PDMR cannot alter the purchases during the closed period.

The point is not that automated transactions are generally exempt. It is that the rules distinguish between an arrangement settled in advance and one over which the PDMR retains later discretion.

That distinction has an obvious share-plan application.

An award might vest automatically. A standing instruction to sell enough shares to meet the tax liability is familiar territory. An agent that decides at vesting how many shares to sell after considering the employee’s available cash, tax position, the share price and their wider portfolio is doing something different.

The first arrangement is mainly execution. The second involves continuing investment judgement.

Company clearance and legal permission must also be kept separate. UK MAR does not create a universal company clearance process for every transaction. A dealing code may impose its own requirements, but company clearance does not settle whether the transaction complies with the insider-dealing rules.

Conversely, a transaction that is not insider dealing may still be prohibited for a PDMR during an Article 19 closed period. Article 19 also imposes separate notification requirements on PDMRs and persons closely associated with them.

And the enforcement consequences are very real. In 2024, the FCA fined former Wizz Air executive András Sebők £123,500 for breaches of the PDMR regime, including trading during closed periods and failing to notify transactions.

No AI was involved, which is why the case matters. An agent must still operate within several existing layers of company and regulatory control. It cannot replace them with one universal permission switch.

The US Rule 10b5-1 regime provides a useful comparison. It can accommodate fixed instructions or formulae, as well as discretion delegated to another person who does not possess material non-public information, provided the insider retains no subsequent influence over whether, when or how trades occur.

That is not a UK safe harbour, but it illustrates the architectural point. The fact that a machine continues to exercise judgement is not necessarily the problem. The important questions are whose judgement it represents, who can influence it and what information reaches it.

Information architecture matters as much as the mandate. A CFO’s working AI may have access to board papers, forecasts and M&A discussions. That information should not flow into an autonomous investment agent. Using separate AI products will not be enough if they share memory, credentials, data stores or integrations.

Segregation can reduce the risk, but it cannot remove the executive’s own knowledge or prevent the executive from influencing the agent through other means.

There is also a financial-services boundary. Managing another person’s investments while exercising discretion is a specified regulated activity. Employee share plan exclusions are targeted, not a general permission for an employer to become an investment manager.

The employer’s role may therefore need to remain narrow. It can provide accurate plan information, dealing restrictions and clearance outcomes, while investment management and execution sit elsewhere.

For listed companies, an agent-ready share plan requires more than adding AI wording to the dealing code. The practical exercise will likely be to map:

• the mandate given to the agent;
• the information it can access;
• where discretion sits;
• the legal and issuer restrictions that apply;
• the route to execution; and
• the reporting that follows.

Reporting is also easily overlooked. An autonomous transaction may occur without the employee touching the platform and still trigger PDMR notification obligations. Agent-readiness is therefore not just about preventing the wrong trade, it is also about recognising that a transaction has occurred and feeding an exercise, acquisition or disposal into the correct reporting process.

Companies will also need to identify which later changes require the arrangement to be reviewed. Changing the mandate, widening the agent’s data access, altering its parameters or giving the employee fresh influence over a supposedly independent strategy may matter as much as changing a conventional broker instruction.

The law is already familiar with automation. What changes is the amount of judgement that can sit between an employee’s broad objective and the eventual trade and, because the technology is so new, the underlying question is probably not:

Who pressed the button?

It is:

Whose judgement caused the trade, what information influenced it and who remained able to intervene?