Everyone's asking the wrong question
When businesses come to me wanting to adopt AI, the first question is almost always about the technology. Which model should we use? Should we go with OpenAI or Anthropic? Do we need a vector database? Should we fine-tune?
These are valid questions. But they're not the right place to start.
The right place to start is much simpler: what work are you actually trying to get done?
Because the technology is a means to an end. And the end is always a process. A workflow. A series of steps that a person currently does, that could be done better, faster, or more consistently with AI involved.
The companies that get value from AI quickly are the ones that understand their workflows deeply before they write a single prompt.
Why workflows matter more than models
Here's what I've seen happen. A team picks a powerful model, builds a prototype, and shows it to leadership. It's impressive. It can answer questions, generate reports, summarise documents.
But when they try to embed it into real work, it falls apart. Why?
Because nobody mapped out the actual process the AI needs to slot into. Nobody asked:
- What triggers this task?
- What information does the person need to do it?
- What decisions do they make along the way?
- What does the output need to look like?
- Who reviews it? What happens next?
Without answers to these questions, you end up with an AI that can do impressive things in isolation but doesn't fit into the way your team actually works.
The intern test
I use this analogy a lot because it works. Imagine you're onboarding an intern to do the task you want AI to handle.
You'd need to tell them three things:
What they need to know. The background information, the company context, the relevant data. For an AI system this is your knowledge base - the documents, databases, and reference material the system needs access to.
What they need to do. The specific steps, the decisions at each stage, the actions to take. For an AI system this is your workflow definition - the structured process the agent follows.
What good looks like. Examples of great work, common mistakes to avoid, how you'll review their output. For an AI system this is your evaluation framework - the criteria you measure quality against.
If you can't clearly articulate these three things to an intern, you can't clearly articulate them to an AI. And if you can't articulate them, the AI won't perform.
The AI is only as good as the workflow it's given. Get the workflow right, and even a simple model can deliver real value. Get it wrong, and the most powerful model in the world won't help.
What a workflow-first approach looks like
When I work with a new client, we don't start with technology. We start with a discovery process that maps out the workflows where AI can make the biggest impact.
That means sitting with the people who do the work. Watching how they handle tasks. Asking questions like:
- Where do you spend the most time on repetitive work?
- Where do things slow down or get stuck?
- Where do errors creep in?
- What information do you need that's hard to find?
From there we identify the workflows with the highest potential. Not the ones that sound most impressive, but the ones where the process is well-defined enough for AI to handle reliably, and where the impact is real.
Then we document each workflow in detail. The inputs. The steps. The decisions. The outputs. The edge cases. The quality criteria.
Only then do we talk about technology.
Your workflows are your moat
Here's the thing that most people miss. The models are commoditising. GPT-4, Claude, Gemini - they're all getting better, cheaper, and more interchangeable. If your AI strategy is built entirely around a specific model, you're building on sand.
But your workflows are yours. Your processes, your domain expertise, your knowledge of what works for your customers - that's not something OpenAI or Anthropic can replicate.
When you build AI around well-defined workflows, you create something that's genuinely differentiated. It's grounded in your data. It follows your processes. It reflects your standards.
And when a better model comes along (it will), you swap it in and everything else stays the same. Your knowledge base. Your workflow definitions. Your evaluation criteria. That's the infrastructure that holds value.
The knowledge layer
Workflows don't exist in a vacuum. Every workflow depends on knowledge. The product specifications your support team references. The compliance guidelines your legal team follows. The historical data your analysts draw on.
Most companies have this knowledge, but it's scattered. Across shared drives, Confluence pages, Slack messages, people's heads.
Structuring this knowledge so that AI can access it reliably is often the hardest part of any AI project. And it's the part that pays the biggest dividends, because once your knowledge is well-organised, it benefits everything - not just AI.
This is where RAG systems come in. Not as a technology to bolt on, but as a way to connect your AI to the knowledge it needs to follow your workflows.
Start small, move fast
You don't need to map every workflow in your business before you start. Pick one. The one where the process is clearest, the knowledge is most accessible, and the impact is most visible.
Build it. Measure it. Learn from it.
Then pick the next one.
This is how the best AI strategies are built. Not through a grand top-down initiative, but through a series of well-chosen workflow automations that compound over time.
The best AI strategy isn't about having the best AI. It's about knowing your work deeply enough to let AI do it well.
Get your workflows right. The technology will follow.