Starting an AI Project? Do the ROI Math First

Of all the AI projects I've watched in warehousing over the past few years, more died in the kickoff meeting than in the technology. The technical pitch sounds dazzling — then the CFO asks one question: "How much does this actually save us per year?" Silence. "Let's study it further." And further never comes.

So this piece isn't about technology. It's about money. Step one of any AI project is the ROI math — and if you can't make it pencil out, don't start.

The ROI worksheet: one table, five rows

Skip the fancy financial models. One table:

Annual cost: software subscription or license, plus implementation, plus internal labor (hours converted to dollars), plus hardware (cameras, servers, and the like).

Benefits, in three buckets:

1. Labor saved: hours eliminated × fully-loaded hourly cost. Note the word eliminated, not theoretically savable. Vision inspection replacing two inspectors at $45K a year each is $90K — but only if those two people actually get reassigned or not replaced. Otherwise it's paper savings.

2. Errors avoided: error-rate reduction × cost per error. Cut the pick-error rate from 0.3% to 0.1% on 10,000 orders a day, at $15 per error (reship + support + returns), and that's $300K a year. Most operators have never run this number. It's usually the biggest line on the table.

3. Capacity gained: incremental margin from higher throughput. Routing optimization that lets the fleet handle 5% more stops produces real margin.

ROI = (annual benefit − annual cost) / annual cost. Payback = annual cost / annual benefit × 12 months.

My red line: if payback exceeds 18 months, don't start. Warehousing moves fast and the technology moves faster — a three-year-payback project can get obsoleted halfway through.

Formulas alone feel abstract, so here's a complete worked example — a vision inspection project:

ROI = (205 − 37) / 37 = 454%, payback 2.2 months. Put that table in front of finance and the kickoff meeting ends in ten minutes. Notice the error line — $115K, the biggest number on the table, and the one most people never calculate.

Three scenarios where the ROI almost always works

Not every AI application deserves funding. From the cases I've seen, these three pencil out most reliably:

One, repetitive visual inspection. Quality checks, pallet counting, jam detection on conveyors — slow for humans, mind-numbing, error-prone. A $30–80K vision system replacing 1.5–2 FTEs typically pays back within a year. The scenario is bounded, the standard is clear, and the technology is the most mature.

Two, routing and route optimization. A 60-truck fleet saves $200K+ a year against a $30–40K subscription — roughly 6:1. The catch: data cleanup and rule extraction count as project cost. Put them in the table.

Three, document OCR. Bills of lading, packing lists, invoices — a clerk keys 200 a day; OCR clears 2,000 at 95%+ accuracy. In port and forwarder warehouses where clerical labor is expensive, this math is easy.

Conversely, the hardest ROIs to justify are AI customer service and demand forecasting. Service saves "experience," and forecasting needs six months of data seasoning before you can even measure it. If the worksheet won't fill in, don't force it.

After the ROI: four things that decide life or death

A positive ROI only buys you a ticket to start. Four more things decide whether the project survives:

First, data readiness. AI eats data. Three categories must be cleaned up in advance — identify what's missing, and cost the cleanup in the budget:

• Historical operations data: at least 12 months of orders, shipments, and returns. No history, no forecasting, no optimization — full stop.
• Master data: complete SKU dimensions, weights, and pack specs. Vision, routing, and slotting projects die on incomplete masters.
• Exception records: errors, late shipments, complaints, and their causes. Exception data is the training diet for "anomaly-detecting" AI — without it, the AI only handles the happy path.

Plenty of projects leave this line out of the worksheet, hit a data wall midstream, and blow the budget by 50%.

Second, put POC acceptance criteria in the contract. "Accuracy above 95%," "mileage savings no less than 5%" — in writing. Miss the target, withhold the final payment. A verbal "ours is very accurate" counts for nothing.

Third, name a business owner, not an IT owner. An AI project is a business project. The warehouse manager owns it; IT supports. Reverse that and it dies, because the people who live with the system are the operators.

Fourth, find a place for the displaced workers. The two inspectors the vision system replaces move to exception handling and auditing. Announce the reassignment plan up front — otherwise the front line resists, the data gets sabotaged, and no technology saves the project.

My four red lines for AI kickoffs

Tape these to the wall:

1. Payback over 18 months — don't start.
2. More than half the "benefits" are paper savings (nobody left, errors unchanged) — redo the math.
3. POC acceptance criteria not in the contract — don't sign.
4. IT leading with no business owner — don't do it.

Whether AI is worth it was never a technology question. It's an arithmetic question. And once the arithmetic is clear, the technology choice gets easy — because eight out of ten vendor solutions can satisfy a scenario whose numbers are honest.

Can your team fill in the ROI worksheet for the AI project you're evaluating? If not, forward this to whoever owns the kickoff.