A human approach to agentic teams

Much of the noise about AI agents across social media is about what they replace, not what they amplify.

Eg, the familiar social media post: This agent team saved me $20,000 instead of hiring an agency to do {insert task}.

Such scenarios get curiously high engagement, but seeing the realities of automation with enterprise clients makes them largely the realms of fantasy. 

We find ourselves in ambiguous economic times, which is causing great hesitancy in job creation. I will refer to the UK’s grey macro economic picture. The 2020s is also likely to be an historically low growth decade.

But the relevant summary is as follows: C-Suites may review off target P/Ls and consider AI to be an answer to cut costs and increase margins. We are seeing quite clearly that the graduate job market is very challenging. It’s tempting to conclude that companies are ‘replacing’ entry level roles with AI. The picture is more complicated, but appears inhuman and job destructive.

My long view is I don’t consider AI and automation to be a cost cutter or job destroyer, rather an amplifier. We are going through a period of unprecedented technological change, mixed with economic uncertainty. This creates an uncomfortable pinch point.

The problem is, even to cut costs, someone, somewhere needs to be able to deploy agents. A few larger companies will be able to align creaking structures to do so. I hazard the majority won’t very quickly. That ‘someone’ (plural) also needs a high level of business and technical knowledge.

The big mistake in my view is to cut staff in the hope of AI transformation, without a coherent approach being realised.

Will enterprise be left behind as they turn the oil tanker? Maybe. But let’s consider what’s needed to increase the probability that they won’t.

For any AI transformation programme, get all the right people in the room, probably many times, for many hours, to agree a strategy. I got a feeling for this reading about when Mark Thompson left The New York Times, and the hours of the debate they went through.

It’s a great example to me. I have been involved in struggles of opinion at many companies big and small, but never have I experienced the long and vital debate for any time longer than an hour. Away days are largely a positive presentation rather than a debate.

And I think this is so well placed on a Friday - to really hammer out the way. That is the human debate we need to have now, and it will take a long time. AND it is worth the time, because for many the next few years is existential.

With strategic alignment, who will ensure easy access to applications? Can someone set up an OpenAI Assistant (or other agent, given there are many options) easily? Can you also integrate the main tech stack quickly - particularly Google or Microsoft Suite? Who has the authority to do so? Map it clearly to avoid delay and repeated pain.

From above, this seems trivial. From below, it appears easy. Neither is true without this alignment occurring.

Now comes deployment - an entirely iterative process.

I propose nimble teams of 3 to be enabled to build agentic solutions fast. We’re talking several agentic solutions per week - even in enterprise. And with the first two doors unlocked - why not?

But what does this human team look like? It’s pretty straightforward really. Good things come in 3s. Like the Three Little Pigs, which my son seems to be obsessed with.

AI product manager - someone with a diverse entrepreneurial skillset who is able to combine business, creative and technical elements, and design processes and translate business needs into technical requirements.

Prompt engineer - I don’t like this term as the role is not very technical, but it’s caught on. What you really need is someone who is good at briefing people in document format. Commonly they will also be a writer, and quite a creative one - with the ability to write with clarity, and failsafe depth, to ensure an agent has the correct instructions. They also need to be on hand to correct and iterate. An agent prompt is not something to write and forget.

Workflow engineer - the high level design is provided from the above, the implementation is done by a technician. They understand variables, a bit (or better a lot) of Python and Javascript, MCP, integrations and API calls.

Working in media I’ve this system work time and time again. In broadcast: the producer, the presenter, the editor. In digital editorial: the commissioner, the reporter, the producer (a different role to broadcast, but one that similarly joins together).

It can easily work for AI. The problem is right now most companies aren’t set up for this nimble structure.

We can operate on borrowed time from elsewhere, but with that comes opportunity cost. With concentration of effort, comes amplification.

And this is what I mean by AI not replacing, but amplifying.

Imagine that your business was able to deploy 20 meaningful agents in 20 days, which amplified the productivity of 100 staff by 10% every day. Over the course of a year, the payback is huge.

Start by building a coherent human team of 3, and the elusive magic is more likely to appear.

From there, the augmented workforce can grow exponentially, with those valued humans driving the growth.