Don't Sell Your Client a Colleague Who Chats. Close the Gap Between Quote and Payment First

The loudest pitch for AI agents this year has been the fully automated digital employee: a colleague who can do anything, chat with customers and make its own calls. Ask companies that already use AI where they actually plan to put agents first, though, and the answer is much duller. Quotes, orders, invoices, purchasing, email support.

A survey that ranked the scenarios

Telekom MMS surveyed 123 decision-makers at German companies this year, at firms with between 50 and 2,500 employees, all from the more AI-literate end of the market. They were shown thirteen scenarios for AI agents and asked which ones they’d prioritise. The ones that made the short-to-medium-term list were order and invoice processing, quote preparation, email service, purchasing, lead research, phone service, and IT support and ticket handling.

What these jobs share is that a person currently sits in the middle acting as glue. A customer enquiry arrives, and someone opens the CRM, looks up the specs, checks the price list and puts together a draft quote. An order arrives, and someone checks it against the warehouse data before passing it on. An invoice arrives, and someone works out which order it belongs to, checks the amount and sends it for approval. No single step is hard. The trouble is that it happens dozens of times a day, across several systems.

The most useful line in the write-up describes the sweet spot. Agents fit best on recurring processes that span roughly two to four systems and two to four steps. A single-step task is usually better served by an ordinary AI tool or classic automation, while a large end-to-end process is too hard to standardise. Quotes, orders and invoices happen to form one chain running from enquiry to payment, with the agent moving data and preparing drafts and a person taking over for exceptions and important decisions.

An unglamorous example

One case from the same series shows what that sweet spot looks like. cec Ingenieure, a firm near Munich that plans and surveys telecom and energy networks, produces a steady stream of site photos: open trenches, house connections, coupling components. Those photos prove work was done, document each construction section and back up the invoices, and until recently someone had to look at each one, classify it and file it under the right project.

Telekom MMS and cec built an n8n workflow for it. An AI model trained with Azure Custom Vision recognises the construction phase and components in each photo, the workflow adds location, time and project data, and the files go into the SharePoint and Power BI setup the company already ran. Nobody in this story is a chatty digital colleague. A gap that people used to bridge by hand has simply been closed.

What to keep in mind when reading it

The material is worth reading, with a few caveats.

The source has an interest. The t3n article presenting the survey is sponsored content produced with Telekom MMS, and Telekom sells an n8n-based T-AI Agent Platform built around permissions, approvals, monitoring and cost transparency. The survey’s conclusions line up with that product. That doesn’t make them wrong, but it’s worth remembering.

The sample is deliberately tilted toward companies that already know their way around AI, and the article itself says it isn’t representative of the German Mittelstand. A company that hasn’t started yet may not begin with its quote process.

And the published material has no hard numbers on results. The cec case says photo sorting became faster and more consistent without saying by how much.

Even so, the direction matches other sources closely. When n8n launched Agents, it described a good workflow as boring and meant it as praise. Jason Wade, a consultant I cited when I compared ten SME AI reports, recommends starting with one boring, frequent, measurable workflow. When the platform seller, the tool maker and the independent adviser all point at the same kind of work, that’s a signal in itself.

Three things to measure before talking about agents

If you run a small business and are wondering whether an AI agent is worth it, I’d start with the three questions the survey write-up itself raises.

  1. How often does this happen in a week? A process that comes up twice a month rarely repays the effort of automating it.
  2. How long does it take now? Include the waiting, the back-and-forth and fixing mistakes, not just the minutes someone spends at a desk.
  3. Which number will show whether it worked? Time from enquiry to quote sent, days from invoice received to booked, corrections made by hand each month. Write today’s figure down, or you won’t know later whether anything improved.

Only once those three are clear does it make sense to talk about tools and models. Within the tooling, the part that deserves most of the attention is governance: what the agent can see, what it may do on its own, and when a person has to press approve. I went into that in more detail in last week’s post on agent permissions. In my experience, getting those three rules right does more for a project’s chances than switching to a smarter model.

For people who sell this kind of work

A last word to fellow consultants and service providers. Clients often open with something like “can you build us a ChatGPT that understands our business?”, and selling them a colleague who chats is the easiest deal to close and the one most likely to disappoint. A steadier opening is to help the client find the gap that comes up dozens of times a week, spans two or three systems and is currently held together by hand, measure it, and then close it. It makes a less impressive demo, but it produces a return you can check, and it makes the client’s decision about what to do next much easier.

Sources