How to Calculate ROI on Custom AI: A Practical Formula for Indian MSMEs

To calculate ROI on a custom AI project, subtract the total cost of ownership from the total measurable benefit over a defined period, then divide by the total cost. The formula is (Benefit − Cost) ÷ Cost × 100. The difficulty is not the arithmetic. It is being honest about which benefits are real cash and which are only capacity, and remembering to include running costs, ramp-up time and internal review effort.

Most AI ROI calculations you will see in a vendor deck are wrong in the same three ways. They count hours saved as if those hours convert directly into cash. They ignore the ongoing cost of running the system. And they assume the benefit starts on day one at full strength.

Fix those three things and you get a number you can take to your board, your bank or your own conscience.

This article gives you the formula, two fully worked examples in rupees, and a method for stress-testing the result before you commit.

Why do standard ROI formulas fail for AI projects?

Conventional capital investment ROI works because the costs and benefits are both cash. You buy a machine for ₹20 lakh, it produces goods you sell, and the money arrives in the bank.

AI projects break this in four places.

The benefit is usually time, not money. If an AI agent saves 200 staff hours a month but you do not reduce headcount or increase output, you have saved nothing in cash terms. You have created capacity. Capacity is valuable, but it is only worth money if you use it.

The costs continue. A machine has a purchase price and maintenance. An AI system has a build cost, model usage costs that scale with volume, hosting, maintenance, and the cost of humans reviewing its output. The last one is invisible in most proposals.

The benefit ramps. Accuracy improves over the first few months. Staff trust builds slowly. Volume routed to the agent increases gradually. Assuming full benefit from month one typically overstates first-year ROI by 30% to 50%.

Some benefits resist measurement. Faster response times win deals you never knew you were losing. Fewer errors prevent complaints you never receive. These are real and you should note them, but do not put invented numbers against them.

Step 1: Establish your baseline cost

You cannot calculate a return without knowing what the process costs today.

Monthly baseline cost = (instances per month × minutes per instance × people involved ÷ 60) × loaded hourly cost

On loaded hourly cost. Do not use the salary figure alone. Include employer contributions, bonus provision, the share of rent and electricity that person occupies, equipment, and supervision time.

Loaded hourly cost = (annual CTC × 1.35) ÷ 2,080 hours

For an employee on ₹4.8 lakh CTC: (4,80,000 × 1.35) ÷ 2,080 = ₹311 per hour. Round to ₹350 for planning if you want a small buffer, or use your own multiplier if your overheads differ.

Also capture the non-labour baseline costs, which are often significant:

• Error and rework cost. Wrong dispatches, credit notes, re-doing work. Estimate the monthly rupee value.
• Overtime and weekend coverage attributable to this process.
• Opportunity cost of delay. If quotations take two days and competitors take two hours, estimate the deals lost. Be conservative here or leave it out entirely.
• Attrition cost. If staff leave because the work is grinding, replacement cost is real. Typically 1 to 3 months of salary per exit in recruitment and lost productivity.

Write all of these down as a single monthly baseline figure.

Step 2: Calculate total cost of ownership, not build cost

This is where most calculations go wrong. Build cost is roughly half the two-year picture.

Year 1 TCO = Build cost + (Monthly running cost × months live) + Internal review cost + Data preparation cost

Your running cost components:

Calculating internal review cost. If your agent handles 3,000 interactions a month and 10% require human review at 2 minutes each, that is 300 interactions × 2 minutes = 600 minutes, or 10 hours a month. At ₹350 per hour, ₹3,500 a month. Small, but include it, because a badly scoped agent needing 30% review changes the picture entirely.

Planning rule: annual running cost typically lands at 25% to 35% of build cost. If a vendor's running estimate is far below that, ask what they have excluded.

Step 3: Separate hard savings from capacity value

This single distinction will make your ROI calculation more credible than 90% of what you will be shown.

Capacity value is real, but only if you actually redeploy the capacity to something that generates revenue. If freed hours dissolve into a slightly less hurried day, the value is zero in cash terms.

The honest approach: calculate ROI twice. Once counting hard savings only, which is your conservative case. Once counting hard savings plus capacity value at full labour rate, which is your optimistic case. The truth sits between them, and showing both ranges is what makes your number believable.

Step 4: Apply a realistic ramp curve

Do not assume full benefit from month one. A defensible ramp for a typical MSME agent build:

If your build takes longer, shift the whole curve right. If you are integrating with fragile legacy systems, extend the ramp by a month or two.

Step 5: The formula

ROI (%) = (Total benefit − Total cost) ÷ Total cost × 100

Payback period = the month in which cumulative net cash flow first turns positive

Calculate over 24 months, not 12. AI projects almost always look mediocre in year one because the build cost lands entirely in that year, and strong in year two when only running costs remain. Judging on year one alone rejects good projects.

Worked example 1: A project that pays back

The business. An auto components distributor in Pune, 42 employees, running Tally and a basic inventory system, with most customer communication on WhatsApp.

The process. Order status enquiries, stock availability questions and invoice copy requests arriving on WhatsApp.

Baseline

• 180 messages per day, 26 working days = 4,680 per month
• Average handling time: 4 minutes, including looking up the order
• Total: 4,680 × 4 ÷ 60 = 312 hours per month
• Loaded cost: ₹350 per hour
• Monthly baseline = ₹1,09,200 (₹13.1 lakh per year)

Expected benefit

The agent is scoped to handle order status and invoice resend automatically, roughly 65% of volume. Pricing enquiries stay with humans because discount logic is contract-specific.

• Gross hours saved: 312 × 65% = 202.8 hours per month
• Human review: 10% of 3,042 automated conversations × 2 minutes = 10.1 hours per month
• Net hours saved = 192.7 hours per month
• Capacity value = 192.7 × ₹350 = ₹67,445, round to ₹67,000 per month

Of this, the hard savings component is the third customer-service hire they no longer need (₹30,000 per month loaded) plus eliminated weekend overtime (₹8,000 per month). Hard savings = ₹38,000 per month.

Costs

• Build cost: ₹3,50,000 (6-week narrow build, two integrations)
• Model usage: ₹12,000 per month
• Hosting: ₹4,000 per month
• Maintenance at 18% of build: ₹63,000 per year = ₹5,250 per month
• Total running cost = ₹21,250 per month, starting month 3

Year 1, optimistic case (capacity value counted)

Year 1 ROI = (₹5,89,600 − ₹5,62,500) ÷ ₹5,62,500 × 100 = 4.8% Payback = month 12

Year 2: benefit ₹8,04,000, cost ₹2,55,000, net ₹5,49,000. Year 2 ROI = 215% Two-year ROI = (₹13,93,600 − ₹8,17,500) ÷ ₹8,17,500 × 100 = 70%

Year 1, conservative case (hard savings only)

At ₹38,000 per month full run rate, year 1 benefit is ₹3,34,400 against costs of ₹5,62,500. Year 1 is negative. Payback arrives around month 26.

How to read this

The honest answer is a payback range of 12 to 26 months, depending on whether the freed capacity gets used for revenue-generating work. That range is the number to present. It is defensible, it is not a sales pitch, and it makes the decision genuinely strategic rather than arithmetic: the project's value depends on what the business does with the time it recovers.

Worked example 2: A project that does not pay back

Worth including because this scenario is far more common than vendors admit.

The business. A 24-person architectural services firm.

The process. Preparing client proposals. 40 per month, 25 minutes each, one person.

Baseline

• 40 × 25 ÷ 60 = 16.7 hours per month
• At ₹400 per hour (senior staff): ₹6,680 per month, or ₹80,160 per year

The proposal on the table

A ₹2,50,000 custom proposal-drafting agent, with running costs of ₹9,000 per month.

The arithmetic

Even assuming a generous 70% time saving, the benefit is ₹4,676 per month. Running cost is ₹9,000 per month.

The running cost alone exceeds the entire benefit. The project never pays back. It loses roughly ₹52,000 a year indefinitely, on top of never recovering the ₹2.5 lakh build.

What they should do instead

• Use a general-purpose AI assistant with good templates, at roughly ₹2,000 per month per user
• Build a structured proposal template library, which costs a week of internal effort
• Revisit a custom build if volume reaches 150+ proposals a month

The lesson. Volume is the dominant variable. A process must be large enough that a percentage saving on it is meaningful. Annoyance is not volume, and the two are easy to confuse because the irritating tasks are the memorable ones.

How do you stress-test your ROI calculation?

Before you present the number, run it against four adverse assumptions. If the project still works, your case is solid.

A project that only works under best-case assumptions is not a project. It is a hope.

What benefits should you record but not put a number on?

Note these in your business case as qualitative, without inventing figures. Fabricated soft benefits are the fastest way to lose credibility with a bank or a partner.

• Faster response times improving win rates
• Consistency of information given to customers
• Reduced dependence on one knowledgeable individual
• Better data capture as a by-product of automation
• Staff satisfaction from removing drudgery
• Capacity to handle seasonal peaks without temporary hiring
• Audit trail and traceability that did not previously exist

That last one is quietly valuable. An agent that logs every interaction gives you records that a WhatsApp-based process never produced.

How do you measure ROI after go-live?

Agree the measurement method before the build starts, or you will argue about it afterwards.

Track monthly:

• Volume of interactions handled by the agent
• Percentage handled without human intervention
• Accuracy, measured by sampling 30 to 50 outputs
• Human review time actually spent
• Actual model and hosting spend
• Baseline process hours, re-measured quarterly

Review at months 3, 6 and 12 against the projection. If month 3 actuals are far below plan, that is a scope or data problem worth fixing early, not a reason to abandon.

Re-measure the baseline honestly. Processes drift. If volumes grew 20%, your savings grew too, and you should claim them.

Three takeaways

1. Calculate ROI twice, conservatively and optimistically. Hard savings are cash that stops leaving your account. Capacity value is time you must actively redeploy to convert into money. Presenting both as a range is more credible than a single confident number, and it forces the real strategic question.

2. Total cost of ownership is roughly double the build cost over two years. Plan for 25% to 35% of build cost annually in model usage, hosting, maintenance and internal review. A proposal quoting only build cost is incomplete.

3. Judge over 24 months and stress-test the assumptions. Year one almost always looks weak because the build cost lands there. If the project only survives under best-case assumptions, reduce the scope rather than the scrutiny.

Your ROI calculation checklist

• Count actual instances per month, from system data rather than memory
• Time the process by observation, not by asking a manager
• Calculate loaded hourly cost as (CTC × 1.35) ÷ 2,080
• Separate hard savings from capacity value and list them apart
• Include model usage, hosting, maintenance and internal review in costs
• Add a data preparation line of 10% to 30% of build cost
• Apply a ramp curve of 0 / 40 / 60 / 80 / 100 across the first six months
• Model 24 months, not 12
• Stress-test: half the benefit, 1.5× the build, 2× the running cost, three-month delay
• List qualitative benefits separately, with no invented numbers
• Agree the post-launch measurement method before signing

Questions to ask yourself before you close this page

1. If you saved 200 staff hours a month, do you know specifically what those hours would be spent on instead, or would they simply be absorbed?

2. Have you ever calculated what your most repetitive process actually costs you in rupees per year, or has it always been described as "too much time"?

3. When a vendor quotes you a build price, do you ask for the twelve-month running cost in the same document?

4. What would have to be true for this project to fail, and how would you know by month three rather than month twelve?

5. If the conservative case shows payback at 26 months, is that a reason to decline, or a reason to reduce the scope until it is 14?