The S.C.A.L.E. Framework I Use to Build AI Agents That Do Real Work

An AI agent, for me, is a self-autonomous, or almost autonomous, thing that does part of a human job faster, easier, and in a more repetitive way.

It can use voice or text. It can live on your website, WhatsApp, email, or on the phone. It can be reactive: someone asks, and it replies. Or proactive: you need to ask something to your customer, and you delegate that to the agent.

This is what I am building with Carla and Andrea for an Italian real-estate auctions consultancy I work with. Carla is a voice agent. She answers the phone line, calls back the people she promised to call back, gives facts about a property, and books a phone consultation with a human consultant. She says she is an AI assistant at the start of every call.

Andrea is a text agent on WhatsApp, email and the website chat. In the team’s words: “Carla is the secretary, fast. Andrea is the assistant, he goes deeper.” Andrea builds a profile of what the person is looking for and proposes properties one by one. Both are built on ElevenLabs, with dozens of tools each: look up a property, find free slots, book / move / cancel an appointment, schedule a callback, send a WhatsApp or an email, update the CRM.

And because Americans love frameworks with an acronym, apparently I need mine too. So here is mine: S.C.A.L.E.

S — Scale without hiring

An agent lets you absorb more demand without turning every traffic spike into a hiring plan.

This lands immediately with business owners. If contacts grow, the classic answer is simple. Hire more people. Train them. Brief them on the product, the process, the edge cases, the tone. Then monitor them, because even good people need context.

In this project, contacts went from about 1,250 a week in August to over 3,600 in September. Roughly three times the load. With agents, the question becomes different: can the system answer, qualify, route, book, and follow up at this volume?

This is why I don’t like “AI reduces cost” as the main point. Sometimes yes. Sometimes the real win is optimization: similar money, much more demand handled. For Carla, a minute costs about 14 cents. And the funny part is that the language model is about two thirds of the bill, not the voice.

C — Consistent, every time

The answer quality of an agent is mostly the quality of the knowledge you give it.

Many people think an agent is a script. If the customer says this, answer that. That is old automation with a better face.

A real agent has your knowledge: product details, constraints, objections, process. Then it uses that knowledge to craft the best answer for that specific customer. Not canned. Consistent.

Before designing Carla, we analysed 79,857 past calls handled by the previous agents. Because that is where the business logic lives. What people ask. Where they get confused. Which information makes them trust you. Which missing information kills the conversation.

And the results are very clear. With the full property sheet in hand, 50.5% of calls ended with a booking; with a listing it could not link, 30.0%; with only a saved search, 18.2%. Same agent. Same voice. Different knowledge.

Your knowledge is the product.

A — Adapt to the customer

A good agent does not force the customer into your favorite channel. It follows them.

Some people call. Some people text. Some people start on the website and continue on WhatsApp. If you design the agent around your internal preference, you create friction.

Carla works on voice because some conversations start with a phone call. Andrea works on WhatsApp, email and website chat because many conversations are better in text. The point is not “voice is better” or “text is better”. The point is: where is the customer already talking?

This is what I mean when I talk about AI in the middle of the conversation. Not AI hidden in a dashboard someone has to remember to open. AI inside the real flow of work.

Adaptation is tone too. If someone sends a short message, Andrea should not reply with an essay. If someone wants a consultation, don’t make them read a brochure. Find a slot and book it.

And when it is you who has to reach out, persistence beats perfection: 60.6% of people answer the first call, and within three attempts 81.1% have been reached.

L — Learn from every conversation

Once the work is handled by a system, you can finally measure the work like a funnel.

This is the part I enjoy the most. A team of people has good days and bad days. A system you can measure, and then fix.

For Carla and Andrea, the prompt is not a text blob edited by hand in a dashboard. It lives in git. The system is treated like product code, because it is product code.

We cut the prompt from 69,000 to 25,700 characters. Not to make it “shorter” because shorter sounds elegant. We did it because large prompts become heavy, hard to reason about, and easy to break. We also built a battery of 36 test scenarios taken from real calls, with a model acting as judge. In one round, the system went from 67 to 80 criteria passed out of 105. We measure the time to the first word too. Target: under one second. Because if the agent waits too long before speaking, the experience feels broken.

And metrics also kill your favorite ideas. We stopped proactive outbound calling because of what it cost. I liked the idea. It felt useful.

The metrics decide, not my enthusiasm.

E — Escalate to a human

The best agent knows when to stop being the main character.

I am not against humans. I want humans supercharged. AI should do the repetitive work, the instant answers, the routing, the follow-up, the scheduling. Then, when the conversation needs judgment, emotion, negotiation, trust, or responsibility, it should pass to a human.

This is where many AI projects fail. They try to automate everything because the demo looks cooler. In this system, every booking lands on a human consultant’s calendar. Carla can book it, but the consultation is human.

We created clear “S.O.S.” paths: urgent, when the auction is close and no slot is free; “we want to bid”, which goes straight to the boss; “call me back”; and staff tasks. We created those after 626 S.O.S. in one week, where more than half only said “wants to talk to a person”.

Andrea also knows when to disappear. If a consultant writes by hand in a thread, Andrea goes quiet for 24 hours, 48 hours or forever. The consultant decides. Then one button gives the thread back to Andrea.

I wrote before that AI won’t replace you. I still believe it. But I would add a piece now: AI will replace the parts of the work that should never have required a human at all.

If you want to build agents like Carla and Andrea for your business, get in touch. I am building them for real, not for slides.