AI in customer services: a draft on every reply

The bit that keeps growing

If you've grown past a certain size, your customer-services inbox is one of the bigger quiet costs in the business.

It rarely shows up as a line on the P&L, because it's spread across people: the warehouse manager who answers "where's my order?", the bookkeeper who fields "can I have a copy of invoice 4023?", the founder who still gets cc'd on the angry ones. Add it up across a year and it's usually somewhere between two and five FTEs of work nobody actually planned for.

This is one of the places AI earns its keep most quickly. Not by replacing the human — that approach goes badly the first time the model gets something wrong — but by doing the boring 80% of the work and putting a draft in front of a person who can sense-check it in seconds.

What it actually looks like

Picture a real morning in your inbox.

A customer emails:

"Hi, I ordered a replacement filter last Tuesday and haven't seen anything yet. Can you let me know what's happening?"

Today, that goes into the queue. Someone has to:

1. Find the customer in your CRM.
2. Find the order in your fulfilment system.
3. Check the courier portal for the tracking status.
4. See if this customer has had similar issues before.
5. Write a reply that's accurate, on-brand, and the right kind of apologetic for whose fault it actually was.

Half a day's worth of that, every day. Across a team. Mostly looking things up.

With AI in the loop, the same email arrives — and by the time the customer-services person opens it, there's already a draft reply waiting in their inbox tool that says something like:

"Hi Mark, thanks for chasing — your order #18432 (replacement filter) went out with DPD last Wednesday. Tracking shows it's been sitting at the Coventry depot since Friday with a 'failed delivery attempt' note. I've asked them to redeliver tomorrow before noon — link below. Apologies for the runaround, let me know if it doesn't land."

The agent reads it, nods, edits one sentence, hits send. Total time: forty seconds.

The same query, last month, was twenty minutes of tab-flipping.

Where MCP comes in

The reason that draft can be that specific is the recently-emerged piece of plumbing that's making a lot of this finally practical: MCP servers.

In plain terms: an MCP server is a standard, controlled way to give an AI safe access to your business systems. Your CRM. Your order management. Your courier portal. Your knowledge base of past tickets. Your invoice records.

The AI doesn't get a wide-open key to everything. It gets a defined menu of things it's allowed to ask — "look up customer by email", "get status for order ID", "search past tickets for similar wording" — and it can only ask those specific questions. Every call is logged. You can audit exactly what was looked at, when, and why.

That matters because the difference between "AI is generally helpful but often wrong about my business" and "AI gives me genuinely useful drafts about my customers" is whether it can see your actual data. MCP is what makes that possible without handing over the keys.

In a typical customer-services setup, we'd usually wire up:

• Your CRM — to look up the customer, their account history, and any current notes
• Your order or fulfilment system — to pull live status on whatever they're asking about
• Your courier or supplier APIs — to get tracking and delivery info in real time
• Your past tickets — to spot whether this question has been asked, and answered well, before
• Your knowledge base or product docs — to draw on the actual product information
• Your tone-of-voice guidelines — so the draft sounds like you, not a chatbot

The AI does all of that lookup quietly in the background, then composes the draft from everything it found. To the agent, it just looks like their inbox already has a sensible answer ready.

The human is always in the loop

This is the part we deliberately don't compromise on, and most of our clients agree quickly once we've talked it through.

Drafts go to a person. The person reviews, edits if needed, sends. They can also reject the draft entirely and write their own — and the system learns from those rejections over time, getting better at the kinds of replies you actually wanted.

The reasons we keep a human in the loop:

• Liability. A wrong reply about a delivery slot is annoying. A wrong reply about a refund, a contract clause, or a safety question is a real problem.
• Tone. Customers can almost always tell when they've been answered by a machine. They can rarely tell when a person used a draft.
• Edge cases. About 5% of inbound messages are genuinely unusual — angry, confused, or about something the AI doesn't quite understand. Routing those cleanly to a human is more important than processing 100% automatically.
• Trust. Internally, your team will use a system more if they're in charge of it than if it's "doing things behind their back".

The AI is doing the looking-up. The person is still doing the customer service.

What changes for the team

The job doesn't go away. It changes shape.

A customer-services person handling AI-assisted drafts can typically work through three to five times the volume per hour. But more importantly, they spend their time on the bits where their judgement actually matters: the angry email, the unusual request, the customer who needs a phone call. The "where's my order?" tide drops to a manageable trickle.

It also tends to surface patterns. Once every reply is being routed through a system that can count things, you start seeing — "45% of our inbox is about delivery delays for this one product line", or "we've answered the same VAT-on-export question 200 times this year". That information used to be invisible. Now it tells you what to fix at the source.

AI in customer services isn't really about answering more emails. It's about getting the same person to spend their day on the half-dozen emails that actually need a person.

Where to start

This isn't a six-month project. A first useful version usually looks like:

• Pick one inbox — typically the one with the highest volume of repetitive lookups
• Identify the three or four systems an answer normally needs to come from
• Wire up MCP access to those systems, with proper permissions and audit logging
• Put drafts in front of the team in the tool they already use — your helpdesk, Outlook, Gmail, wherever the work happens
• Measure reply time, edit rate, send rate, and customer satisfaction. Tune from there

A working version of this for a single inbox is usually four to six weeks of work. Payback is typically inside a quarter, sometimes faster if the inbox is large.

If your customer-services queue is the kind of thing your team complains about on Mondays, this is one of the better-shaped places to put a small, careful AI system to work — with humans firmly still in charge.

If you'd like to walk through what that might look like for your inbox, say hello.