Bringing AI into your business? What you need to know

Someone in your company has to decide which AI ideas are worth pursuing, what needs to be built and who should build it. Then, usually without much ceremony, everyone turns to you.

You now have influence over the money, the supplier and whether the system delivers anything useful. This edition gives you the AI knowledge required to make those decisions well.

So let’s get you ready.

An AI Champion turns worthwhile business problems into controlled, measurable AI systems.

So you may have found yourself in the position of being the company’s ‘AI person’. Your actual title might be operations manager, product lead, finance director, transformation lead, or simply the person who asked one sensible question about ChatGPT in a meeting.

Congratulations. The job is now yours.

For this newsletter, I’m going to call that person the AI Champion. It means the person inside the company who finds, shapes and leads AI initiatives, regardless of the title printed on their email signature.

That gives you a sizeable chunk of influence. You will help decide which problems get attention, which ideas receive money, which suppliers get trusted, and whether the finished system creates value.

When it works well, the title starts to make sense.

The job

Most of this job will feel familiar to anyone who has led a software project. The technology is newer and some of the risks are different but the way it works is more or less the same.

Your responsibility is to:

1. Find a problem that matters and understand its root cause.

2. Work with the people who experience the problem to understand how the work happens today.

3. Define what success would look like in business terms.

4. Shape and specify a solution that an internal team or external provider can build.

5. Get realistic costs and choose who should deliver the work.

6. Run the quickest sensible pilot that can prove or disprove the idea.

7. Integrate the system into the company’s processes, controls and quality management system, then measure the result.

The sequence matters because starting with the problem gives you a way to judge whether AI belongs there in the first place. Other approaches are like beating around the bush.

The five skills

You probably already have parts of these skills from your current role. Now you just have to learn to apply them to AI projects.

1. Business problem discovery and root cause analysis

A useful AI project begins with a problem that affects cost, revenue, risk, service or the ability of people to do their jobs. Now that you have a meaningful problem we need to find the root cause of it. After that we have to decide whether the problem is big enough, affects many people or happens many times over to solve.

2. Solution design and specification

You need to work with the affected team to describe

• how the solution should fit in,

• what information it needs,

• what decisions it can make,

• what actions it can take and where a human must remain involved.

Then you need to document that clearly enough for an implementation team to cost and build.

3. Commercial and supplier judgement

AI is new, in demand and has a lot of hype around it. That makes vague pricing very easy. You need to compare realistic options, understand the ongoing costs, test supplier claims and control scope so the project doesn’t cost an arm and someone else’s leg.

4. Delivery, adoption and business integration

A product creates very little value when its not used. You need to:

• manage delivery,

• involve the users and other stakeholders,

• redesign the surrounding workflow and

• make sure the new system fits the company’s security, privacy, governance and quality processes.

Change is hard, so adoption has to be designed into the work.

5. Measurement, governance and reporting

You need a baseline, a small set of useful measures and a clear view of who needs to know what. That means measuring the business result, the quality of the AI’s work, the risks that need attention and whether people are actually using the system.

The engineering technical knowledge can sit with developers, architects and specialist service providers. Your job is to understand the problem, shape the solution and make sure the work delivers a result.

A turtleneck and saying “hmmm, interesting” repeatedly  remains optional.

The AI part

That brings us to the knowledge you need to make good decisions about AI. A clear mental picture of what the tech can do, how an AI system comes together and where it can fail gives you enough technical fluency to work well with specialists.

1. AI outputs are probabilistic

Traditional software is generally deterministic. The same input, processed under the same conditions, produces the same output because developers have defined the rules it follows.

Generative AI is probabilistic. It estimates a useful response from patterns learned during training, so the same prompt can produce different wording, levels of detail or conclusions across separate runs.

The variation may be minor, such as different wording in a WhatsApp message draft.

It can also affect the substance. Give a model an ambiguous transcript twice and it may assign an action to different people, include different details or reach a different conclusion.

A well-designed AI system uses both approaches. Deterministic software handles fixed rules and predictable actions. The model handles work requiring interpretation, judgement or generation. This distinction helps you decide where AI belongs, where ordinary software is safer, and which steps need human approval.

A good-looking answer is only one example of the output. You need to test the system across representative cases and decide how much variation the business can accept.

2. The model is one component of the system

ChatGPT is an application. GPT is a family of models used inside applications. Claude is also a product built around a family of models.

A model can receive instructions and context, then generate text, structured data or a request to use a tool. It is the reasoning engine inside the solution.

A complete system also needs:

• an interface,

• software logic,

• data,

• tools,

• storage,

• controls and

• a way to measure whether it is working.

Choosing a model matters but the overall solution around it determines whether it generates value for the business or not.

3. Context determines quality

A capable model can still produce bad work when it lacks the information required for the task. Context is the material supplied while the system is working:

• business rules,

• examples,

• current data,

• project history,

• customer information and

• the specific request in the form of the prompt.

A context pipeline decides what information to retrieve, how to structure it, when to supply it and how to keep it current. This is where much of the real implementation work sits. The model can only reason over what the system makes available.

4. Agents can decide and act

An agent combines a model with a goal, instructions, tools, state and boundaries. It can decide the next useful step, use a tool, inspect the result and continue until the task is complete or a human decision is required.

Some steps should remain deterministic. Filing a transcript in an agreed folder or sending approved data to a known system can follow fixed software rules. Other steps, such as deciding which project a discussion relates to, may require the model’s judgement.

A useful agentic system combines both and makes the boundaries explicit.

5. Evals turn ‘looks good’ into repeatability

An eval is a repeatable test of how an AI system performs across representative cases. The result may be a score, classification, rubric or threshold, depending on what good performance means for the task.

For meeting minutes, you might test whether decisions were captured, actions were assigned to the right people, dates were extracted correctly and sensitive information stayed where it belonged.

When an eval exposes a failure, error analysis helps you trace the cause. The problem may sit in the source data, context, instructions, workflow, tool connection or model.

Monitoring and feedback then show whether the system continues to perform once real people begin using it.

The components of an AI solution

These five concepts become easier to place when you can see the whole system. Most AI solutions contain some version of the following components:

• The user interface, where a person starts work, reviews a result or gives approval.

• Orchestration, the software logic that coordinates the steps.

• Agents, which use judgement to work towards a goal.

• AI models, which interpret information and generate outputs or tool requests.

• Context pipelines, which find and supply the right information at the right time.

• Tools, MCPs and APIs, which connect the AI to other systems and let it take action.

• Storage, state and memory, which record data, progress and useful history.

• Evals, monitoring and feedback, which measure quality and expose failures.

• Infrastructure, which runs, secures and scales the system.

I’m developing an AI System Basics Map that explains the five concepts, goes into more detail on what makes up AI systems and shows how these components fit together in one AI solution. If you want it, let me know and I’ll send it over when it’s ready.

What we are building

Over the next few editions, we will use one simple keep-in-contact bot as the running example. Its job is to help you maintain valuable relationships before they quietly go cold.

The bot can review your contacts and recent conversations, identify who may need a follow-up, bring the relevant context together and draft a message for your approval. You can start with ChatGPT, Claude or Copilot and a simple spreadsheet, then connect it to a CRM, email or calendar as the solution becomes more useful.

It is small enough for someone to build, start using and develop further. It is also valuable because relationships are easy to neglect when the day gets busy. Through this one example we can work through the full implementation lifecycle: defining the problem, setting the rules, supplying context, designing memory, connecting tools, keeping a human approval step, testing quality, measuring value and improving the system.

Your first action

Start mapping the problems and use cases in your business with the use-case scoring spreadsheet from the welcome email. If you did not receive it, reply ‘Yes’ and I will send it over.

Later, we will use the framework to refine and select the strongest use cases. I will also choose one reader-submitted use case for a future public teardown, where we will work through the problem, solution shape, risks, costs and measures together.

You now have the role, the five skills, the five AI concepts and one place to begin: find a real problem worth solving. Once you can describe that problem properly, you can start shaping a system the business can actually use.

Whats happened this week

Bill gates essay - we need more high-profile people saying this because it's important. We have been saying from the start of the AI boom that this transition will be turbulent and that most businesses are not prepared for it. Its to help the company and its people prepare now, while there is still time to shape how the change happens.

±6000 execs were surveyed by NBER and 90% reported no impact on employment or productivity over the last three years. So don't be one of the companies that just buys licences and expects a business result to appear. Be deliberate, pick the use cases well, rebuild the workflows around AI and make sure you can measure a gain. A task getting faster means very little if the business process around it stays exactly the same.

Anthropic notes that its automated AI researcher outperformed proposals from experienced human safety researchers in a controlled test. It also found an alignment method that used roughly 15,000 times fewer training examples than Anthropic's production process. This is narrow research. It is also an early sign that AI can beat human experts at parts of expert work. So the line around which steps still need human judgement is starting to shift. Keep an eye on this in your own area of expertise too.

116 organisations signed a letter warning that we have a limited window to strengthen cyber defences before AI-enabled attacks become more widespread and sophisticated. The message for every business is simple: fix the basic weaknesses now and raise the security bar for anything you buy, build or deploy, including AI-generated code. Waiting for the first serious attack will be a very expensive way to find the gaps.

Mathematician Terence Tao coined the term ‘proof indigestion’ for what happens when AI produces mathematical proofs faster than people can verify, explain, publish and absorb them. The same problem will show up in business. AI will make it cheap to produce reports, analysis and recommendations, while human attention stays expensive. So the new bottleneck will be deciding what is good, what matters and what gets acted on.