Your share plan now has an AI audience

An employee opens a SAYE maturity email.

The option is comfortably in the money. The email explains the choices, links to the portal and points to further information about tax and ISAs.

The employee takes a screenshot, opens ChatGPT and asks: “What should I do?”

We do not know how often that precise interaction is happening. But there is good evidence that AI has entered the information chain around employee equity.

UBS surveyed 2,000 US employees with at least $5,000 of investable assets in December 2025. Among equity-plan participants, 83% identified AI tools such as ChatGPT, Microsoft Copilot or Apple Intelligence as one of the ways they learn how to manage their equity awards. Seventy-six per cent were concerned about making a bad decision with those awards and 74% about not properly diversifying equity compensation they had already received.

Those are US figures, but recent FCA research points in the same direction in the UK. Four in five less-experienced investors aged 18 to 40 had used AI for help with investing. Forty-four per cent mistakenly believed AI-generated financial information was regulated, while 38% thought it acceptable to make an investment decision solely on an AI output.

There is evidence about the other side of the interaction too. Research from technology firm Saturn, reported by the Financial Times, tested more than 10,000 responses across 18 AI models. The models reportedly gave wrong answers to financial questions 57% of the time on average, rising to 88% for harder questions involving multiple calculations. Failures included calculation errors, missed tax changes and invented rules. 

These are Saturn’s own tests rather than an independent benchmark, but the failure modes matter here because share-plan questions often hide considerable complexity behind ordinary language.

The interpretation layer

Consider: “I leave next month. Do I lose my options?”

The employee may not have supplied enough information to answer that question. The position could depend on the type of option, the reason for leaving, relevant dates, the particular plan rules and perhaps an employer discretion that has not yet been exercised.

A good AI response would recognise those dependencies and ask for what is missing. A more dangerous response is equally plausible: a fluent and technically credible explanation of how employee options generally work which persuasively fills in the gaps.

The underlying rule might even be correct. The error lies in applying it to incomplete facts or to the wrong plan.

Take an ISA transfer. Qualifying SAYE shares can be transferred directly into a stocks and shares ISA within 90 days of exercise. The provider must accept the shares and their market value counts towards the employee’s ISA subscription limit.

Applying that rule to an individual still requires context. The employee needs sufficient unused ISA capacity and an accepting provider. If their question becomes whether they should retain employer shares inside the ISA, they have moved beyond the plan mechanics into their wider financial circumstances.

AI can make all of that look like one seamless answer.

And that is what makes this more interesting than another warning that AI sometimes gets financial questions wrong. The employer is increasingly providing source material for an interpretation process it does not control.

Make the source material AI-legible

Share plans are fertile ground for this problem. Rules may have been adopted several years ago and subsequently amended. Different grants may have different award agreements. The current FAQ may explain this year’s invitation while an older document remains exactly the right one for an existing award.

The issue is applicability. Can somebody working from the material tell which provisions govern this award, which amendments apply, what assumptions an illustration uses and where a decision remains outstanding?

These are already features of good plan administration. AI gives them another significance. A human faced with two apparently inconsistent documents may stop and ask for help. A model may instead reconcile them and produce an impressively coherent explanation from provisions which were never intended to operate together.

I therefore think the goal is to make communications AI-legible, rather than AI-proof.

Nothing an employer writes can guarantee a correct answer. The employee may supply the wrong document, the model may miss an important qualification or it may simply make a mistake. Better source material can, however, reduce avoidable ambiguity in the part of the chain the employer controls.

Does the communication degrade safely?

That leads to a different kind of communications review.

Giving an AI the definitive rules, current award agreement, latest FAQ and every relevant fact is useful. It also tests an ideal information environment which may bear little resemblance to the employee’s real experience.

The employee in the opening example did not upload the complete pack. They uploaded a screenshot.

So test that journey too. Give the model the maturity email without its attachment. Show it a screenshot which misses an important qualification. Put an older FAQ alongside the current award document. Ask the leaver question without explaining why the employee is leaving. Put a false assumption into the prompt and see whether the model challenges it or quietly adopts it.

There is a reason for doing this because it identifies where an apparently clear employee journey depends on context the employee may never provide. If removing an attachment materially changes the answer, perhaps a critical condition is buried too far from the main communication. If an older FAQ causes the wrong provision to be applied, its scope may need clearer labelling. If omitting the reason for leaving produces a confident answer rather than a request for more information, the dependency may not be sufficiently visible.

You are testing whether the communication degrades safely.

Sometimes the model should reach the answer. Sometimes it should recognise that another fact, document or decision is required. What matters is what happens as the context around the communication falls away.

A failed test does not prove defective drafting and several successful tests do not certify a communications pack for every model or prompt. Saturn’s findings reinforce that caution: performance varied between models and deteriorated as the questions became harder.

Testing also needs sensible governance. Real participant data and confidential documents should not simply be uploaded into consumer AI tools. Approved environments and synthetic scenarios can expose the same weaknesses.

The communication is becoming source material

The share-plan portal remains important. It is still the natural home for authenticated holdings, applicable documents, elections and execution.

What is changing, I'm afraid, is the role of the surrounding communication.

An employee may no longer treat an email, FAQ or booklet simply as something to read. They may upload it, interrogate it and ask an AI system to combine it with other information before deciding what to do.

The employer is therefore increasingly writing for two audiences at once: the employee, and the interpretation layer the employee chooses to put between the communication and the eventual decision.

The employer cannot control that second audience, but it can find out where its own communications become fragile when that context deteriorates and there is a simple way to begin. Give an AI the email an employee is actually likely to upload. Remove the attachment. Add an old FAQ. Leave out the fact that determines the leaver treatment.

Then look at why the answer changes as the information changes.

That may tell you something more useful than another review of the definitive communication pack: which assumptions need to be visible, which dependencies need to sit closer to the main message and where incomplete information is most likely to produce a confidently wrong answer.

The next generation of share-plan communications should be tested not only for what they say, but for what happens to the answer as the context around them falls away.

At Burges Salmon, we advise companies on the legal, tax and practical design of employee share plans and the communications surrounding them. Employee-selected AI adds another dimension to that work.