Most AI projects fail on math nobody did up front. Here are the four numbers, the floor-and-ceiling method, and the costs everyone forgets when estimating AI ROI.
See how Leadify AI can grow your business.
No spam. Your information stays confidential.
A 60-person logistics company in Ahmedabad approved a budget of ₹42 lakh to build an AI system that would auto-classify their inbound shipment emails. Nine months later, the tool worked beautifully. It just never saved anyone any money, because the two staff members it was meant to free up got reassigned to a new warehouse and the email volume they were classifying turned out to be seasonal. The CFO asked one question in the review meeting that nobody could answer: "So what did we actually get back for this?"
That question should have been answered *before* the first line of code, not nine months after. This is the single most common failure in enterprise AI, and it has almost nothing to do with the technology. It has to do with math that people skip because it feels premature.
Here is how to do that math, before you commit a rupee to building.
Return on investment is not a mysterious concept. It is (gain minus cost) divided by cost, expressed as a percentage. The trouble with ai roi specifically is that both the gain and the cost are slippery, and people tend to be optimistic about one and vague about the other.
The gain is slippery because AI rarely replaces a whole job. It shaves 15 minutes off a two-hour task. It catches 8% more fraud. It answers 40% of support tickets without a human. Those are real, but they are partial, and partial gains are easy to overstate.
The cost is slippery because the build is only the beginning. There is data cleanup, integration, model monitoring, retraining, and the very real cost of the thing being wrong and someone having to fix it. Most ROI cases die on the costs people forgot to count.
So the discipline is not "estimate ROI." It is "estimate ROI honestly, with the pessimistic version written down next to the optimistic one."
You do not need a 40-tab spreadsheet. You need four numbers, and you need to defend each one.
What is the current process costing you *right now*, in money, per year? Not the theoretical cost. The actual one.
Take that logistics company. Two people spending three hours a day classifying emails. That is six person-hours daily, roughly 1,500 hours a year. At a fully-loaded cost of ₹450 an hour, the problem costs about ₹6.75 lakh annually. That is the baseline. Everything the AI does gets measured against it.
If you cannot write down a baseline in rupees, stop. You are not ready to build, because you have no way to know if you won.
AI will not eliminate the whole baseline. It will capture a fraction. Be brutal here.
Will the tool handle 90% of emails, or 55%? Will it do it with zero human review, or will a person still glance at every classification for the first year? A 55% capture rate with light review is a completely different business case than a 90% capture rate with none, and the difference is often the entire ROI.
For the logistics example, assume 65% capture with spot-checking. That is roughly ₹4.4 lakh of the ₹6.75 lakh baseline that AI could realistically take off the table each year.
The build quote is the number vendors love to quote. It is also the number that gets you in trouble. Total cost of ownership over three years usually breaks down like this:
A ₹42 lakh build can easily carry ₹8-10 lakh a year in run-and-maintain costs. Over three years, the true number is closer to ₹65-70 lakh. If your ROI only pencils out against the ₹42 lakh, it does not pencil out.
Divide the annual gain by the annual cost, and figure out when the cumulative gain crosses the cumulative cost. If the honest answer is "year four," and your business plans in two-year horizons, the project is a no regardless of how exciting the demo looked.
A rough rule from what we see work: enterprise AI projects that pay back inside 18 months get built and stay built. Projects promising payback in three-plus years tend to get killed halfway through when priorities shift, which is the worst possible outcome because you eat the cost and never reach the gain.
Never present a single ROI number. Present two.
The ceiling is the optimistic case: high capture rate, low maintenance, no scope creep. The floor is the pessimistic case: capture comes in low, data cleanup runs long, and you need more human review than you hoped. If the *floor* still clears your payback threshold, you have a genuinely good project. If only the ceiling clears it, you are gambling.
For the logistics case, the ceiling might show a 14-month payback and the floor a 26-month payback. That spread is the actual conversation to have with your CFO, not the single rosy figure.
Three line items sink more AI business cases than any modeling error:
If your ROI is strong but the problem is genuinely common (transcription, generic chatbots, standard document extraction), you should probably buy an off-the-shelf tool rather than build. The ROI on building something a vendor already sells is almost never there.
Building makes sense when the problem is specific to how *your* business works, when your data is a competitive advantage, or when no tool fits the workflow. That is the moment to look at custom AI software rather than forcing your process to bend around a generic product. And if you are not sure which side of that line your problem sits on, an honest AI consulting engagement will usually save you more than it costs by killing the bad projects before they start.
Before any AI build, fill in one page:
That last line matters more than people expect. Sometimes the cost of inaction (competitors pulling ahead, a manual process breaking at scale) is the strongest part of the case. Sometimes writing it down reveals that doing nothing is perfectly fine for another year.
If you want a sanity check on your own numbers, our pricing page lays out what different scopes actually cost, which is a better starting anchor than a vendor's opening quote.
The logistics company eventually rebuilt their business case the right way. The rebuilt version showed the original email-classification project never should have been the first thing they built. The AI that actually paid back, a route-optimization tool, had been sitting third on their list the whole time. The math would have told them that on day one, if anyone had done it.
What is a good ROI for an AI project?
There is no universal number, but a useful bar is a payback period under 18 months and a positive return even in your pessimistic ("floor") scenario. A project that only looks good in the optimistic case is a gamble, not an investment. Percentage returns matter less than payback speed, because faster payback survives shifting priorities.
How do I calculate AI ROI when the benefits are hard to quantify?
Convert soft benefits into a range, not a point estimate, and be conservative. "Better customer experience" becomes "a 2-5% reduction in churn," which becomes a rupee figure against your revenue. If a benefit genuinely cannot be tied to money or time saved, leave it out of the core ROI and list it separately as an intangible, so it does not quietly inflate your numbers.
Should I include data preparation in my AI cost estimate?
Absolutely, and it is usually 20-40% of the build cost by itself. Skipping data prep in your estimate is the most common reason AI projects blow their budget. Dirty, scattered, or incomplete data has to be cleaned and connected before a model can do anything useful, and that work is real, billable, and time-consuming.
Is it cheaper to build custom AI or buy an existing tool?
Buy when the problem is generic and a proven tool already fits your workflow. Build when the problem is specific to your business, when your data is a genuine advantage, or when no tool matches how you actually work. The ROI on building something a vendor already sells well is almost never positive once you count maintenance.
How long should an AI ROI calculation take?
For a first pass, a few days of honest work, not weeks. You need four defensible numbers: today's cost of the problem, a realistic capture rate, three-year total cost of ownership, and payback period. If you cannot produce those four numbers, that is a signal you are not ready to build yet, which is itself a valuable finding.
