If you lead operations, product, or a growing business, you have probably been offered “an AI layer” in the last year. The pitch is usually a dashboard: live insights, predicted demand, a chatbot that can answer anything. Six months later the dashboard is open on one monitor and the real work still happens somewhere else.
That is not an intelligence problem. It is a product problem. Practical AI automation starts with a narrower question: which recurring task should disappear?
If the system does not take a job off someone’s week, it is a report, not automation.
Where work actually hides
In most companies the expensive work is not the strategy meeting. It is the middle: copying an order from a message into a system, chasing a missing field, re-explaining a status, or reconciling two tools that should already agree.
Those tasks survive because they look cheap in isolation. One message. One exception. One “quick check”. Across a week they become a second job. Across a year they become headcount.
Look first at work that is:
- frequent, not rare;
- structured enough to describe, even if the input is messy;
- painful when it is late or wrong;
- currently held together by a capable person, not a process.
Fresh food wholesale is a clear example. Buyers still place orders in WhatsApp because it is fast. The cost appears later, when someone has to turn those messages into confirmed lines. That is why we built dmcart: not to add another commerce dashboard, but to turn informal messages into clean orders.
A four-step way to choose the first AI project
1. Watch the work before you name the model
Sit with the people who do the job. Record the inputs they actually receive: emails, photos, voice notes, half-complete forms. If you start from a vendor demo, you will automate the tidy version of the process, not the real one.
2. Quantify one loop
Pick a single loop and put numbers on it. How often does it happen? How long does it take? What breaks when it is late? “Customer support is messy” is not a brief. “Eighty order amendments a day, each taking six minutes, three of them creating a delivery error” is a brief.
3. Automate the boring middle
The highest-leverage AI systems do not replace judgement. They prepare it. Extract the lines. Draft the reply. Match the SKU. Flag the exception. Leave the unusual case with a person who can see the context.
This is also how you keep trust. Staff adopt a tool that finishes the dull part of their day. They resist a tool that claims to do their job and then fails on the one order that matters.
4. Measure hours returned, not model novelty
Accuracy matters, but it is not the business outcome. The outcome is time given back, fewer corrections, and a cleaner handoff into the system of record. If you cannot name the hours or the error rate you expect to move, you are not ready to build.
Write the before-and-after week of one role on a whiteboard. If the after still contains the same copy-paste, status chase, or rekeying, the AI idea is decoration.
What not to automate first
Skip the wide, impressive use cases until a narrow one is in production. A company-wide knowledge bot, a strategy copilot, or a “single brain for the business” sounds ambitious. It also has no owner, no success metric, and no natural place in the working day.
Delay projects that:
- need perfect data you do not have yet;
- touch regulated decisions before the workflow is stable;
- try to replace a trusted relationship, not a repetitive task;
- cannot be rolled back if the model is wrong on Tuesday.
Good intelligent systems are boring on purpose. They sit in the path of work that already exists. They make that path shorter.
Keep people in control
Automation without a review path creates silent failure. The operator needs to see what the system believed, change it, and teach the next pass. That is not a nice-to-have interface. It is the product.
Design the exception queue first. Then design the happy path. Teams that reverse this order ship a demo. Teams that start with exceptions ship an operation.
How we approach this at Techturf
From Bengaluru and Swindon we help companies turn messy operational work into systems people will actually use. The work is part product strategy, part engineering, and part sitting close enough to the process to see where time leaks.
If you have a process that is held together by chat, spreadsheets, and one indispensable person, that is usually the right place to start. Not because it is fashionable. Because removing that work changes the week.
Read next: why building the right thing too late still fails, or talk to us about an automation brief.