Working Out the ROI of AI Automation (With Real Numbers)

Every AI vendor will show you a return calculation. Almost all of them start from a benefit and work backwards to a justification. This is the version you do yourself, on paper, with numbers you already have, before anyone sends you a proposal.

An AI automation ROI calculation has four steps: size the leak, estimate what gets recovered, total the real cost, and divide. It takes about twenty minutes and it will tell you whether to keep reading proposals or stop.

Step one: size the leak in your own figures#

Automation only pays when it recovers something you are currently losing. So the first number is the loss, and for most service businesses the loss lives in enquiries that go unanswered or get answered too late.

Take a plumbing business as the worked example. Four numbers, all of which you can find in an afternoon.

• Inbound calls per week: 40, from the mobile log or the phone system report.
• Share currently missed or unreturned: about 30 percent, so 12 calls a week.
• Share of those missed calls that were real jobs rather than sales calls or wrong numbers: 20 percent, so 2.4 jobs a week.
• Average job value: $480.

That comes to roughly $1,150 a week, or about $4,600 a month walking out the door. Getting an honest missed-call count is the step most people skip, and what missed calls are really costing you goes through how to pull it from a mobile or a PABX rather than guessing.

Two warnings on this number. If you take it from memory it will be wrong, usually low, because the calls you never answered are the ones you do not remember. And if you count every missed call as a lost job you will inflate it wildly, which is why the third line above exists.

Step two: recover less than you think#

No system captures the whole leak. Some callers will hang up on an AI voice. Some enquiries were never going to convert. Some jobs you would not have wanted.

I use 60 percent as a realistic recovery rate for enquiry capture, and I would rather you use 50. On the plumbing example at 60 percent, the recovered revenue is about $2,760 a month.

Revenue is not profit, so take it one step further. If the business runs at a 35 percent gross margin on labour and materials, the recovered gross profit is roughly $966 a month. That is the honest number to compare against a subscription, and it is the number vendors almost never use.

Step three: total the real cost#

Three costs, and most quotes only show you one.

The subscription is the obvious one. For a configured system answering calls and chats, expect $400 to $1,200 a month in New Zealand, with voice minutes pushing the upper end. The bands and what drives them are broken down in what an AI chatbot costs in New Zealand.

The setup fee is the second, usually $1,000 to $5,000 depending on how much of your process has to be documented before anything can be built.

The third cost is yours, and it is the one that ambushes people. Somebody has to review conversations, correct wrong answers and keep prices current. Budget two hours a week for the first month and half an hour a week after that. At an owner's effective rate of $80 an hour, that is roughly $170 in month one and $130 a month ongoing.

For the plumbing example, call it $700 subscription plus $130 of oversight, so $830 a month against $966 of recovered gross profit, on a $2,500 setup fee.

Step four: do the division#

Monthly gain is $966 minus $830, which is $136. Divide the $2,500 setup by $136 and the payback is about eighteen months.

That is a bad deal, and the arithmetic just told you so before you signed anything.

This is the part where most articles would quietly change the assumptions. Instead, notice what is actually wrong. The margin haircut and the oversight cost are fixed. The lever that matters is the price band: this business does not need a $700 configured system, it needs the $150 tier, because 40 calls a week is not complex volume. Rerun it at $150 plus $130 oversight and a $600 setup, and the monthly gain is $686 with a payback under a month.

Same business, same leak, completely different answer depending on what you buy. The build-or-buy decision is doing more work here than the AI itself, which is the subject of choosing between off-the-shelf and a custom build.

A case where it genuinely does not pay off#

Take a boutique architecture practice. Six enquiries a week, average project value high, decisions made over months, and every enquiry is a conversation about a unique site.

Size the leak: they answer nearly every enquiry already, because six a week is manageable. Say one a month slips. Recovery at 60 percent is 0.6 enquiries a month, and their close rate on enquiries is one in five, so 0.12 projects a month.

Even at a large project value the recovered gross profit does not clear a $600 monthly system reliably, and worse, the thing being automated is the first conversation with a client who is choosing an architect partly on that conversation. The cost of a mediocre first impression here is not on the spreadsheet at all.

The correct answer for this practice is not an AI receptionist. It is a faster quoting process and better follow-up on the enquiries they already have. Automation applied to the wrong bottleneck buys you nothing, which is worth checking with the scoring test in how to choose your first AI project.

What most AI automation ROI calculations leave out#

On the cost side, the ones that get missed are integration work when your job management software has no proper connection, and the fee to change your mind and reconfigure in month four. GST on overseas SaaS catches people out too.

On the benefit side, two things get left out and they matter. Speed of response affects close rate independently of coverage, because the business that replies first usually wins the job. Your own evenings also have a price, even if you have stopped charging for them.

I would not put a dollar figure on either of those in the calculation. Work out the payback on the hard numbers only, then let the soft ones break a tie.

What to do with the answer#

Under six months of payback, it is worth doing. Six to nine, do it if the underlying volume is stable and growing. Over nine months, the assumptions are carrying too much weight and you should either buy cheaper or fix a different problem.

Before you commit, rerun the whole thing at half your recovery estimate. If it still works at 30 percent recovery, you have a decision you can defend. If it only works at 60, you have a hope.

Then take a baseline. Record your enquiry count, your answer rate within an hour, and your close rate for two weeks before anything is switched on, because otherwise you will never be able to prove the calculation was right. The metrics worth tracking covers what to capture, and the wider view of what AI does for a small business covers which problems are worth putting through this arithmetic in the first place.