The $37K Email Summary

Today I want to share an example of how a profitable AI case develops in customer support. This isn't an actual case study, but more of a realistic example based on what I see repeatedly in my work with clients.

The starting point:

"We have 10 people doing customer support. AI can write emails better than we can. So let's build a solution that does the work for us."

Sensible approach?

Let’s find out!

We're in a B2B service business. A team of 10 people handles client support, most of it via email. Everyone works out of a shared support@ inbox in Outlook that has evolved its own rituals over time – flags for important emails, flags for messages that need further digging, flags for emails currently being processed by someone.

It mostly works.

The main problem happens when an email thread goes back and forth and the original handler is out of office or on leave. Someone else has to dig into the conversation history, figure out what's happening, and make sure the issue still moves toward resolution. On a busy Monday, this can eat an hour before anyone has actually responded to anything.

Everyone knows that at some point they'd need a proper ticketing system. But that’s a story for another day.

Until someone heard "AI".

Nobody in the organization wants AI to handle the entire customer interaction. At least not yet. What people want is a little helper – something that brings them up to speed quickly, especially on longer threads and complicated cases.

For this, the AI would need to be able to read the full thread and figure out what's going on. Instead of reading 15 messages, the agent reads 4 sentences.

If that works, the next step is obvious: since the AI has a good understanding of the problem, why not also draft a reply? Look up relevant information from the company's support documentation, combine it with the thread context, and suggest a response.

Best case: the agent quickly reviews the draft and hits send. Worst case: they scrap the suggestion and write their own. Realistically, it's somewhere in the middle – the draft needs some edits, but the structure, context, and relevant product info are already there.

Naturally, the team looks at solving this with a tool they already have – Microsoft's Copilot. It works in Outlook natively, it's already in the Microsoft ecosystem, and the cost seems reasonable. So why not give it a try.

The problem is that Copilot out of the box doesn't work well for shared inboxes. Besides having each of the 10 agents need their own M365 Copilot license – roughly $20/user/month = ~$2,400/year, the main problem is that you would need to customize the Copilot solution for each agent (since it’s Productivity AI). How do you ensure 10 individual Copilots produce consistent summaries? You can't. Besides that, waiting each time for the Copilot to pass the email thread and formulate the reply feels slow.

What they need isn't a personal assistant for each human agent. It's one central AI workflow that serves the whole team consistently.

The custom version would roughly work like this:

No auto-sending or customer-facing AI yet. Just a behind-the-scenes assistant that reads, researches, and prepares — so the agent can focus on the response. And the agent would automatically see the drafted reply the moment they see the email in the shared inbox.

If this works, the team estimates that each agent could save roughly 20 minutes per email on complex threads. Given the current volume, that would be about 2.5 hours per agent per week.

$37,500 – that's the number you saw in the title.

But here's the problem: we can't say where those 2.5 hours go. People don't save 2.5 hours and then do 2.5 hours of measurable extra output. The time will likely just dissolve into the day. So $37,500 is the ceiling. Not the business case.

Instead of arguing about the exact savings, there's a simpler way to think about this.

I apply a filter to every AI opportunity, it’s called the 10K Threshold. This gives you an easy way to find out if an AI solution clears a certain value hurdle with the lowest bar being $10K per year (~3 hours per week). Below that, savings are too diffuse, the ROI is too hard to prove, and the project will die in the next budget review. I haven’t seen any scaled Engineered AI solution that lived below $10K/year business impact.

Given our original $37,500 on paper value, the next closest 10K threshold below that would be $10K per year.

This gives us an entirely different way to think about it.

Instead of asking "what do we do with these 2.5 hours saved per rep per week," ask the team:

Few people would approve $10K to save people 10 minutes. But protecting response times and service quality when the team is understaffed is an entirely different conversation that clears $10,000 p.a. easily in this case.

If we assume that this case earns us at least $10K p.a., the total running costs of this solution must stay well under that number. We could be conservative and define that the solution must not exceed $5,000 per year in running costs — API fees for processing ~2,500 emails/month, hosting, and basic maintenance overhead. That would give us another $5K per year in real value earned.

With $5K per year net profit from this solution, we'd have $5K development budget if this solution should pay off in one year or $10K if we're looking for a two-year period. Or $2.5K if we're looking for a 6-month payback. Which is the scenario I prefer most.

Here's how the economics of our solution work out:

The math works on paper. Before you build, 3 things must be true:

1. The emails must be complex enough to matter. If most support emails are 2 sentences, there's nothing to summarize.

2. Support docs must exist and be usable. If the company's product knowledge lives in people's heads and not in documents, the AI has nothing to look up.

3. Agents must support it. If no one reads the drafts or just rejects every AI response, the solution won’t have any impact.

Since the implementation budget is very tight, we won’t have much room for long prototyping sessions. The first version must be built in a day which means we must build on top of an out-of-the box platform like Power Automate or n8n depending on your preference. Or even copy-pasting some long email threads and relevant support docs into ChatGPT to see how well current models can handle the conversation out of the box given clean context and data.

Show the results to your human agents, collect feedback, and once you clear the Prototype-to-Production Checklist you’re good to go to build the full integration.

To summarize, here’s the profitable AI case for our support email solution:

I hope you enjoyed this little breakdown.

If you found it useful, leave a note or reply to this email and I’ll make sure to share more examples like this in the future.

See you next Saturday!

Tobias

P.S. I wrote a book on how to find and evaluate AI opportunities like this one. It’s called The Profitable AI Advantage.