Can AI Run the Inbox Without Annoying the Customers?
Answering email is easy. Knowing what the customer actually wants is much harder.
For many small businesses, the inbox is effectively the front door.
A new customer may not telephone. They may never visit the premises. Their first contact with the business might simply be an email asking:
“Do you teach A Level Physics?”
“Could you produce a short promotional video for our company?”
“My daughter is struggling with GCSE Maths. Do you have any availability?”
“How much would it cost to photograph our products?”
The quality and speed of the reply can determine whether that enquiry develops into a customer or disappears.
But answering email takes time.
Could artificial intelligence take over some of this work?
I think it can — but there is an important distinction between AI helping to run an inbox and AI being allowed to communicate unsupervised with customers.
Those are not necessarily the same thing.
The Inbox Is More Complicated Than It Looks
At first sight, email administration appears straightforward:
Read the message.
Work out what it is about.
Write a reply.
Send it.
The difficulty lies largely in Step 2.
Consider these messages:
“Hi Philip, just checking whether Thursday is still OK?”
“My son is in Year 12 and is really struggling with Physics. His school has suggested he might need some extra help. Could you let me know what you offer?”
“We've looked at the first version of the film. Everything is fine except the interview at the beginning. Could that section be changed before Friday?”
All three are emails, but they require completely different actions.
The first needs context. Which Thursday? What was previously agreed?
The second is potentially a new customer and needs a helpful, reassuring and informative response.
The third may be a production deadline and therefore potentially urgent.
A conventional automated system might recognise a few keywords.
A capable AI system can attempt something more useful: interpretation.
The First Useful Job: Sorting the Inbox
I would probably not start by asking AI to answer every email automatically.
I would start by asking it to organise them.
Imagine opening the business inbox in the morning and instead of seeing 37 unread messages, seeing something more like:
URGENT
Production client needs an amendment before Friday.
NEW ENQUIRIES
Parent asking about Year 12 Physics tuition.
Business requesting quotation for product photography.
ACTION REQUIRED
Existing student requesting a change of lesson time.
Supplier asking for confirmation of an order.
WAITING FOR CUSTOMER
Filming client has not yet supplied the logo files requested last week.
INFORMATION ONLY
Order confirmation.
Newsletter.
Automated software notification.
That could immediately make the inbox more manageable.
More importantly, AI could potentially identify emails that have quietly disappeared down the list but still require action.
That is often where the real value lies.
Finding the Message I Forgot to Answer
One of the most useful questions AI could answer is:
“Which emails still require something from me?”
That is subtly different from finding unread messages.
I may have read an email without replying.
I may have replied asking the customer a question.
The customer may have answered, meaning the conversation has returned to me.
A client may have asked three questions and I may only have answered two.
A parent may have enquired about tuition, received an initial reply and then asked about available lesson times.
The interesting problem is therefore not:
Is this email unread?
It is:
Whose turn is it to act?
That requires understanding the conversation rather than simply examining the status of the latest message.
AI as an Inbox Detective
Suppose a production customer writes:
“Thanks, that looks excellent. Could we have the final version in both 16:9 for the website and vertical format for social media? Also, could you change the telephone number on the final screen? We need everything by Wednesday.”
There are at least three actions hidden inside that short message:
produce a 16:9 version;
produce a vertical version;
change the telephone number.
There is also a deadline: Wednesday.
A useful AI system could extract those tasks and perhaps present them as:
Client: XYZ Ltd
Project: Promotional film
Actions: Create 16:9 export; create vertical export; amend telephone number
Deadline: Wednesday
Status: Action required
That starts to blur the distinction between an email system and a business-management system.
And that is where AI becomes particularly interesting.
Tuition Enquiries Are a Good Test
Tuition enquiries provide an excellent example because the messages are often quite conversational.
A parent might write:
“My daughter is starting Year 13 and has been finding Chemistry increasingly difficult. She did reasonably well at GCSE but has lost confidence during the first year of A Level. She is doing OCR and would prefer face-to-face lessons if possible. Do you have anything available?”
A useful system should be able to identify:
Student: Year 13
Subject: Chemistry
Exam level: A Level
Board: OCR
Problem: Difficulty/confidence
Preference: Face-to-face tuition
Customer wants to know: Availability
That is far more useful than simply labelling the email “Education”.
The AI could then prepare a draft response based on the services actually offered.
But this is where human judgement becomes important.
Drafting Is Different From Sending
I would be much more comfortable allowing AI to draft a reply than allowing it automatically to send one.
For example, AI could prepare:
“Thank you for getting in touch. I teach A Level Chemistry and can provide face-to-face tuition. Lessons can include syllabus work, exam technique and practical science where appropriate. I currently have…”
I can then read it, correct anything necessary and press Send.
That may save most of the writing time while retaining human responsibility for what actually reaches the customer.
This is particularly important because AI can produce language that sounds extremely confident even when it has misunderstood something.
A polished wrong answer is still a wrong answer.
Consistency Without Sounding Like a Robot
There is another potential advantage.
Small businesses often answer similar questions repeatedly:
“How much do lessons cost?”
“Where are you based?”
“Do you teach online?”
“Can you film on location?”
“Can you photograph small products?”
“What subjects do you teach?”
AI could use approved business information to prepare consistent answers.
That reduces the risk of accidentally giving one customer an old price, another an outdated availability slot and somebody else information copied from an obsolete email.
However, consistency should not mean that every customer receives the same impersonal paragraph.
Compare:
“Thank you for your enquiry. We provide GCSE Mathematics tuition. Please see our website for further information.”
with:
“Thank you for getting in touch. From what you've described, it sounds as though your son understands much of the Maths but is losing marks when applying it to longer exam questions. That is something we can specifically work on during tuition.”
The second response demonstrates that somebody has actually understood the enquiry.
AI can help create that personalisation — provided it is given enough context and its response is checked.
Recognising Urgency Is Harder Than Looking for the Word “Urgent”
An email does not have to contain “URGENT” in capital letters to be urgent.
“Just checking you're still coming to film tomorrow morning.”
That is probably important.
“We've noticed the venue address on the call sheet is wrong.”
Very important if filming is tomorrow.
“My exam is next week and I wondered whether you have any availability before then.”
Time-sensitive.
AI could potentially combine language, dates, deadlines and previous conversations to assign priority.
But again, I would be cautious about allowing AI to make the final decision.
A long-standing customer asking a small question may deserve attention before an apparently more important automated message.
Relationships are difficult to reduce to an algorithm.
What About Chasing People?
This is another area where AI could be extremely useful.
Imagine asking:
“Show me all enquiries from the last 14 days where the potential customer has not replied.”
The system might find five.
It could then prepare a polite follow-up:
“I just wanted to check that you received my previous message regarding A Level Physics tuition. Please feel free to get in touch if you would like to discuss availability or have any further questions.”
For a production business, it might identify that a quotation was sent seven days ago but there has been no response.
However, automated chasing needs restraint.
Nobody wants:
DAY 2: Just checking in…
DAY 4: Did you see my last email?
DAY 6: I haven't heard from you…
DAY 8: Final reminder…
That is an excellent way for AI to make a business remarkably efficient at annoying its customers.
AI should help maintain relationships, not relentlessly pursue them.
The Importance of Tone
A tuition enquiry from a worried parent should not sound like a corporate sales email.
A quotation for a commercial filming project should probably be more formal.
A message to an established customer might be considerably more relaxed.
Therefore an AI inbox assistant needs more than factual information. It needs guidelines about tone.
For my businesses, I might specify:
friendly but professional;
clear rather than overly formal;
avoid unnecessary jargon;
answer the customer's actual questions;
do not make promises about availability without checking;
do not invent prices or services;
do not pressure customers;
identify anything requiring personal judgement.
Those rules become part of the business's communication policy.
AI Should Know When It Does Not Know
Perhaps the most valuable behaviour an AI system can learn is:
“I am not sufficiently certain to answer this.”
Consider:
“Could you guarantee that my daughter will get a Grade 8?”
The correct response requires care.
Or:
“Can you definitely finish our promotional film by Monday if we send the material on Friday?”
That depends on workload and the complexity of the project.
Or simply:
“Can we move Tuesday's lesson to Thursday?”
That sounds trivial until we realise the AI needs access to the diary before promising anything.
In each case, the correct automated action may be:
Human decision required.
That is not a failure of AI.
It is good system design.
A Practical Experiment
A very revealing experiment would be to give an AI assistant a fictional day's inbox containing perhaps 20 messages:
three new tuition enquiries;
two existing students changing lessons;
one production deadline;
a supplier invoice;
several newsletters;
a customer asking for a quotation;
an unanswered question from last week;
a parent confirming a lesson;
a spam message;
an automated receipt;
a customer complaint;
a filming client requesting changes.
Then ask the AI to produce five things:
What needs attention immediately?
What requires a reply today?
Which messages are simply informational?
Which conversations are waiting for somebody else?
Which replies can be safely drafted, but should require human approval?
I suspect this would tell us much more about the usefulness of AI than simply asking it to write an email.
Could AI Eventually Send Some Messages Automatically?
Possibly.
There are certain low-risk messages where automation may make sense:
“Thank you. I have received the files.”
“Your message has been received and I will reply shortly.”
“Here is the information you requested.”
But I would introduce automation gradually.
A sensible progression might be:
Stage 1 — Observe
AI analyses the inbox but takes no action.
Stage 2 — Organise
AI categorises messages, identifies tasks and highlights priorities.
Stage 3 — Draft
AI prepares suggested replies for human approval.
Stage 4 — Limited automation
AI sends only tightly defined, low-risk responses.
Stage 5 — Review
Regularly examine what it is doing well and where it is making mistakes.
That approach allows trust to be earned rather than assumed.
There Is Also a Data Question
An inbox can contain personal information, quotations, addresses, financial details, student information and commercially sensitive material.
Therefore the question should never simply be:
“Can this AI read my email?”
It should also be:
“Should this AI have access to this information, how is that information handled, and what controls are in place?”
Any business using AI with customer communications needs to think carefully about privacy, security, access permissions and data protection.
Convenience should not automatically override confidentiality.
What I Actually Want From an AI Inbox Assistant
I do not really want an artificial intelligence that impersonates me.
I want something closer to an extremely capable assistant sitting beside me saying:
“You need to answer this one first.”
“This customer asked two questions and you only answered one.”
“You sent that quotation a week ago and have heard nothing.”
“This parent is asking about A Level Biology and prefers an evening lesson.”
“This production customer has changed the deadline.”
“You promised to send those files yesterday.”
“I've drafted replies to these six messages. Would you like to check them?”
That is considerably more useful than simply generating paragraphs of text.
AI Should Reduce Administration, Not Remove the Relationship
This is the central issue.
Customers do not necessarily object to businesses using technology.
They object when technology makes dealing with the business harder.
Nobody wants to explain the same problem three times because an automated system cannot understand the conversation.
Nobody wants a cheerful automated response to a serious complaint.
Nobody wants to receive a completely irrelevant follow-up because an algorithm misunderstood an earlier message.
The test for AI should therefore be simple:
Does this make the customer's experience better?
If AI helps me respond more quickly, remember important details, avoid missing enquiries and provide better information, then it is useful.
If it merely allows me to send more automated emails, it probably is not.
Conclusion — The Best AI Inbox May Be the One Customers Never Notice
The interesting future of AI in business email is not necessarily a robot answering every message.
It may be an invisible layer of assistance behind the person running the business.
It reads.
It sorts.
It identifies.
It remembers.
It drafts.
It reminds.
But when judgement, empathy, negotiation or responsibility are required, the human remains involved.
For a small business, that could be enormously valuable. The owner gets some of the organisational benefits of having an additional office assistant without surrendering the personal communication that helped build the business in the first place.
The ultimate measure of success is not how many emails AI can answer.
It is whether customers receive faster, more accurate and more thoughtful service.
Perhaps the best compliment for an AI-powered inbox would therefore be that the customer never realises AI was involved at all.
They simply think:
“That business is remarkably good at replying.”
No comments:
Post a Comment