AI
Before you start an AI project, Part 4
This article draws on hands-on experience from real AI projects across different industries. It’s not a how-to guide. These are observations we picked up along the way.
Most companies track to the minute who worked on what project. Reports are drawn up showing how many hours the team spent on a task, where each process could be tightened up. Then an AI subscription lands in the budget and suddenly that same rigour evaporates completely. We bought it, so we must be more efficient. Who actually uses it, and for what? That rarely gets measured. The tracking stops at the exact point where the most expensive tool enters the picture.
There’s a neat parallel with the microwave. When they first appeared,publishers rushed out thick cookbooks explaining how to prepare an entire Christmas dinner in one. And what do we actually do with them today? We shove in yesterday’s leftovers and keep pressing the “+30 seconds” button until the food is roughly warm.
Smartwatches follow the same story arc. We strap them on with the conviction that from tomorrow we’ll be fitter and more mindful. The device delivers flawlessly, measuring heart rate, counting steps, analysing sleep. But how many people’s habits actually changed just because they bought one? We love generating data. We assume that if there’s enough of it, we’ll be productive, because surely we’ll process it all at some point. (Quick question: how many photos are sitting on your phone that you haven’t glanced at in over a year?)
The pattern is that we don’t really buy the technology. We buy the possibility of what it can do for us, and quietly trust that the hard part will sort itself out.
Smart tools vs. smart use
Most companies today are using advanced tools, but not in a particularly effective way. While adopting new technology is an important strategic step, on its own it rarely leads to meaningful efficiency gains.
At the beginning, it’s often more valuable to focus on how tools are used rather than how advanced they are. Throughout history, even the simplest tools proved powerful when applied skillfully. A basic tool doesn’t automate thinking or produce results on its own, but in the hands of someone who knows how to use it well, it becomes highly effective.
This is the key idea: smart use matters more than smart tools. Before investing further in more sophisticated solutions, it’s worth asking whether you are already making the most of the tools you have.
What the numbers show
Enterprise AI surveys from all across the globe, reinforced by our own ongoing research, paint a clear and unfiltered picture:
Around 75 to 78% of organisations proudly announce that they “use” GenAI. In practice, only 10 to 14% of employees actually reach for these tools on a daily basis. More than 80% of surveyed CEOs and CFOs report seeing no measurable impact on either headcount or productivity over the past three years. Corporate spending on AI stood at 37 billion, investment grew sixteenfold and daily use sits at roughly a tenth of the workforce.
There is one notable exception: GitHub Copilot, where 67% of licensed developers use it five or more days a week. That is not an accident. Copilot is woven into one specific task, right where the developer works, and it gives immediate, tangible feedback. It was deliberately built to fit inside the workflow rather than sit alongside it. Like a good hammer, it is just there when you reach for it.
The rest got put on the shelf.
The most expensive AI packages end up being used to spell-check emails or rewrite a paragraph to sound “a bit more formal.” People ask it for definitions instead of just looking them up. This is the microwaves’ “+30 seconds” button of the AI world, and most of us are pressing it on repeat.
And when a task feels like it is not moving fast enough, the reflex is to go browsing. Maybe Claude would be better for this? Should we pick up a video generation tool? The instinct to shop for a solution is itself the problem.
What smart use looks like in practice
Let’s look behind the button through three specific roles, using essentially the same set of tools.
But before we get into it, let’s talk about some basics. When you open ChatGPT, Gemini, or Claude, a default model greets you, but the subscriptions typically include several models, and there are real differences between them. Not necessarily that one is better than another across the board, but that each one is better at certain things.
If you don’t bother checking, you end up using the same setting for every task, just like on the microwave: you burn through your quota, slow yourself down, and sometimes even get worse results, because a large model can overthink a simple task. You don’t need the most powerful, most expensive model to rephrase a quick email.
On top of that, different providers’ models respond with different “personalities”. One is more concise, another more detailed. One thinks in a more structured way, another writes more naturally. Much like with colleagues: with some people collaboration just flows because they get how you ask questions, with others it doesn’t. If a tool isn’t working for you, try a different one, or simply ask in a different way.
With that backdrop, let’s look at the three roles.
HR: the knowledge base that answers on your behalf
HR typically gets bombarded with the same questions every week. Where’s the holiday request template? What’s the company’s position on remote work? When does the loyalty bonus kick in?
This is where knowledge base building comes in. If you upload internal policies and the most common questions into a document-based AI tool (NotebookLM is currently one of the strongest for this), it can answer these questions on its own. Internal wiki, onboarding materials, contract collections. It solves the age-old problem of knowledge living inside people’s heads rather than in a system. The time freed up can go where it’s actually needed.
The same toolkit helps with writing job adverts, once you’ve trained it on the company’s tone of voice and legal requirements. The deep research features now available in ChatGPT, Claude, and Gemini can also be used to research salary bands and market benefits. What used to take days of trawling now comes back as a coherent analysis built from multiple sources in parallel. It can still hallucinate, but as a starting point and baseline research it saves a great deal of time.
There’s room to move on CV processing too, though you need to be aware of the AI Act boundaries here. Automatic scoring of candidates is prohibited, but structuring and filtering experience summaries (e.g. “who has at least three years of relevant experience?”) is doable with AI, with a human making the final call. This approach is called HITL, Human-In-The-Loop: the AI prepares, the person decides.
Project manager: the summary that’s ready in minutes
A significant chunk of a PM’s day is spent gathering information from different places and passing it on in some form.
If you have a Google Workspace subscription, Gemini can already search within your company’s own materials. Emails, documents, spreadsheets, files on Drive. It has access to everything you do. Through the Workspace integration (“summarise all the project’s Drive documents and last week’s email threads”), the weekly summary can be cut down to minutes. This isn’t a chatbot any more. It’s a search engine across your own organisation.
On both major platforms (Google Workspace, Microsoft 365 Copilot) you can create pre-prompted assistants. In Gemini they’re called “Gems”, in Copilot they run as agents. The idea is the same: you write up once what it’s for and how it should behave (e.g. “Weekly status report writer that automatically drafts a summary from project data”), and then you don’t have to start from scratch every time. If you’ve worked out one well-crafted prompt, that alone can save hours each week.
Agentic features show their real strength here. They don’t just answer a single question at once. Instead, they work through a task iteratively, in multiple steps. They can compare several incoming bids, tender documents, or client requirements in parallel, and produce an overview that would take hours to assemble by hand. This isn’t a chatbot-level question-and-answer. It’s delegation: you hand over the task and get a summary back.
Developer: the hammer they got first
For developers, GitHub Copilot is a well-known example, and it’s no coincidence that it’s the only AI tool people genuinely use daily. But it’s just one tool.
With Claude Code or similar agentic solutions, you can get a legacy codebase understood, documented, and refactoring suggestions drawn up. These are the kinds of tasks teams typically put off because doing them by hand takes a disproportionate amount of time. Generating test cases, writing API documentation, running a first pass on security risks, all within reach. The developer won’t become redundant, but they can step out of some of the dull, repetitive technical debt.
What might surprise you: developer environments aren’t only suited to development tasks. The same tool lets someone in HR run through a collection of contracts, or a PM processes a folder of tender documents. The boundaries are less sharp than you’d first expect.
Connect, don’t isolate
The real problem with “+30 seconds” AI use is the constant copy-paste: from Word into ChatGPT, back again. In all three roles.
Smart use begins when you treat AI not as an island but as part of your systems. If you connect an existing subscription to an automation platform (n8n, Zapier) and hook it into your CRM or email, incoming customer complaints get categorised automatically, the AI drafts a response based on previous guidelines, and you decide and approve. Two tools intelligently linked together are more powerful than six premium AI packages that colleagues are trying to use in isolation.
A note on data
Before uploading anything to an AI system, it’s worth pausing for a moment. The EU AI Act and the GDPR together govern what data can be fed into these tools and under what conditions, and it also matters which subscription tier you’re on. Personal data, employee data, client data can’t just be pasted into a chat window. It’s not as complicated as it sounds, but there are mandatory steps involved. The opportunities are real, but the prerequisites are worth sorting out in advance, not after the fact. We’ll cover this topic in more detail in a separate part of the series.
Where to go from here
If you already have the tool, it’s time to use it properly before buying another one.
Take stock. Feature lists change often, and it’s easy to lose track of what that package you bought a year ago can do today. Have you used 80% of it? Almost certainly not. That’s not a reproach, it’s a starting point.
Build skills beyond prompting. Don’t roll out software, build capability. It’s easy enough to have a chat with AI, but using it well in your own profession is a different matter. You don’t learn that from videos. You learn it through practice: how does AI fit into my specific work, my specific tasks.
AI can deliver a lot, but on its own it’s blind. Every new conversation is a fresh start: it doesn’t know you, doesn’t know the company, doesn’t know your data. It’s like a colleague who’s just arrived and needs to be briefed. We steer. As long as we’re stuck on the “+30 seconds” button, the world’s most advanced language model remains nothing more than an overpriced reheater.
Not sure how to get the most out of the AI tools you already have, or how to connect them effectively?
Often it’s not a new subscription you need, but a clear picture of what’s already there. A short conversation can save months of frustration and unnecessary licence fees.
Get in touch with our team: https://attrecto.com/contact
Our AI Enablement programme is designed to help teams move past technology hoarding and onto the path of smart use.
