Once a customer has read a message, it has done its work, whatever you send next. A promise of a refund creates an expectation. A wrong balance prompts a decision. A letter to the wrong address discloses something private. A correction can limit the damage, but it cannot remove what the first message started. That is why outbound communication is an irreversible action, even when it looks like the safest thing an agent does.
Why do messages feel reversible when they are not?
Messages are cheap to send and easy to follow up. If one is wrong, another can be sent. Inside the firm, the error may be corrected in minutes, and the case notes updated. From the firm's side, the problem seems fixed.
From the customer's side, nothing has been undone. They read the first message. They may have told someone, moved money, cancelled a plan or relaxed because a problem seemed solved. The correction arrives into a situation the first message created. At best it causes confusion. At worst it reads as the firm going back on its word.
What kinds of harm does a message cause?
The irreversible harm from a message falls into a few distinct kinds. Each calls for a different control.
| Kind of harm | Example | Why a correction does not cure it |
|---|---|---|
| A commitment | "We will refund the fee by Friday" | The customer now expects it; withdrawing feels like a broken promise and may be treated as one |
| A wrong fact the customer acts on | A balance, a due date or an eligibility answer that is wrong | The customer may already have made a decision on it |
| A disclosure | A letter to an old address, an email to a shared inbox | The recipient has read it; privacy cannot be restored |
| Distress | A collections message to someone who has just reported a bereavement | The experience has happened |
| A regulatory communication | A financial promotion or a formal notice sent without approval | It has been issued, with whatever obligations follow |
Commitments are the subtle one. An agent trying to be helpful will reach for reassuring language: "this will be resolved", "you will get your money back". Those phrases are commitments. If the agent has no authority to make them, it has still made them, in the firm's name.
Where is this sharpest?
Some situations make the gap between a message and its correction especially wide.
When a customer has sent money to the wrong account, the first reply shapes everything that follows. Telling them the money will be returned, when recovery depends on another bank and the recipient, sets an expectation the firm may not be able to meet. The approach in how to handle payments sent to the wrong account with an AI agent is careful about exactly this wording.
In collections, contact itself is the irreversible act, and timing and recipient matter as much as content; irreversibility in collections covers this in depth.
In markets where communications are recorded and monitored, every message an agent sends becomes part of the firm's regulated record. Communications surveillance applies to what agents write as much as to what people write.
How does communication compare with other sectors?
In IT changes, a rollback leaves behind anything that happened while the change was live. Communications are the purest form of that: there is nothing to roll back, only consequences. Unlike payments, where recovery is sometimes possible through the receiving bank, there is no recall mechanism for a message that has been read. The wider comparison is in how irreversibility differs between sectors.
How far should an agent be allowed to go?
The answer depends on what the message does, not on the channel.
- Depth 2, draft for a person to send: any message that makes a commitment, explains a decision, responds to a complaint, goes to someone showing signs of vulnerability, or is the first contact on a sensitive matter.
- Depth 3, send within bounds: templated or tightly bounded messages that state facts from sealed sources, make no commitments beyond published policy, and go to verified contact details. Balance alerts, appointment reminders and document requests are typical.
- Never, without a person: messages that waive a right, promise money outside policy, issue a formal notice or constitute a financial promotion.
The test for depth 3 is whether every factual statement in the message comes from a source strong enough to support it, and whether every commitment is one the firm has already made in policy.
What controls make agent messages safer?
- Verify the recipient at the moment of sending, not when the conversation began, because contact details change.
- Check every factual claim against its source and refuse to send if a load-bearing fact is external or stale.
- Scan for commitment language and hold any message that promises an outcome, a time or an amount outside the agent's authority.
- Check vulnerability and contact rules before each outbound message.
- File every message with its basis: the facts it relied on, their provenance, and whether a person approved it.
Where Pryme fits
In Pryme, sending a message is a capability with its own depth, separate from drafting one. The minimum chain checks the recipient, the provenance of every fact and the agent's authority to make any commitment at the moment of sending. Messages outside bounds are held for a named person. See why autonomy in regulated industries is a question about reversal.
Questions people ask
Can an agent send a correction automatically?
It can draft one immediately. Whether to send it, and how to word it, should be decided by a person, because a clumsy correction can cause more harm than the original error.
Are chat replies different from emails or letters?
The channel changes speed, not irreversibility. A chat reply is read within seconds, which leaves even less time to catch an error, so the same bounds apply.
Should an agent apologise on the firm's behalf?
A plain apology for inconvenience is usually fine. An apology that admits fault or implies compensation is a commitment, and should go through a person.