AI that suggests or AI that replies alone on WhatsApp: how to decide by conversation type

When the AI suggests, it drafts a reply and a person reads it, edits it if needed and decides whether it goes out; when it replies alone, the message goes straight to the customer without anyone in between. The first setup is called copilot mode: the AI proposes, the person decides. The second is an autonomous agent: it answers on its own and hands off to a person when it detects it should.

It is the same technology with a different level of control. The market usually describes a copilot as an assistant that suggests replies without taking over the conversation, which is a good way to fix the idea: your team keeps the last word and the AI saves them from writing from scratch. For the full definition inside heysteff, see what copilot mode is.

The key point is that this is not a one-time, all-or-nothing choice. You can run a conversation about opening hours on automatic and one about a refund in copilot on the same WhatsApp number, and change each one once you have evidence of how it behaves.

Copilot mode makes sense when a mistake costs more than the delay of someone reviewing the message. There are three clear signals: the message commits money or an exception, the customer is upset, or you are not yet sure what the AI knows well.

The first signal is commitments. A refund, a discount outside policy, a delivery date promise or a warranty condition are replies your team wants to see before they go out, even if the AI words them well. The second signal is tone: a customer who writes angrily or has already complained twice needs a reply that reads the emotional context, and there the AI's suggestion is a good draft that a person makes more human.

The third signal is the stage you are in. In the first weeks, when the knowledge base has just been loaded, copilot works as a trial period: you see what the AI suggests, fix what is wrong and find the gaps before a customer does. If you have not built that base yet, start with the heysteff knowledge base.

The AI can reply alone when the question is repetitive, the answer is objective and verifiable, and a mistake is cheap and reversible. Opening hours, location, order status, product availability, requirements for a procedure or how to use something already documented all meet those three conditions.

There is a second argument in favor, and it is time. A customer who writes after hours or during a message spike does not wait for someone to approve a suggestion; for a simple query, what they value is an immediate answer. In those cases requiring human approval cancels the very benefit the AI was brought in for. To see why speed matters so much, read why first response time matters.

Replying alone does not mean replying without limits. Even on automatic, the AI should operate inside escalation rules: if the customer asks for a person, if a certain keyword appears or if the conversation enters a sensitive topic, the chat moves to a human. You set that up with human escalation rules.

The simplest way is to score each conversation type with three questions: how repetitive is it, how expensive is a mistake and how easy is it to fix afterwards. A highly repetitive conversation with a cheap, easily corrected error goes on automatic; a rare one with an expensive or irreversible error goes in copilot.

Applied to a real case: asking whether a size is in stock is repetitive, the error is cheap and a second message fixes it, so it goes on automatic. Promising a delivery date to a hard-to-reach area is less repetitive and a mistake creates a complaint, so it goes in copilot. A price negotiation with a wholesale customer is rare and expensive, so a person probably handles it from the start, with the AI only helping to summarize the history.

One detail that is often forgotten: the rule can depend on channel and moment. An Instagram message about a post tends to be more casual and exploratory than a WhatsApp message from someone who already bought, and a launch with heavy traffic changes what your team can approve by hand. Revisit the rule when the context changes, not only when something goes wrong.

Picture an online clothing store that gets five types of messages: size and stock questions, order tracking, exchanges and returns, complaints about delays and wholesale requests. With the rule above, the first two categories are answered by the AI alone, because they are repetitive and can be checked against the catalog and the orders.

Exchanges and returns start in copilot: the AI drafts the reply using the store's policy and a person approves it. After a few weeks, if the team sees it approves the simple cases with almost no edits, it moves those within the return window to automatic and keeps only the exceptions in copilot. Delay complaints always stay with a person in charge and the AI as the drafter, and wholesale requests are handed off right away.

Notice that the store did not switch the whole AI on or off: it moved categories one at a time, based on evidence of what the team was correcting. That gradual adjustment is what avoids both the runaway bot and the overloaded team approving messages that never needed review.

The most common mistake from too much control is leaving everything in copilot forever. The team ends up approving dozens of identical suggestions, approval turns into a chore nobody reads closely, and you lose both the speed and the real oversight, because a person approving out of fatigue is not supervising anything.

The opposite mistake is releasing everything to automatic on day one. If the AI answers alone about policies you never reviewed, the first wrong message is seen by a customer and not by your team. Between the two extremes, a third mistake is not measuring: if nobody looks at how many suggestions are edited and how many are approved as they are, there is no way to know when a category is ready to move to automatic.

Reviewing conversations regularly closes that loop. For the product side, the heysteff copilot mode page shows how it is turned on, and the AI customer support overview puts it in the context of the whole service.

Does copilot mode slow down the service? It adds the time for a person to review, but saves the time of writing from scratch. That is why it is best kept for conversations where a mistake costs more than that delay, leaving repetitive ones on automatic.

Can I combine copilot and automatic replies on the same number? Yes. The usual setup gives each conversation type its own autonomy level, for example automatic for stock questions and copilot for refunds, with escalation rules so any chat can move to a person.

When should a category move from copilot to automatic? When your team approves almost all of that category's suggestions without editing them over a reasonable period and the remaining errors are low cost. If corrections stay frequent, the cause is usually the knowledge base, not the autonomy level.

• heysteff copilot mode
• What is copilot mode
• How to set human escalation rules
• AI to human handoff on WhatsApp without repeating