Most ROI calculations for AI automation are fiction. Someone multiplies "hours saved per week" by an average hourly wage, adds it up over a year, and produces a number that would make any CFO squint. Then the tool gets bought, the hours don't actually disappear, and six months later nobody wants to talk about the spreadsheet.

I've built the model I actually use before recommending an automation to a client. It's not clever. It just refuses to lie to you. If you work through it honestly before you sign anything, you'll kill about a third of the automations you were excited about, and the ones that survive will pay back faster than you expected.

Start with the work, not the tool

The first mistake is starting from "we should use AI for X." The correct starting point is a specific process that already exists, that someone is already doing, and that you can describe in steps. If you can't describe it in steps, you're not ready to automate it. You're ready to document it.

Pick one process. Write down who touches it, what triggers it, what systems it lives in, and what "done" looks like. For a dental front desk that might be insurance verification: patient books, someone pulls the insurance card image, calls or logs into the payer portal, confirms benefits, notes them in the practice management system, and flags anything unusual for the office manager. Five steps, three systems, one exception path.

Now you have something you can price.

The four numbers you actually need

Forget generic ROI templates. You need four numbers, and you need to be honest about all of them.

1. Fully loaded cost of the current process

Not just wages. Include benefits (roughly 1.25 to 1.4x salary for most US employers), the software licenses that person uses, the manager time spent supervising the work, and the cost of errors. If your biller currently spends 12 hours a week on claim follow-up and their fully loaded cost is $32/hour, that's $19,968 a year in labor alone. Add the two hours a week your office manager spends untangling denials that got missed, and you're at $23,000 before you count the revenue lost to claims that aged out.

Do this per process, not per person. A single employee usually touches five or six processes and you're only automating one of them.

2. Realistic automation coverage

This is where most models break. A vendor says "automates insurance verification." What they mean is it handles the 70% of cases where the payer portal returns clean data and the patient's info matches. The other 30%? Still a human problem. Sometimes a bigger one, because now the human only sees the weird cases and has lost context on the easy ones.

Before you buy, ask for the actual coverage rate on cases that look like yours. Then cut it by 15% for your first six months, because your data is messier than the demo. If a vendor won't give you a specific number, assume 50% and see if the math still works.

3. Total cost of ownership, not sticker price

The subscription fee is usually the smallest line. Add:

  • Implementation time from your team (typically 20 to 80 hours for a real automation, spread across ops, IT, and whoever owns the process)
  • Integration costs if the tool needs to talk to your PMS, EHR, CRM, or accounting system
  • Ongoing supervision: someone has to review outputs, especially in the first 90 days. Budget 3 to 5 hours a week.
  • Change management: retraining the team, updating SOPs, handling the person who liked doing it the old way
  • Compliance and security review, which for healthcare and financial data is not optional

A $500/month tool often has a real first-year cost between $15,000 and $25,000 once you count the human time to launch it and keep it honest.

4. What happens to the recovered time

This is the number everyone lies about. If you save your front desk 8 hours a week, you have not saved 8 hours of labor cost unless you actually reduce headcount or reallocate that time to revenue-generating work. In most small operations, that time gets absorbed. It's real relief for a stressed team, which has value, but it is not cash back.

Be specific about the recovery mechanism. There are basically three:

  • Headcount avoidance: you were about to hire and now you don't. This is real, bankable savings.
  • Revenue capture: the automation books appointments you were missing, recovers claims that were aging out, or follows up on leads that were falling through. This is the biggest lever and the easiest to measure if you have baseline numbers.
  • Reallocation: your team spends the recovered hours on higher-value work you can name specifically. If you can't name it, don't count it.

The model, in one paragraph

Annual value equals (current process cost x realistic coverage rate x recovery factor) plus (new revenue captured or losses prevented) minus (total cost of ownership). Recovery factor is 1.0 if you're avoiding a hire, 0.7 to 0.9 if you're reallocating to measurable revenue work, and 0.2 to 0.4 if you're just giving your team breathing room. That last case is still worth doing sometimes, just don't put it on the ROI slide.

A worked example

A 4-location medical practice wanted to automate appointment reminders and no-show recovery. Current state: front desk staff at each location spent about 6 hours a week on reminder calls and rescheduling no-shows. Fully loaded cost per hour was $28. Across 4 locations, that's 24 hours a week, or roughly $35,000 a year.

No-show rate was 11%. Each no-show represented an average $180 in lost revenue. At around 320 appointments a week per location, that's roughly $1.3M in annual no-show losses across the group.

The automation (SMS reminders with two-way rescheduling and a callback agent for no-shows) had a realistic coverage rate of about 75% on reminders and reduced no-shows to 6%. Total first-year cost of ownership, including integration with their scheduling system and 90 days of supervised rollout, was about $42,000.

Labor recovery: the time didn't eliminate anyone, but it let them cancel a planned hire at one location, worth about $52,000. No-show reduction: from 11% to 6% recovered roughly $590,000 in revenue that would have walked. Net first-year value, conservatively: about $600,000 against $42,000 in cost.

That's the shape of an automation worth buying. Boring process, clear baseline, measurable revenue mechanism. If your model doesn't look like that, keep asking questions before you spend money.

Red flags in your own math

  • You can't state the current process cost in dollars without guessing
  • The ROI depends on "productivity gains" you can't attach to a specific outcome
  • The vendor's case studies are all from companies 10x your size
  • Nobody on your team has time to supervise the rollout
  • The savings assume 100% coverage of edge cases

If you want a second set of eyes on a specific process before you commit, talk to our team at Qintara Corp. We'd rather tell you an automation isn't worth building than sell you one that quietly underperforms.

Frequently Asked Questions

How long should payback take on an AI automation?

For operational automations in small and mid-sized businesses, I'd want to see payback inside 6 to 9 months on a conservative model. Anything projecting 3-year payback is either enterprise-scale or wishful.

Should I count "soft" benefits like employee morale?

Count them, but in a separate column. If the hard-dollar case doesn't stand on its own, morale won't save the business case when your CFO asks why the number didn't move.

What if I don't have baseline data for the process today?

Spend two weeks measuring before you buy anything. Have the team log time, count exceptions, and note error rates. Automations built on guessed baselines get evaluated against guessed outcomes, which is how you end up renewing tools nobody can defend.

How do I account for risk, especially with patient or customer data?

Add a line for compliance review and expect ongoing governance work. For healthcare, confirm HIPAA specifics with your own counsel and your vendor's BAA before you sign. Security is a cost of ownership, not an afterthought.

What's the single most common mistake in these calculations?

Assuming recovered hours automatically convert to saved dollars. They don't, unless you have a specific plan for what happens to that time. Name the mechanism or don't count the money.