
What are the true costs of integrating Object Detection into an FMCG manufacturing company?
Earlier this week, we broke down "How DeepSeek AI will change the game for FMCG manufacturing" in this blog post here. But noticeably there was no mention of Object Detection. This is because DeepSeek focuses on LLMs not Object Detection models. To truly create a step by step winning strategy for AI implementation we need to deep dive into What are the true costs of integrating Object Detection into an FMCG manufacturing company.
And to build a robust Object Detection model that your competition will envy, let’s sprinkle in the expertise of Glenn Jocher wherever we can. If you didn’t already know, Glenn Jocher is an internationally recognised Object Detection expert and creator of Ultralytics, YOLOv5 & YOLOv8.
Step 1: Begin Image Collection Now
Image collection is the foundation of your object detection project, starting early and staying consistent is critical. Training a model requires extensive labeled data, start collecting ASAP.
A common pitfall is pausing or abandoning data collection mid-way due to competing priorities.
Tesla, on the other hand, leveraged its massive data collection to set a high barrier for competitors. For FMCG leaders, the takeaway is clear—start today, even if resources are limited. Future access to proprietary data can dramatically improve accuracy and performance of your object detection project. Here is Glenn Jocher’s ironic comment on Tesla’s large datasets, (this was 5 years ago… imagine how much bigger Tesla’s dataset is now).
Quick Tip: Test existing solutions, such as YOLOv8-world, to assess whether off-the-shelf models meet your MVP needs. This saves time now, while you gather your images for training your own future models.
Step 2: Map Your Current Processes
This is simple - hopefully your current processes are already mapped. The objective is to understand how data flows and how goods are managed. This should be used to influence your design.
You’ll need to ask the product owners to access two key documents:
Data Processing Models: What happens to the data today?
Product Overviews: What happens to the goods today?
If these documents aren’t readily available, it signals a knowledge gap that could be the cause of this process's inefficiency, regardless it should be corrected.
Quick Tip: Don’t take on responsibility for creating these documents yourself—it’s the domain of other teams. But this is an indicator of significant Technical Debt which should be addressed.
Step 3: Measure the current results
You should have an idea of what your business values most, optimise for that. It is important to align your project to the business strategy to increase stakeholder engagement.
Quick Tip: Speak to the operators and understand if there are any hidden costs that the C-Suite will not know about.
Step 4: Design an MVP—Fast and Simple
An MVP (Minimum Viable Product) is your entry point to test object detection’s feasibility. Avoid overengineering; the goal is to develop something functional, not perfect. This early prototype serves to validate your approach without excessive time or financial investment.
Quick Tip: Include the team in the design, make it fun and create a competition to improve adoption. The fewer moving parts the better.
Step 5: Build an MVP—and choosing Object Detection model
After physically building your MVP you will need to use some level of programming to collect the images and feed them to an Object Detection model, choosing a pre-trained model will be less accurate in your specific business case but is very quick to get started.
Quick Tip: Leverage pre-trained models to reduce development time. A faster launch ensures you can start gathering performance metrics sooner.
Step 6: Measure MVP Performance
Your MVP must demonstrate clear, measurable improvements over current processes. For FMCG businesses, key metrics often include quality, speed, and waste reduction. Identify which metrics matter most for your business—stacking these in favor of your MVP will build a stronger case for scaling.
For example, a reduction in noise might seem impressive, but if your primary goal is reducing waste, focus your analyd a Data-Driven Business Case
Your business case should be based on the outcomes of Steps 1–6. Instead of investing heavily in theoretical ROI models, use real-world data from your MVP to make the case for scaling. Demonstrating tangible results will resonate more with stakeholders than a polished pitch deck.
Solicit feedback from teams to refine the business case and ensure alignment. Patience is key—set realistic expectations and recognise that the journey to full AI adoption is iterative.
The Cost Breakdown: What to Expect
For C-suite leaders, the true cost of AI goes beyond the initial investment in hardware or software. Here’s a high-level breakdown of where costs typically arise in an object detection project:sis there.
Quick Tip: Prioritise metrics aligned with strategic objectives to keep stakeholders engaged.
Step 7: Understand CPO Before Moving into Production
Process | Cost | Purpose | Outcome |
Data Collection | £ | Foundation for training AI models | Enables future accuracy improvements |
Data Storage | £ | Manage large datasets efficiently | Supports scalability |
Data Labeling | ££ | Annotate images for training | Critical for model performance |
Current Process Mapping | 0 | Should already exist | Reduces disruption during integration |
Designing an MVP | 0 | Utilise existing talent (e.g., interns) | Boosts adoption and speeds up testing |
Creating an MVP | £ | Build and test early prototypes | Validates feasibility |
Building a Business Case | 0 | Based on MVP data | Secures leadership buy-in |
Future Process Mapping | £££ | Implement at scale | Significant ROI potential |
Building & Implementing | £££ | Deliver on the plan outlined | Noticeable realised increase in measurables |
The costly data labelling stage can be postponed until after your project has proved to be successful if you use an existing Object Detection Model like YOLO and have results at an acceptable level.
By far the most expensive stages are future mapping, build and implementation.
The true cost of these are largely dependent on the software you choose, along with the correlating technical debt. The Hardware you choose, along with ongoing maintenance and repair. The scale at which you implement your solution.
But by validating the MVP, you only encounter the larger costs once your project has proved it can generate the expected increase in measurables with a high likelihood of success. Which you have already demonstrated is a viable project.
Take action now
To get started you will need to answer these questions:
Who do you need to go to, to get the Data Processing Models of the process?
Can you already start collecting more data immediately - if not what is stopping you?
What do you have to do to overcome this hurdle?
What is the problem you are trying to solve?
Who are your key stakeholders?
Is this project aligned to the business strategy?
What is the measurable(s) you will be tracking?
What is the current state of the measurable(s)?
What is the cost(loss) of the current state?
Who is going to help you design, build and test your solution?
More details
If you’re ready to assess your company’s readiness for object detection, our team can help. Contact us to get started with our Smart Industry Readiness Index and take the first step toward transforming your operations.
Look out for our next series of posts: