PODCAST · SALES · AI AGENTS • 7 min read •

The client wanted an agent on its website. The problem was in the sales team

One of the large suppliers of printed T-shirts wanted an AI agent on its website. The website brought in probably under 10 percent of sales and worked well. Most sales went through sales reps, over the phone. The company had procedures, but the data from the CRM and the emails showed that not every rep stuck to them. The pilot we built was an agent meant to prepare a draft reply to a customer's email for the rep.

A headset with a microphone resting on a stack of folded dark T-shirts, on a desk under a green desk lamp

On the Polish podcast “I to się liczy”, Adrian Andrzejczyk asked whether the return on an investment in AI can be calculated at all. I said I thought so, and that we actually have to. Then I told him about a client who came wanting an agent on its website.

The whole conversation is on YouTube, in Polish. The client's story starts at 22:30.

This is the seventh post from that conversation. Here I describe how we worked out where the agent should go, and what that agent was meant to do.

First, answer why

A company can treat AI as a sandbox. It takes one project or one department and experiments. If it burns the money, it burns the money, and maybe something comes of it. That is a good approach, and I often recommend it when a company has the money, the people and the time for it.

It is different for a company with a proven process and a stable situation. That company has to answer why it is deploying AI: to cut costs or, for example, to scale. To start with, you need to know your process and the parameters that describe it. You also need to know why you are doing this.

As a consultant, I most often meet companies that do not have this mapped. Five questions worth answering in one sentence each before the first dollar goes into AI are collected in a separate list.

AI is a tool for automation, and automation is how we scale. If you start scaling a mess, you scale the mess. You can also use AI to clean up that mess first: map it, untangle it, put it in order. That is a completely different project, though. Scaling comes only in the next one.

An agent where sales were small

I was doing analytics for one of the large suppliers of printed T-shirts. Their customers pick a T-shirt and set the print. I am not naming the company. They had a few dozen sales reps, and each of them handled dozens, hundreds of customers a month.

Most sales went over the phone, through the reps. They also had an online shop with a configurator. They came with an idea: let's build an agent on the website. We already have the configurator, but now we're going into agents.

I asked how much of their sales came from the shop. It turned out to be probably under 10 percent, maybe even 5. The configurator worked really well and people had no problem with it.

So I asked why they would put AI where sales are small and everything works, when on the other side they have the reps, where they themselves say there is some total problem. Maybe that is where we should go in and see what is happening. It clicked. We said: fine, let's check.

Where the agent was meant to go

The client's idea and the place where the problem was

The client's idea

An agent on the website

  • Online shop with a configurator
  • Probably under 10% of sales
  • The configurator worked, people had no problem with it

The problem was here

Sales reps

  • Most sales, over the phone
  • A few dozen sales reps
  • The procedures existed, but not everyone followed them

The shop's share of sales is approximate, without exact figures.

WORK WITH ME

This is what I do hands-on: advising on AI strategy and building agents that survive the demo.

What the CRM and the emails showed

I got access to the sales data and analysed the problem with AI. We pulled everything out of the CRM, which also held the emails, and on top of that we had the procedures in PDFs, Excel files and access to the shop's database.

What stood out was that a lot of this work could be automated. The basics came out too:

  • Reps who replied within a few hours had better results than those who put it off.
  • Reps who were unable to match the offer to the enquiry and said “we don't have it” had worse results than those who said “we don't have it, but we can offer some alternative”.

I asked whether they had processes for this. They did. They sent me one document, then another. Except that people did not follow those processes. So the problem was not a lack of processes, but making sure someone carried them out.

The data showed this, too. Those who stuck to the processes had much better sales results than those who did not.

Newsletter

The Agentic Architect

Practical patterns, case studies, agentic projects taken apart and AI news. Zero spam, once a week.

You'll get one email to confirm. Unsubscribe with one click. Privacy policy

An agent meant to write a draft for the rep

The solution had two steps. First, map it and analyse it. Then build an agent that supports the reps.

A fast reply is, in practice, a few tasks. You have to prepare the email and check the offer and what is in stock. If something is not available, you have to find an alternative. Today AI can do all of that.

The agent was meant to prepare a draft reply for the rep, for example within a few minutes of the customer's email. The rep would only have to look at it. The draft was meant to have a set form and to follow set rules.

The point was emails of the quality the company expects, rather than each rep trying to reply in their own way, or not replying.

How the agent was meant to work

From the customer's email to a draft the rep looks at

  1. 1Email from the customer
  2. 2A draft reply in a set form, following set rules
  3. 3The rep looks at the draft

Pilot built and run internally.

In this setup the agent was meant to propose and the rep to look at the result. How to match the split of work between AI and a person to the cost of an error and the repetitiveness of the task is something I described alongside the hybrid patterns.

It was a pilot. We managed to build it and run it internally. It worked.

What an agent like this costs

Adrian also asked about the cost. I think costs have dropped a lot. Programmers use AI themselves, so the whole process has become shorter and faster.

How I work since agents started writing my code is in one of the earlier posts from this conversation.

When asked for an amount, I said every case is different. The price depends on the problem and on how complicated the processes are. The three things an agent's price depends on: integrations, autonomy and upkeep after launch are broken down in a separate post.

In my view, the most important thing for a reliable quote is that the client comes in openly and shows all the documents, their history and how they work. Then we have full analytics. We can prepare scenarios, a set of benchmarks, and even predict the return on investment. We can also run AI on the historical data and see how it does.

For that to be possible, you need data and you need to know what is in it. The worse the data and the less of it there is, the more of the solutions people have to propose. The better it is and the more of it we have, the more we can leave to AI.

SP

Szymon Paluch

Claude Certified Architect · ex-CTO

Not sure where an AI agent should go?

30 minutes on where the problem in your company is and whether an agent has anything to do there. No sales pitch.

Book a call
Related posts