It’s the weekend, plans are being made around you, people are going to the pub and asking if you’re coming. Maybe you’re a bit knackered from the week. I’ve got a bit of info that might sway you to come along.
Working from home, it can take a lot to get me out after work, but the last few times I’ve been out, the conversation has inevitably turned to AI. Which, fair enough, may sound less interesting than it used to. What’s striking is how much the conversation has moved on in the past few months.
Not that long ago, a lot of AI talk was still about models: which one was better, what the APIs could do, whether the capability was real, and what might happen next. Then it moved into personal tools and early experiments: writing assistants, image generation, video tools, research helpers, and plenty of software that was basically a chatbot front end with not much happening behind it. Now, the conversation is getting more operational.
People are talking about orchestration, MCPs, agent workflows, local models, pipelines, and the tools that sit between raw AI capability and actual work. ComfyUI is one example from the graphics side. Instead of treating image generation as a magic prompt box, it lets people build workflows: models, inputs, outputs, adjustments, and repeatable steps stitched together in a way that starts to look much more like production practice.
ExoLabs points in a similar direction from the infrastructure side, asking whether useful AI can be run locally, across ordinary machines, without immediately depending on someone else’s cloud platform.
These are cool tools, but the specific tech is less important than the signal. People are finding ways around the gaps, limitations, and sometimes terrible product directions of the major platforms. Underneath all of this is a practical question:
How do I use AI in a way that is useful to me, affordable, context-specific, and safe enough for the work in front of me?
And sometimes sitting just underneath that is a second question:
Do I really need the big vendors for all of this?
That’s a very different conversation from “which model is best?” It’s also a different conversation from the one many organisations are still having about AI strategy, rollout, and approved tooling.
Informal networks are becoming AI infrastructure
In the Allianced Cultivation piece, I wrote about people asking urgent questions in response to constant change, including the obvious one: “How will AI impact my career?” One answer, it turns out, is that AI is putting more responsibility on individuals to stay ahead of the curve than ever before.
With change happening so fast, people aren’t waiting for perfect organisational clarity. They’re using trusted professional networks to compare notes, test assumptions, and work out what’s useful. Informal networks carry practical knowledge before it’s standardised — a place where formal channels often struggle.
Peer-specific knowledge goes beyond generic claims about “AI productivity.” Instead, you’re hearing how someone in a similar role, with similar constraints, used a particular tool to solve a particular problem. That kind of context matters. It helps people judge whether something is actually useful, whether it’s safe enough, and whether it’s worth the effort.
But informal networks also do something else. They move ideas across disciplines.
A conversation that starts with AI-assisted software development can suddenly become relevant to graphics, photography, research, operations, compliance, or product work. Someone describes a workflow from one domain, and another person recognises the same pattern in theirs. The tool may be different. The output may be different. But the underlying strategy travels.
That’s why informal networks matter. They don’t just spread tips. They help people recognise patterns that would otherwise look like isolated experiments, personal preferences, or random tool chatter. And once those patterns start shaping how people actually work, they stop being background noise. For leaders, that’s when the battles start.
When formal strategy lags, good governance helps
Inside organisations, AI adoption is still moving through the formal machinery of change. Strategies are being written. Tools are being reviewed. Policies are being drafted. Procurement, governance, risk, training, data protection, security, and measurement all have to be worked through.
That work matters and shouldn’t be dismissed. Organisations have responsibilities that individuals don’t always have to think about in the moment.
But this is also where the friction starts. Informal practice doesn’t always arrive politely, in the right template, at the right committee, with a business case attached. It arrives as people trying things, finding shortcuts, comparing tools, and quietly changing how work gets done.
AI is already on every agenda. The harder question is whether organisations can absorb what people are learning fast enough to do anything useful with it.
That gets uncomfortable when informal conversations start turning into experiments. At that point, learning is no longer just happening outside the formal structure; it’s starting to change how work gets done. Organisations often have a natural Not Invented Here reflex. If learning didn’t originate inside the approved structure, it can look messy, unofficial, unsafe, or simply inadmissible.
The discomfort isn’t only that unofficial learning looks messy from the centre. It’s that it can challenge decisions already made. Strategies may have been set, vendors selected, budgets committed, and senior people may have put their credibility behind a particular direction.
So when useful knowledge arrives from side projects, peer conversations, tool communities, or patterns people have noticed before they’ve been validated internally, it can feel threatening. From a governance perspective, it’s tempting to treat that kind of knowledge as inadmissible until it has passed through the proper channels.
That’s the Not Invented Here reflex kicking in.
In normal circumstances, that instinct can be useful. It protects the organisation from fads, vendor nonsense, uncontrolled risk, and random personal preference dressed up as strategy.
But with AI, a pure Not Invented Here posture won’t work. Too much of the practical knowledge is forming outside the formal boundary first.
That doesn’t mean every informal practice should be accepted. It means governance has to become better at listening to signals it didn’t create, then turning the useful ones into formal practice.
I’ve had to work through this problem directly when designing security and governance around analytics, machine learning, and AI platforming in large organisations. One of the simplest requirements tells the story: data science tools, ML libraries, AI packages, and related components may need to be reviewed or updated constantly. Sometimes daily. That reality doesn’t fit neatly into governance models built around slow approval cycles, fixed tooling lists, and occasional review boards.
The answer isn’t to let everyone pull whatever they like into production. But it also isn’t to pretend that a static approval model can keep up.
The formal practice has to be designed around movement. That means clear ownership, regular communication between data science, engineering, security, architecture, and leadership, escalation paths people understand, and enough shared responsibility for teams to act quickly when something is inside agreed boundaries — and slow down when it isn’t.
That’s the important distinction. Good governance doesn’t mean every decision starts from the centre. It means the organisation knows how decisions are made, who owns the risk, when something needs escalation, and how new learning becomes part of normal practice.
Informal exchange: the hype killer
Not everybody enjoys the pub, of course. But these informal conversations matter now more than ever. It could be a coffee, a meetup, a conference side conversation, a Slack backchannel, or a remote call with an old colleague. I, personally, miss a good diner.
These are the places where people say the things that don’t always survive official channels:
- “This is saving me a bunch of time … like this.”
- “That tool looks impressive … but falls apart when the work gets complicated.”
- “Don’t put client data in there … let me tell you this story, just don’t repeat it.”
That kind of exchange matters because it’s grounded in actual work. It isn’t a vendor claim, a productivity slogan, or a polished internal success story. It’s evidence from people testing tools against the reality of their own jobs.
It also carries a different kind of honesty. In formal settings, people often talk about AI in the language of strategy, risk, transformation, or approved use cases. In informal settings, they talk about the awkward middle: what helped, what broke, what felt risky, what they’d use again, and what they’d quietly avoid.
That’s where some of the most useful signals live.
Turning signals into capability
For leaders, the point isn’t to romanticise informal AI use. It’s to ask where the organisation’s AI ideas are really coming from.
Some will come from vendors, strategy decks, analyst reports, and formal programmes. Some should. But some of the most useful signals will come from competent people solving real problems: the tools that save time, the demos that don’t survive contact with actual work, the risks that appear in ordinary use, and the judgement that still needs to stay human.
The useful move isn’t to clamp down on that signal. It’s to make it safer and more visible without killing the trust that made it valuable in the first place.
That might mean internal demos, lightweight show-and-tell sessions, clear red lines around sensitive data, safe experimentation spaces, and feedback loops between users, security, legal, data, and leadership.
Good governance should turn informal learning into organisational capability, not smother it.