The Best AI Projects Start With a Workflow Everyone Hates
Most companies don’t need an “AI strategy.” They need one painful workflow replaced.
That sounds almost too simple, but it’s the pattern I keep seeing. The businesses with the clearest AI opportunities usually don’t start with a model, a chatbot, or a big transformation roadmap. They start with a process that everyone already knows is broken.
The workflow is slow. It depends on one experienced person. It involves too many handoffs. It lives across PDFs, spreadsheets, emails, screenshots, internal systems, and tribal knowledge. And because the team has been dealing with it for years, it starts to feel normal.
That is where AI can actually matter. Not as a vague “let’s use AI somehow” initiative. As a workflow replacement.
The hidden AI opportunity is usually boring
When people talk about AI, the conversation often gets abstract fast. What model should we use? Should we build an agent? Do we need RAG? Should we fine-tune something?
Those are real questions eventually, but they are almost never the right place to start. The better starting question is:
What workflow is costing the team hours every week?
That question is much more useful because it points at actual operational pain.
A few examples:
- A production team recreates the same customer-facing asset over and over.
- A finance team reconciles messy reports by hand every week.
- A sales or ops team copies data between PDFs, spreadsheets, and internal tools.
- A manager depends on one person to translate messy source material into a usable report.
- A customer support team manually classifies requests that could be structured automatically.
- A telecom operator compares rate decks, routing data, and CDRs with too much human glue in the middle.
None of these sound glamorous. That’s the point.
The highest-value AI work is often hiding inside boring operational workflows that already have obvious costs.
A real example: apparel proofing
I’m working on one of these right now in apparel production. The workflow is customer proofing.
Before an order can move forward, production artists need to prepare visual proof/mockup assets for customer signoff. That means working with garment templates, customer artwork, placement, sizing, visual presentation, and the details needed for the customer to approve what will be produced.
The existing process can take production artists anywhere from 4 to 8 hours for a single proof. That is not because the people are slow. It is because the workflow is heavy.
There are artifacts to collect, templates to use, artwork to position, decisions to make, and a final output that has to look good enough for a customer to trust.
This is exactly the kind of work where “just add AI” is the wrong framing. The goal is not to remove the production artist from the process. The goal is to remove the repetitive setup work so the production artist can spend their time on judgment, polish, and exceptions.
The first version of the system I built can generate the initial proof direction in about 30 seconds. There is still human review and arrangement involved — and there should be. But if the final workflow becomes 15 to 30 minutes instead of 4 to 8 hours, that is a major operational improvement.
That is the difference between AI as a demo and AI as leverage.
The playbook is not magic
The playbook for this kind of work is straightforward:
- Find the bottleneck.
- Understand the real process.
- Build the smallest useful system.
- Keep humans in the loop where judgment matters.
- Measure the before and after.
The hard part is not usually the model. The hard part is understanding the workflow well enough to build something the team will actually use.
That means asking practical questions:
- What does the team do today?
- Where does the work slow down?
- What inputs are messy?
- What outputs need to be trusted?
- What does a human still need to review?
- What would make this production-ready?
- What would save real time this week?
Once you know that, the technology decisions get much clearer. Maybe you need a vision model. Maybe you need OCR. Maybe you need structured extraction. Maybe you need a simple internal app. Maybe you need an agentic workflow with multiple steps and human approval. Maybe you don’t need anything fancy at all.
But you cannot know that until you understand the bottleneck.
The best AI systems keep people in the right places
A common mistake is assuming automation means removing humans. In real operations, that is usually not the right goal.
The better goal is to move humans out of repetitive setup and into review, judgment, and exception handling.
For the apparel proofing workflow, the production artist still matters. They understand quality. They understand what a customer will accept. They understand edge cases. They can see when something feels off.
The system should not pretend to replace that judgment. It should get them to the judgment point faster.
That is a much more practical way to think about AI adoption inside real businesses. Do not ask:
How do we automate the entire job?
Ask:
Which part of this workflow should a human not have to do from scratch every time?
That question usually leads to better systems.
How to spot a good AI workflow candidate
Not every workflow is worth automating. A good candidate usually has a few traits:
1. It happens repeatedly
If the workflow only happens once a year, it may not be worth building a system around. But if it happens every week, every day, or multiple times per day, the ROI can add up fast.
2. The inputs are messy but patterned
AI is useful when the source material is not perfectly structured, but still follows recognizable patterns.
Examples:
- PDFs
- screenshots
- spreadsheets
- emails
- customer files
- call records
- product specs
- invoices
- design assets
3. The output has a clear definition of “good”
You need to know what the system is trying to produce: a proof, a report, a cleaned dataset, a reconciliation result, a customer response, a routing recommendation, or a prioritized lead list.
If nobody can define the desired output, automation will drift.
4. A human can review the result
The best early systems usually have a human approval step. That makes the system safer, easier to adopt, and easier to improve.
5. The time savings are obvious
If a workflow takes 4 to 8 hours today and can become 15 to 30 minutes, you do not need a complicated ROI story. The value is visible.
Why “AI strategy” often misses the point
The phrase “AI strategy” can be useful, but it often pulls teams too far away from the work. A lot of businesses do not need six weeks of strategy before they can learn something useful.
They need to pick one workflow and test whether AI can create leverage. That first workflow teaches you more than a deck ever will.
It reveals:
- how clean or messy the data really is
- how the team actually works
- where human review is needed
- which integrations matter
- whether the ROI is real
- what the next workflow should be
That is why I like starting small. Not small as in unserious. Small as in concrete.
One workflow. One system. One measurable before and after.
The question I’d ask any operator
If you run a business or a team, the question I’d ask is simple:
What workflow does your team hate doing, but still has to do every week?
That is the starting point. Not because every hated workflow should be automated, but because that question usually surfaces the places where time, expertise, and attention are being wasted.
From there, you can ask:
- How often does it happen?
- Who owns it?
- How long does it take?
- What tools does it touch?
- What files or data does it depend on?
- What does the final output need to look like?
- What would a 50% improvement be worth?
- What would an 80% improvement be worth?
Now you are having a real AI conversation. Not a hype conversation. A business conversation.
The opportunity
I think a lot of the best AI work over the next few years will happen in businesses that are not trying to look like AI companies: apparel companies, telecom operators, manufacturers, distributors, agencies, finance teams, and operations-heavy companies with messy internal processes.
The opportunity is not to sprinkle AI on top. The opportunity is to find the bottlenecks that are already costing time and build useful systems around them.
That is the lane I care about: AI systems that survive contact with real work.
Want to test one workflow?
I’m opening up a few free 15-minute AI workflow audits.
Bring one manual workflow your team hates. I’ll tell you whether I think it is a real automation candidate, what I would build first, and where the ROI might be.
No slide deck. No vague AI transformation talk. Just one bottleneck and a practical next step.
Start here: