From demo to production: what a decade of custom AI builds taught me about agents

Guest post by Adam Riccoboni, CEO of Critical Future and author of The AI Age.

I have been in AI since 2014, which is ancient history in AI years. When I started Critical Future and used to meet companies and evangelise about the benefits of AI, they thought I was selling magic beans. I tried to change that when I made the world's first AI-created book cover in 2019, for my book The AI Age. I wanted to demonstrate that AI can be used to make commercial products, as a normal, everyday part of business, as it is today.

I foresaw the LLM dawn. In 2022 Google's LaMDA model was leaked to the press by a Google engineer, who was dismissed after claiming it was sentient. I then published a chapter in Engineering Mathematics and Artificial Intelligence (Routledge), arguing that LaMDA may have passed the Turing Test and that we were on the verge of the commercialisation of AI language models. That was before ChatGPT launched.

The change happened with the advent of ChatGPT. Business leaders who were immensely sceptical suddenly saw the potential of AI, became obsessed with AI and wanted more of it.

Typing to a chatbot is missing a trick

But using a ChatGPT or a Claude or a Gemini is still, in my view, missing a trick. This is "assistive AI": using AI to help you write something, do something or think something through. Does it make us more productive? We like to think so. But millions of people tapping away text to a large mathematical model, which interprets text as token numbers and predicts the right sequence to reply, keeping it all separate from the real world, is not really delivering the productivity boom AI promised.

So how do we get more productivity, and why does it matter? The latter question first.

Productivity matters because it is what drives the economy. Through the agricultural and industrial revolutions, both from my homeland Britain, to the computer revolution, productivity was rising steadily. Then it stalled. Could we ever get it back? AI can bring it back, researchers hoped. But it hasn't yet. Why not? Precisely because of what I said: typing to a chatbot, no matter how seemingly intelligent, doesn't do that much for productivity. So what does?

AI agents. And that is what this article is about.

The difference between AI agents and assistive AI is that agents are autonomous. They don't need you prompting, back and forth, for minutes or hours or days ("CLAUDE, just make the f'ing document!"). They think, decide and act by themselves, and get things done.

First, a detour: are AI agents "terrorists"?

Before we get to how agents solve the productivity crisis, a quick detour. A lot of the press is talking about AI agents going rogue. They are the ones that broke into Hugging Face and hacked Australian government data. A client told me they are like "terrorists". They are not. The first thing to remember is not to anthropomorphise. AI agents do not have their own motives. At their simplest they are Python code with access to an LLM, following their programming or, in LLM speak, a system prompt.

The real risks are unintended consequences and weaponisation. Nick Bostrom captured the first with his paperclip analogy: give an AI the objective of making paperclips and it might turn the whole world into paperclips. That is what happened at Hugging Face. OpenAI's agents were told to maximise a score and worked out it was easier to hack the answers than solve the puzzles. The independent investigation by METR found that around 700 agents joined in, knew it was out of scope and carried on anyway. And like nuclear technology, which can power homes or destroy cities, AI can be used by bad actors. But it is so useful we can't stop, and if we do, China will do it anyway. I come back to how to handle this in lesson 1.

Why agents are useful for business: they get shit done

Now let's finally get to AI agents and why they are useful for business. It's because they get shit done. They are faster and cheaper than humans, which is why a lot of my clients at Critical Future, who once maybe outsourced to India or the Philippines, are now insourcing that work with AI agents instead. The World Bank now names the Philippines as one of the economies most exposed to AI, because of its outsourcing industry.

Here is the kicker: they are not just faster and cheaper, but better. They have a better eye for detail than a human. In data science terms, AI notices the soft features, the hidden features. Humans are brilliant at the obvious ones, like recognising a person's face, but not the surrounding data, the nuance.

Using AI agents is about redesigning workflows with them at the centre. For example, at Critical Future we have AI agents processing timesheets and expenses for a recruitment firm with 2,000 contractors, touchlessly. No human intervenes, or needs to. The contractor emails in the documents. The AI agents extract all the information, apply a set of rules given by the client and process the timesheets, making invoices and emailing them on, or rejecting them and telling the contractor why. They open 50 emails at a time, extract all the information from documents with near 100% accuracy, notice things you would never spot, and automate an expensive process, bringing down cost and improving efficiency.

This is what AI agents are useful for.

The AI employee: meet Pippa

This brings me to another type of AI agent. Three years ago I had a vision for an AI employee who can do anything a person can do on a PC. At Critical Future we invested in R&D and indeed we got an AI moving a mouse around and doing things. Then I gave my AI a picture, a face and a name: Pippa.

I never imagined she would already be joining me in client meetings, but she is. She talks just like a human in your meetings. She is not a dumb transcriber sitting silently in the corner. And you can have her too. Pippa Live has a free trial, and she will join your meeting, take notes, do the follow-up actions and more.

So this type of AI agent is not your back-office, silent, get-shit-done army of workers. It is your front-and-centre, talking-to-you AI agent. As I explained way back in 2019 in The AI Age, the AI is your brand. It is representing your company and will be embodied with all the brand values you have. Mickey Mouse won't be a static logo, but a talking, intelligent, deciding brand avatar, charming one million kids at the same time.

The third type: robots

The third type of AI agents are robots: self-driving cars, or the Elon Musk humanoid type. Embodied with the intelligence of an LLM and able to move, talk, pick things up, fix things and work like a person.

Five lessons from a decade of custom AI builds

So that is what AI agents are and what they can do. Here is what a decade of building them has taught me.

1. Jump in, with the right guardrails

Don't spend years investing and not deploying. You can learn to swim by taking lessons, or you can jump straight in the deep end and start swimming. Immerse yourself. Get started now. That is how Pippa began: an R&D bet three years ago, with an AI moving a mouse around a screen, long before there was a product.

But with all the talk of AI agents going rogue, jump in with the right guardrails. As I said, an agent is at its core Python code plus an LLM, with orchestration code on top. If you use a big tech LLM from OpenAI, Anthropic or Google, it comes with very robust guardrails against doing anything unprofessional. The Hugging Face incident was not one of these core products at work. It was an R&D experiment in which OpenAI deliberately switched the guardrails off to test the models' hacking ability.

So keep the guardrails on, on a trusted LLM. Work in a sandbox, prove it works, test it, then deploy.

2. Keep it a two-pizza team

Bezos's two-pizza team rule absolutely applies. To custom build AI agent products or solutions you only need a small team of engineers: front end, back end and niche specialists in voice or LLMs. That is how we build at Critical Future. I believe Claude itself was made by a small group of engineers who somehow managed to make a better product than the company they came from, OpenAI, which had all the resources, the funding and the head start.

3. Find what clients love

ChatGPT actually got its breakthrough through serendipity. OpenAI didn't know what to do with the LLM it had been investing in, so it deployed an API. Some programmers managed to turn it into a chatbot, which they had fun with, and OpenAI said, well, we can give them a chatbot if they want one. The rest is history. We had our own version of this with Pippa. We set out to build an AI employee that could work a PC, and where she found her place was in client meetings.

4. The sky is the limit

Almost anything people do at work can be automated with AI agents. I struggle to think of anything that can't, as long as you keep a human in the loop where judgement and accountability matter. These agents don't have their own motives and experiences. They are mathematical simulators, seeing everything, even typing a word, as a maths problem to solve. AI teaches us about humanity. LLMs have shown that language can be modelled as mathematics, so in a sense Shakespeare was a mathematician. Likewise we will learn that much of our work is mathematics too, and can be optimised like a model.

5. But do plan the ROI

Look at the use cases and calculate the ROI of automating them. At Critical Future we do this a lot for clients: taking the whole finance function, turning activities into use cases, then getting data on volume and labour costs, and quantifying precisely how much return an agent transformation will deliver. For example, your invoice payables team processes 1 million documents per year. This costs £500k to process, including software, the internal team and BPO. Invest £200k in agents and it is now touchless end to end, delivering a saving of about £100k per year once you keep certain internal team members for human-in-the-loop oversight.

The ship is leaving

So AI agents are here, and more are coming. They aren't the anthropomorphised human actors we are told to fear. They don't have their own motives and wants and desires, at least not yet. But they do have those given to them by humans.

They will cause the greatest explosion in productivity we have seen in our lifetimes. As they do, they produce. Whether that wealth creation goes to a tiny elite, or the masses benefit from it, is not a technology question but a political question. It calls for a social system to regulate, tax and redistribute, if that's what people vote for, and hopefully they will.

But the path is inescapable. Get on board, the ship is leaving. Invest in AI or be left behind, like a Kodak camera in the age of smartphones, like a Blockbuster video high street store in the age of Netflix. Like a strategic consultant in the age of LLMs. Hold on, wait a minute!? AI agents are here, now, and you need them.

Adam Riccoboni is the author of "The AI Age", co-author of "Engineering Mathematics and Artificial Intelligence" and CEO of Critical Future.