AI suggested replies: a draft answer waiting for every customer message
How AI suggested replies work in email, helpdesk and chat: the AI reads the thread, the order and your rules, drafts the reply, and a person approves it.
Vlado Pandžić · Founder · Senior .NET architect
Published · 5 min read
Monday morning, the shared inbox has sixty new messages. Most of them are the same five kinds: where is my order, the delivery arrived damaged, can we change the delivery date, please send a quote, I need a copy of the delivery note.
Every reply means the same routine: open the order system, find the order, check the courier, look up the rules, and write a polite answer. Ten or fifteen minutes each, and by Wednesday the customers who wrote on Monday are writing again.
AI suggested replies change one thing: when someone opens a message, a draft answer is already waiting in the reply box, with the facts filled in. The person reads it, adjusts it if needed, and sends it.
What it looks like
Hello,
the package arrived today and two of the six glasses are broken. Photo attached. What happens now?
Laura
Proposed: 2 replacement glasses, shipped Thursday, no return needed
Dear Ms Carter,
we are sorry the glasses arrived broken. We have checked order 20417, and we will send two replacements on Thursday at no cost. There is no need to return the broken ones.
Kind regards, Mark
Sources: order 20417 · delivery note · returns policy, damage in transit · awaiting approval
The AI did not invent the solution. It found the order, saw that the delivery was confirmed, and applied the company’s own rule for damage in transit. If the rule had said “credit note above €100 needs a manager”, the draft would have said so and waited for the manager.
Which messages it helps with
| Message | What the AI looks up | What the person decides |
|---|---|---|
| Where is my order? | Order status, courier tracking | Usually nothing: check and send |
| Damaged or wrong delivery | Order, delivery note, photo, returns policy | Replacement or credit above an agreed limit |
| Change of delivery date or address | Order and shipping status | Whether the change is still possible |
| Quote request | Price list, stock, standard discounts | The price, if outside standard terms |
| Complaint about the service | The customer’s history | Tone, compensation, escalation |
The first three are often the majority of all messages, and they are also where the suggestions save the most, because the answer depends almost entirely on facts from your systems.
Built-in suggestions, or connected to your systems
Helpdesk and email tools such as Zendesk increasingly offer AI suggestions out of the box. They are useful for general answers from a knowledge base: opening hours, return conditions, how to reset a password.
But they do not see your order system, your stock, your prices or the delivery status, unless someone connects them. And that is exactly where the value is: a suggestion that says “we will send two replacements on Thursday” saves ten minutes, while “we are sorry to hear that, we will look into it” saves nothing.
Connecting them is a small, well-defined piece of software: a service that, for each new message, fetches the relevant order, delivery and rules through your existing systems, and hands the draft back to the inbox or helpdesk. If the messages are handled in your own application, the same thing can live directly in its screens, as described in the article on an AI assistant inside a Blazor application.
How it works behind the scenes
- 1A message arrives
- 2The AI recognises the type and the order
- 3Your systems provide the facts and rules
- 4A draft with sources waits in the reply box
- 5A person edits and sends
The model does not learn from your messages on its own. It gets better because you improve its instructions and examples, for instance by looking once a month at the drafts people changed the most.
What stays with people
- Nothing goes out without a click. The AI prepares, and a person sends.
- Money above a limit. Credits, discounts and compensation above an agreed amount go to whoever is allowed to decide.
- Escalations and legal matters. Angry customers, threats of legal action and anything sensitive are flagged, not answered automatically.
- Key customers. For the most important customers, the tone is set by the people who know them.
- “I don’t know” is allowed. If the AI cannot find the order or the rule, the draft says so instead of guessing.
How to start
Pick the most common type of message, which is usually “where is my order”. For two weeks, measure how long a reply takes today. Then switch on suggestions for that type only, and compare. If the time per reply halves, add the next type.
Speed matters even more for sales enquiries, where a slow reply means a lost customer, as we covered in the article on answering enquiries faster.
How we work
This is exactly what we do: we bring AI into business processes, including suggested replies connected to your order system, stock and rules, with a person approving every message. We start with one type of message and a measurable saving. More on our AI work, and the first step is a free 30-minute call.
Sources
- Data, privacy and security for Azure OpenAI, Microsoft Learn
- Microsoft.Extensions.AI libraries, Microsoft Learn
This article is general information only, not legal, tax, financial or other professional advice. Scenarios, examples and calculations are illustrative. Terms of use and disclaimer.