How to Implement AI in Your Project Without Wasting Money

I've reviewed about a few AI implementation plans in the last few months. Most of them start with the same mistake… they begin with the AI.

They pick a tool like ChatGPT or Claude, maybe even something custom, and then work backwards to find a problem it might solve. It's understandable. AI is exciting. But it almost always leads to building something clever that nobody actually needs.

Here's what actually works when you implement AI into a project: start with the problem, then the technology.

Step 1: Define the Problem

Before you touch any AI tool, write down the specific problem you're trying to solve. Not "we want to use AI" or "we need to be more efficient." Those aren't problems — they're vague aspirations.

A proper problem statement looks like this: "Our customer service team spends four hours a day answering the same twelve questions, which delays responses to complex queries by an average of six hours."

Notice the specificity. You can measure that. You can tell if AI actually helped.

I worked with a health tech team last year who wanted to implement AI for patient triage. When we dug into it, the real problem was that reception staff were spending 20 minutes per patient manually entering data from PDFs into their system. The AI solution wasn't a triage bot, it was document parsing. Completely different implementation, same underlying frustration.

Your problem statement should answer three questions: What's happening now? Who's affected? What's the cost (time, money, opportunity)?

If you can't answer those, you're not ready to implement AI yet. You're ready to do discovery.

Step 2: Map the Workflow

Most AI implementations fail because people skip this step. They assume the workflow is obvious when it rarely is.

Draw out — literally, on paper or a whiteboard — what happens now. Every step. Every decision point. Every handoff. Include the bits that feel trivial, because that's often where AI can actually help.

Then draw what you want to happen. Where does AI sit in that flow? What does it do? What does a human still need to do?

Start with one use case. Don't try to implement AI across five workflows at once. Pick the one that's highest value and lowest complexity, prove it works, then expand.

Here's the critical question: does the AI replace a step, augment a step, or create a new capability?

Replacing a step means the human stops doing that task entirely (like automated document parsing). Augmenting means the human still does it, but faster or better (like AI-assisted writing). New capability means you can now do something you couldn't before (like real-time translation for customer support).

Be honest about which one you're doing, because each has different implementation requirements and change management needs.

In practice, most successful AI implementations are augmentation, not replacement. The AI drafts the response; the human reviews and sends it. The AI flags the anomaly; the human investigates. That's a solid implementation, building trust and catching errors before they reach customers.

Step 3: Choose the Simplest Tool That Works

Once you know the problem and the workflow, now you can pick the AI.

People often look to the latest and most exciting things first. I've seen teams spend three months training a custom model for something ChatGPT could do out of the box. They did it because "we need to own the AI."

I disagree with this approach. Don't build a custom model if an API call to GPT-4 does the job. Don't use GPT-4 if a simpler classification model works. Don't use a model at all if a well-designed form and some basic logic solves it.

Don’t let ego get in the way of strategy.

Here's a simple decision tree:

If your task is general-purpose (writing, summarising, translating, answering questions): Use a pre-built LLM like GPT-4, Claude, or Gemini via API.

If your task is domain-specific but you have training data: Fine-tune an existing model or use retrieval-augmented generation (RAG) with your documents.

If your task requires consistent, repeatable outputs with high accuracy: Consider a smaller, specialised model or even rule-based logic with AI as a backup.

Most projects I work on land in the first category. They're using OpenAI or Anthropic APIs, wrapped in a simple interface, with guardrails to catch hallucinations.

Step 4: Build, Test, Govern

Now you build. Start with a prototype. A working example that handles the core workflow, even if it's rough around the edges. Show it to the people who'll actually use it — not just leadership.

I worked with an operations team who built an AI assistant to help with scheduling. Leadership loved it. The ops team hated it, because it didn't account for juggling shift swaps and last-minute changes. We rebuilt it in a week based on their feedback. If we'd gone straight to production, it would have been months of wasted effort.

Once the prototype works, test it properly. How accurate is it? What happens when it's wrong? How do users know when to trust it and when to double-check?

Then — and only then — put governance around it. What data can it access? Who can use it? How do you handle errors? What gets logged?

Governance is sometimes seen as a means of killing innovation. But in this case, it's about making sure your AI implementation doesn't accidentally leak customer data, give bad advice, or create legal liability. This step isn't optional.

The teams that do this well treat AI like any other system: clear access controls, audit trails, and a process for when things go wrong.

What Success Looks Like

You'll know your AI implementation worked when people stop talking about the AI. They’ll just use it, because it’ll become a natural part of the workflow.

The customer service team won’t say "we use AI" — they’ll say "we respond faster now." The ops team won’t show off the tool — they’ll just get the schedule done in half the time. That's the goal. A problem that used to exist and now doesn't.

If you're trying to figure out where AI actually fits in your project, I run a half-day session that walks through this exact framework with your team. No sales pitch, just clarity on whether AI is the right move and what to build first. Book an AI Basics Session here.