AI integration strategy

Many organizations are exploring AI integration in their products or internal processes. In a new and fast-moving field of modern AI systems, that can be hard to do for a variety of reasons:

• There is a lot of hype, unrealistic expectations and claims of massive benefits (such as Facebook’s plan to "replace junior developers with AI"), and bold predictions of imminent AGI (Artificial General Intelligence) or the Singularity. Meanwhile, real-world results are mixed. While AI has clear potential, distinguishing between genuine potential and overblown promises can be difficult.

• Large Language Models (LLMs) are getting most of the attention, but they’re just one type of AI. Other machine learning (ML) models may be better suited (faster, cheaper, more accurate...) for certain use cases, or a layer above the LLM (RAG, agentic, reasoning, ...) may be required.

• All of these approaches are evolving rapidly, making "best practices" and in-house expertise quickly outdated. You should be open to "thinking outside the box" and applying AI in novel ways, but at the same time critical while evaluating its feasibility.

At the end of the day, AI is just another tool: an implementation detail, not a goal in itself. Companies don’t benefit from AI adoption: they benefit from product or process improvements AI enables, but only when it’s the right fit. Don't fall into the trap of chasing "AI transformation" without a clear purpose and plan. You'll just waste time and money and result in failure.

To avoid this, you need a clear, practical AI strategy that defines why and where AI will be used, how it will be implemented, and how results will be measured against expectations:

1. Start small until you build up expertise and confidence.

When selecting tasks for AI optimization, start with initiatives that require minimal effort but offer substantial returns. These early wins build confidence and experience, allowing your teams to refine their approach before tackling more complex implementations. Once familiarity with AI grows, you can pursue higher-effort projects that promise even greater rewards.

2. Analyze your current workflows to find a good fit.

Closely examine existing workflow in these tasks, to understand parts that can be improved for higher quality, spending less time, overall efficiency, and so on. Visualize an ideal future workflow, compare with your current state, and find where the gaps are that could potentially be solved with AI.

To do this, you'll need up-to-date knowledge of current AI strengths and weaknesses to understand which tasks may be improved without sacrificing quality or reliability. This is the part where hype can most easily mislead you, so be on the lookout for real vs. advertised capabilities. Keep it real and ignore the hype.

3. Do not ignore data governance.

You need to consider data confidentiality, privacy, and availability. You'll have to figure out what data is required, whether it is already accessible, and how it can be collected and processed.If data ingestion is necessary, you'll have to ensure compliance with any applicable data laws, regulations, or internal policies.

Noisy, low-quality or bad data is often cited as number one reason for unsuccessful AI initiatives, and data gathering, cleanup and preparation may take a sizable portion of time and resources from your AI budget. To help with this, in some cases we can use AI to clean up the data, or even help generate it (with human input) to unblock the rest of AI implementation.

4. Choose the best tech for the task.

Based on task and data analysis, you have solid ground to design a workflow that integrates AI into the product or process.

When designing the solution, understand there isn't a one-size-fits-all solution, so you should be aware of the different potential technologies you might leverage, with their strenghts and weaknesses. For example general-purpose LLMs vs. an optimized model, vision LLMs vs. classical OCR solutions, RAG vs. fine-tuning, agentic vs reasoning models, and so on. On another axis, you have a choice between AI-as-a-service vs. running your own infrastructure.

Whatever your choices, you'll need to balance speed, latency, cost, and quality according to the specific needs of the task.

5. Have a human in the loop.

Modern AIs are very capable but they make mistakes. With the current LLM architecture, it's basically impossible to create an AI system that will guaranteed, 100%, always be correct. This does not make them useless, but we do need to account for this.

There are two approaches to this:

First, have a human in the loop. A trained operator that can spot and correct or block incorrect responses or actions from the AI can ensure the high quality of the output while leveraging AI for productivity gains.

Another approach, for cases where errors are not large and can be easily undone or fixed after the fact, is to automate fully, recognizing there will be some errors (which we still want to minimize with proper AI engineering). Those errors can be spotted after the fact, escalated to a human operator, and fixed manually.

6. Measure to understand impact.

How will you know if your AI implementation is successful? Many AI systems (like chatbots) are measured using "vibes" - anecdotal evidence from scattershot manual testing. That "looks fine to me" kind of approach can't provide you confidence your AI strategy. Instead, measure.

Before implementing AI, measure the systems that will be impacted to understand the baseline level of quality, productivity, etc. Then, implement a proof-of-concept (PoC) solution, and measure again to see improvement. If it compares favorably to your expectations, proceed to full implementation, integration and rollout. Otherwise, do a post-mortem investigation to see why reality differs to your expectations and how you should course-correct.

Like any technological transition, integrating AI into your product can be challenging, and requires having clear goals and strategy for adoption.

What are you trying to achieve, and how will you know if you've succeeded? This structured approach to AI integration will help you create a clear, informed strategy and avoid the pitfalls and the hype.