What is a reasonable ROI to expect from automation?

For SMBs, a year 1 ROI between 2x and 4x is reasonable. Above 5x is usually overestimated (re-check the numbers). Below 1.5x, reduce scope. Typical payback is 3-6 months.

How do I measure opportunity cost if I cannot see it?

Three ways: (1) A/B test — measure 2 weeks with slow reply vs 2 weeks with fast reply; (2) industry benchmarks (there are reports on conversion-by-response-speed); (3) conservative probability estimate. When in doubt, use the lower number.

Does the framework work for non-operational processes (marketing, sales)?

Yes, with one tweak: in marketing/sales the upside is not saved time but earned conversion. Replace 'hours saved' with 'extra expected sales with probability'. The ROI vs payback logic is identical.

What if I am just starting and have no historical data?

Measure 2 weeks in the current state before calculating. That is the minimum. If you do not have volume yet (early-stage startup), evaluate a low-touch automation (Make, Zapier) and revisit ROI at 90 days.

How often should I re-evaluate ROI of a live automation?

Month 1, month 3, month 6, then every 6 months. In the first 6 months savings are usually underestimated (you are still finding inefficiencies) and operating cost sometimes goes up (LLMs, scaling). Then it stabilizes.

How to Measure Automation ROI: A Concrete SMB Framework

TL;DR: Real automation ROI = (monthly measurable savings × 12) ÷ (initial investment + annual operating cost). If the ratio is above 2x in 12 months, it is worth it. Below 1.5x, scope it down or wait.

A marketing agency in Montevideo asked us 4 months ago whether it was worth automating their sales proposal generation. They were doing 30 proposals a month, each taking 90 minutes. They wanted a system using an LLM to generate drafts from a 10-question brief.

We ran the ROI before writing a line of code. Turned out the savings were real but smaller than expected (filling out the brief still took 30 minutes, not 0). We scoped down, simplified, and shipped something that paid back in 5 months instead of the 2 months they originally projected.

This post is the framework we use. It works for SMBs to decide if an automation is worth it before spending money on it, and to measure it afterwards.

Why almost nobody measures ROI right

Three typical mistakes:

Overestimating time savings. "This would save Maria 4 hours a day" — Maria almost always still spends 1-2 hours on something (reviewing exceptions, fixing what the bot does not do well). Real savings are 60-70% of the projection.
Ignoring hidden costs. Automation is not free after kickoff: APIs cost money, LLMs cost money, servers cost money, maintenance costs money, training costs time.
Confusing "automate" with "lay off staff". In 80% of SMBs nobody is fired — people get reassigned to higher-value work. That counts as opportunity savings, not payroll savings. Accounting must reflect that.

The 5-step framework

Step 1 — Quantify the current problem cost

Measurements:

Monthly person-hours spent on the process × loaded hourly cost (not just salary: salary + benefits + space + supplies ≈ 1.4-1.7 × net salary).
Errors that end in direct costs: credit notes, returns, client churn, fines. Conservative estimate.
Opportunity cost: unanswered leads, delayed collections, missed sales. Calculate with probability: "if reply is in 1 hour vs 1 day, conversion goes from X% to Y%".

Sum it all: current monthly cost = C_current.

Step 2 — Estimate the post-automation cost

Not zero. Includes:

Person-hours that still exist (review, exceptions, oversight). Usually 20-40% of original.
Automation operating cost (N8N, LLM, WhatsApp BSP, hosting). USD 50-300/month depending on complexity.
Maintenance cost (internal or external). For an SMB with no technical staff: USD 100-400/month from a software factory keeping it running.

Total: post-automation monthly cost = C_post.

Step 3 — Calculate real initial investment

Not just the build fee:

Project cost (internal or external). USD 2,500-15,000 depending on scope.
Team time during implementation (meetings, definitions, QA). Usually 30-60 hours.
Operational team training. 5-15 hours.
Opportunity cost: month 1 has productivity drag from the change.

Total: initial investment = I.

Step 4 — Calculate 12-month ROI

Formula:

Monthly savings = C_current - C_post
Annual savings = Monthly savings × 12
Year 1 total cost = I + (C_post × 12)
Annual ROI = Annual savings / Year 1 total cost

And payback period:

Payback (months) = I / Monthly savings

Step 5 — Apply the reality checks

After the math, multiply by:

0.7 on estimated time savings (you always overestimate).
1.3 on estimated operating cost (you always underestimate).
1.2 on initial investment (projects always have extras).

The resulting number is realistic. If after that the ROI exceeds 2x and payback is under 8 months, it is worth it. Between 1.5x and 2x: depends on strategic context. Below 1.5x: reduce scope or skip.

A real numerical example

The Montevideo agency case, sales proposal generation:

C_current: 30 proposals × 90 min × USD 12/h loaded = USD 540/month. Plus 2 proposals/month that get delayed and lose the deal (client signs with a competitor): USD 1,000 × 2 = USD 2,000/month. Total: USD 2,540/month.
C_post: 30 proposals × 30 min (brief + review) × USD 12/h = USD 180/month. Operating: USD 30/month (OpenAI). External maintenance: USD 100/month. Total: USD 310/month.
Monthly savings: USD 2,230/month (no reality check) → with reality check (0.7): USD 1,561/month.
Investment: USD 4,000 build + USD 900 internal time + 1.2 factor = USD 5,880.
Payback: 5,880 / 1,561 = 3.8 months.
Year 1 ROI: (1,561 × 12) / (5,880 + 310 × 12) = 18,732 / 9,600 = 1.95x.

Decision: go. ROI close to 2x, payback under 8 months.

Template to use

Simple Google Sheet template:

A — Manual hours/month × cost/hour = USD __
B — Errors and direct monthly cost = USD __
C — Monthly opportunity cost = USD __
CURRENT COST (A+B+C) = USD __

D — Hours/month that remain × cost/hour = USD __
E — Monthly operating cost (LLM + infra + maint) = USD __
POST-AUTOMATION COST (D+E) = USD __

MONTHLY SAVINGS = CURRENT COST - POST COST
REALISTIC SAVINGS = MONTHLY SAVINGS × 0.7

INITIAL INVESTMENT × 1.2 = USD __

PAYBACK = INVESTMENT / REALISTIC SAVINGS
Year 1 ROI = (REALISTIC SAVINGS × 12) / (INVESTMENT + POST COST × 12)

If ROI > 2x and Payback < 8 months → go. Between 1.5x and 2x → evaluate. < 1.5x → reduce scope or wait.

How to apply it in your company

Identify the process you want to evaluate. Just one, well-defined. "Automate everything" cannot be measured.
Measure one real week. Who does what, how long, what errors occur. Do not guess — log it.
Get 2 quotes (internal and/or external) with a specific scope. Take the median.
Apply the framework with reality checks. Be conservative on savings, generous on costs.
If you proceed, measure monthly. Compare real vs estimated. Adjust expectations. If real savings are 60% of original estimate, it can still be worth it.

Mistakes we see often

Forgetting LLM operating cost. An automation with GPT-4 can cost USD 200-500/month in tokens at high volume. Use GPT-4o-mini or Claude Haiku as default and only step up where needed.
Not including maintenance. An automation needs 2-5 hours/month of maintenance minimum. If you do not have someone internal, hire support.
Assuming the team absorbs change for free. There is a learning curve, resistance, reassignment. That is time and it counts.
Comparing against the "dream version" instead of the real baseline. If you do not measure today, you are comparing against intuition. Start by measuring the current state.

Related services: AI Automation · Tech Consulting · Tech Audit

Available in: Buenos Aires · Montevideo · Asunción · Santiago

When NOT to automate (even if ROI looks good)

If the process is going to change in the next 6 months. Automating something that will mutate is a waste.
If you depend on one person "who knows how to do it". First document, then automate. Otherwise you automate that person's knowledge and stay even more dependent.
If the team is very change-resistant. An automation poorly accompanied by change management fails even if the code is perfect.

Want us to help you evaluate?

At Deepyze we run this evaluation free in the first call: we look at your process, tell you whether to automate it, in what order, what stack fits, and what payback to expect. Then you decide whether to build it with us or your own team. Tell us what you have in mind.

More on how we approach projects: AI Automation and tech consulting.