A guide to incorporating AI in your organisation
If you have reviewed a colleague's work, assigned someone a task, summarised a report, prepared for a meeting, checked a document or asked someone to improve a draft, you already have most of the skills you need to work with AI.
The easiest way to think about it is this: AI is another member of the team. It can contribute ideas, analyse information and challenge your thinking. And like any team member, it needs clear direction and supervision.
Start with an approved tool, and mind what you share
Begin with an AI chatbot your organisation has approved. Check what the approval covers and what the guidelines allow you to upload. If there is no guidance, treat the model as a supplier bidding on a tender, and share accordingly.
You would not hand a supplier everything you know about a sensitive project simply because they asked for context. You would work out what they actually need, what they are authorised to receive, and what protections are in place. Treat AI the same way.
The reason is practical. When you enter information into an AI service, you are sharing it with that service. What happens to that information depends on the product, plan, contract and settings. It may be retained and, with some consumer services and settings, may be used to train the model. Approved tools should have been vetted for exactly this. There may be contractual protections governing your data, or specific instructions you are expected to follow. AI adoption should work within your normal business governance, not around it.
Choosing which AI to use
There is more than one AI model, and they do not all behave the same way. Picking one is a lot like recruitment.
When several people apply for a role you look at their experience, you ask questions, you test their knowledge and you judge how well they fit the job. You can do exactly this with AI. Give several models the same question. Use identical wording each time, otherwise you are not comparing anything.
Put your own terminology to work. Every team has industry-specific language and techniques, so ask about them and watch whether a model applies them correctly or just talks around them. Then ask about a recent development to see whether it can access current information and whether it is clear about the limits of what it knows. Then give it something well outside your business area to see how it copes. Compare the results.
You are interviewing them. You may find one model is best at analysing documents, another at research, another at coding, and another for a quick answer. No single AI is best for everything. And unlike a hiring round, you can repeat the interview whenever your needs change.
Match the capability to the task
We already do this with people. You would not hand a complex, high-risk project to someone who has never managed one.
AI models differ in capability too. Some are built to answer quickly. Others spend more effort on a problem before replying. Some are general-purpose, others are built for particular kinds of work. More capability is not automatically better. An advanced model can hand you far more analysis than the task actually needs.
So try the same request across different models and settings, compare the results, and keep the one that fits the work. The principle is familiar: match the capability to the task.
Put it to work on your own documents
A good place to start is the information your teams already handle. Businesses produce and receive a constant stream of reports, presentations, procedures and meeting notes, and the same information often has to be presented differently to different audiences. Senior management may want the main decisions and risks. An operational team may need the detail.
Many AI tools can work with documents you provide and let you ask questions about them. Instead of a flat "summarise this", you can ask:
- "What are the three things a department head needs to know?"
- "Identify the decisions, risks and unresolved questions."
That is not sophisticated prompt engineering. It is simply giving a colleague a clearer instruction. And when a model is working across your own material, it can surface a connection, an inconsistency or an issue that nobody involved had noticed.
Refinement is a conversation
Working with AI is rarely a one-question, one-answer exercise. It is a conversation. You ask for something, you read the response, and often it is too long, or it misread you, or it needs more inputs or you simply want a different angle. So you tell it:
- "Make this shorter."
- "Focus on the operational impact."
- "Explain what you think is missing."
- "Challenge my assumptions."
That is very close to reviewing work with a person. There is one difference worth noting. People sometimes hide the fact that they have not understood an instruction, hoping the confusion sorts itself out later. AI can be told to flag uncertainty and say plainly when it does not know. That does not guarantee that it will always recognise when it is wrong, but it gives you another way to question, rephrase and steer it towards what you actually wanted. The output may still need work, but that back and forth is where the value is.
Crossing languages
AI also helps multinational teams communicate. An organisation may have an official business language, but people still understand information most naturally in their own language.
Translation tools have handled documents for years. What AI adds is that you can keep working with the material, asking questions of it and summarising it in either language, rather than just producing a translated copy. That makes information easier to distribute and easier for local teams to consume, and language becomes far less of a barrier to sharing knowledge across the organisation.
Prompting, without the fear
Prompting is the part that makes AI look harder than it is. You may have seen elaborate prompt formulas telling you exactly how to speak to a model. Those techniques have their place, particularly for repeatable tasks, but you do not need them to start.
Just ask. Use normal language and explain what you want much as you would explain it to a colleague. The clearer you are, the better the result, simply because there is less room for misinterpretation. Modern AI is also good at working through imperfect sentences, half-formed ideas and spelling mistakes.
Let the AI build the prompt. If you do want something more structured, ask the AI to write the prompt for you. Tell it what you are trying to achieve, who the audience is and what output you need. It may ask you questions; answer them, and together you refine the request. You can read the result, adjust it, and hand it on to another model.
Use images, not just text. Some models can interpret images. Take a screenshot, paste it into an approved tool, and ask "what is wrong here?" or "explain what I should check next." The goal is never to become an expert prompt writer. It is to describe the business problem well enough for AI to help.
The AI committee
For more important work you can go a step further. I call it the AI committee, and it just means putting reasonably comparable models to work on the same thing.
There are two ways to run it:
- In parallel. Ask two models the same question and compare the answers. If you have a third model, give it both answers and let it do the comparing: "Compare the two answers, identify agreements and conflicts, check the important claims against authoritative sources, and produce a single document containing the strongest supported points from both."
- In series. Take the answer from one model and give it to another: "Verify the statements made here, identify any errors or missing information, and suggest improvements."
Agreement between models is not proof that something is correct. For important claims, verification still means going back to the source.
One rule matters. The models should be reasonably well matched. There is little value in asking a basic, quick-response model to challenge detailed analysis from a far more capable reasoning model. It is like expecting a beginner to critique a chess grandmaster. And always tell the models to suggest changes only where they add value, otherwise you get a pile of cosmetic edits. The point of the committee is not more text. It is to find mistakes, expose gaps and improve the result.
Keep a human in the loop
AI makes mistakes. It can misunderstand a question, use incorrect information, and sometimes present a wrong answer with real confidence. So its work needs to be reviewed.
The principle is not new. People make mistakes too, which is why we review reports, check calculations and challenge proposals. If an intern prepares a piece of work, the manager is still responsible for what leaves the department. AI is no different. Saying something was "made with AI" does not move responsibility away from the person or team using it. If anything, reaching for AI signals that the team meant to go beyond the ordinary, so the output deserves the same ownership as anything else you put your name to.
AI should not be seen only as a tool to turn a one-hour task into a five-minute one. The first question should be whether AI can help you produce a better result: sharper analysis, a clearer document, better preparation for a customer discussion, or a risk you might have overlooked. Saving time is welcome, but the real objective is better business outcomes.
A look at where this is going
Now take the idea of AI as a team member one step further. Picture a board meeting a few years from now. On the table sits a small AI unit. It is listening to the discussion, with permission, and it has access to the information relevant to that meeting.
Someone asks, "what risks haven't we considered?" The AI answers through the room's audio while supporting information appears on a screen. It points out that two proposals conflict. It recalls a decision from an earlier meeting. It produces a figure nobody had to hand. Perhaps it offers another option.
It does not chair the meeting and it does not make the decision. It contributes, and the leadership team stays accountable. We can already talk to AI in natural speech, and models keep getting better at working with documents, images and organisational knowledge, so this is not hard to imagine.
The organisations that start using AI seriously today will understand how to manage that moment when it arrives. They will know which models suit which tasks. They will have prompt libraries for repeatable work. They will have their own AI agents supporting defined business processes. Most importantly, their people will already know how to work with AI, and that is where the operational advantage will come from.
Where to start
None of this needs a major transformation programme. Start with something useful, low-risk and measurable. You want to learn how AI behaves in your organisation before putting it anywhere near a process where a mistake has serious consequences.
It needs a first step:
- Start with an approved tool.
- Choose an appropriate task.
- Give it clear instructions.
- Review the result.
- Challenge it.
- Improve it.
- Capture what works and make it repeatable.
Treat AI as you would a new member of the team. Give it useful work, manage its limitations, and stay accountable for what it produces.
It all starts with one prompt.
This article accompanies my video on the same topic. You can watch it here: [coming soon].
Follow This, That and (Maybe), the Other:
Comments
Post a Comment