At the start of this year, three AI vendors came knocking, one after another. Big models, visual recognition, smart scheduling — each slide deck flashier than the last. I asked them all the same question: how much money will this save me per month? Not one of them could answer. The most honest one said "it depends on your data quality". So I pay first, and the results show up later. Right. The warehouse AI project ROI business case has to be written before anyone talks to a vendor.

That is the most common way an AI project dies: buy the technology first, then go looking for a use case. The vendor's job is to sell you the system, not to make the numbers work for you. The numbers are your job, and they have to be done before you sign anything.

Three Numbers for the Business Case, Worked Out Before You Sign

The ROI of a warehouse AI project is not one formula. It is a table with three rows, and if any row is blank, the numbers in the pitch meeting are fiction.

Row one: what does this job cost in labour right now? Take the pick verification step in my own warehouse. Around 800 orders a day, two people on verification. Their wages plus benefits add up to a visible monthly number. Worse, the role has high turnover — peak season forces extra hires, the new hires are slower, and mis-picks actually go up. You do not need a vendor to calculate this. Three months of payroll records will tell you.

Row two: how much can the AI verifiably save? Note the word verifiably. A vendor saying "99.9% accuracy" is quoting a lab number. In your warehouse the question is different: out of 800 orders a day, how many wrong picks does it actually catch? And of the ones it flags, how many are real errors and how many are false alarms? Every false alarm costs your verifier several minutes, and if false alarms pile up, the labour you saved goes straight back out the door. I made vendors run a two-week live test in my warehouse first. Numbers come from running, not from a slide deck.

Row three: implementation plus maintenance plus the cost of people changing. The software fee is only the beginning. Do you need extra cameras, better lighting, a steadier network, several rounds of staff training, and someone to fix the system at midnight when it breaks? All of it costs money. The easiest cost to miss is staff turnover: when an experienced worker leaves, the replacement has to learn a whole new workflow, and productivity dips for the first month. Nobody writes that into a quote, but you pay it anyway.

Put the three rows together and the decision gets clear:

Which number How to get it If you cannot get it
Current labour cost Three months of payroll and mis-ship compensation records No data? Run one month of manual tracking first
What AI can save Two-week live test: caught errors and false-alarm rate Vendor refuses a live test? Find another vendor
Implementation, maintenance, staff change List hardware changes, training, and failure response Anything missing from the quote, assume you pay for it

That table is pinned on my office wall. Any AI project that wants approval has to fill in all three rows first. It is not that I don't believe in AI — I only believe in numbers my own warehouse produced.

The Approval Gate: One Metric, One Owner, One 90-Day Checkpoint

Once the numbers are clear, the project still has to pass a second gate. My rule has three parts: one measurable business metric, one named owner, and one 90-day checkpoint. Miss any one of them and the project does not start.

The metric has to be tied to the business, not fluff like "improve intelligence levels". For our visual verification project, the metric was one sentence: within 90 days, mis-shipped orders cut in half, false alarms kept within an acceptable range, and verification staff reduced from two people to one. Written on paper, checked on the due date, no negotiation.

The owner has to be someone from operations — not IT, and not the vendor. The vendor's team leaves when the project ends, and then who do you call? Our project owner was the verification team lead. He knows best whether the system actually works, and he is the first to refuse sign-off if it doesn't.

The 90-day checkpoint is the most important part. The killer of AI projects is "let's run it a bit longer" — one year later, the money is gone and there is still no verdict. At day 90, the metric is met and we expand, or it is missed and we stop and review. Stopping is not embarrassing. Dragging it out is what burns money.

The move that worked for me was starting with one small scenario. Visual verification went live on a single outbound line first. After 90 days, mis-ships were genuinely down and false alarms were under control — only then did we roll it to a second line. When we rolled it to the second line, the first two weeks were rough: false alarms spiked while the cameras adjusted to the new lighting, and for a week the verifiers were slower than before. We almost pulled the plug. Then the numbers settled, and by week six the second line was matching the first. The lesson: budget for a dip in the first fortnight of any rollout, and judge the checkpoint on the last month, not the first two weeks. I have seen the opposite too: another seller bought the full platform on day one, spent six months failing to get it running, and the system now sits idle gathering dust. The money was already paid.

Here is the contrarian point: a small scenario that actually works is worth more than a big platform. The bigger the AI rollout, the faster it dies; the smaller the first cut, the longer it lives. Let one step genuinely save money first, and only then decide whether the next step deserves funding.

And one honest truth: the most expensive part of an AI project is not the software, it is your own messy process. If your picking logic is a mess and your location codes haven't been updated in ten years, AI will only recognise the mess faster. Sort the process out first, then talk about AI. Don't get the order backwards.

So before you sign, ask yourself three questions: have I worked out these numbers? Am I willing to put the checkpoint in the contract? Can my process survive being amplified by AI? If you can answer all three, sign. If you can't, no matter how pretty the vendor's deck is, put it aside for now. Write the warehouse AI project ROI business case first. Then talk tech.

This approach worked for me — the money saved is real, and the lessons cost real money too.

By Ah K, E-commerce Seller (5 yrs)