How to measure the ROI of an AI agent.

At some point in every AI agent conversation, someone asks: "What's the ROI on this?" And at that point, most vendors either hand-wave about "transformative efficiency" or produce a spreadsheet so optimistic it belongs in a pitch deck rather than a business case.

Neither helps you make a real decision. So here's how we actually model ROI before we build anything — and how you can apply the same framework whether you're evaluating a vendor, building in-house, or trying to justify the investment upward.

Start with the denominator: what does the task cost today?

The first number you need is the fully-loaded cost of the status quo. Not the salary — the fully-loaded cost, which includes benefits, overhead, management time, and the cost of errors.

Here's the calculation:

• Hours per week spent on the task across everyone who touches it
• Fully-loaded hourly rate (annual comp + 30–40% for benefits/overhead, divided by 2,000 working hours)
• Error rate and rework cost — how often does the task produce an error, and what does it cost to fix?
• Throughput ceiling — is there work you can't currently take on because this task is a bottleneck?

For a concrete example: a three-person team each spending 8 hours per week on document processing, at a fully-loaded rate of $60/hour, costs $1,440 per week, or $74,880 per year, just in labor. That's before you count the cost of errors or the backlog that builds during peak periods.

Then calculate what the agent actually saves

AI agents don't eliminate all labor on a task — they change the nature of it. A well-built agent typically handles 70–90% of cases autonomously and routes the remainder to humans for review. So the real question is: what percentage of the task goes autonomous, and what does the remaining human review cost?

Using the same example: if the agent handles 80% of documents autonomously and reduces the human review time on the remaining 20% to 30 seconds per document, you might go from 24 person-hours per week to 4. That's 20 hours saved per week — $1,200/week at $60/hour, or $62,400 per year.

Now you have a savings numerator. The next question is what it costs to get there.

The full cost of building and running an agent

This is where most ROI calculations go wrong — they count the build cost but not the run cost, or they count neither and just assert value.

A complete cost picture includes:

• Build cost: agency fee or internal development time. For a focused, well-scoped agent, this is typically $8,000–$25,000 depending on complexity and integrations.
• LLM inference costs: at current pricing, a well-optimized agent processing 200 documents per day typically costs $50–$200/month in API calls, depending on document length and the model used.
• Monitoring and maintenance: even a stable agent needs someone reviewing performance dashboards, catching edge cases, and deploying prompt updates. Budget 2–4 hours per month.
• Integration overhead: if the agent connects to multiple systems (email, CRM, document management, etc.), each integration adds ongoing maintenance surface area.

For our example agent, assume a $15,000 build cost and $150/month in ongoing costs ($1,800/year). Total first-year cost: $16,800. Annual savings: $62,400. Net first-year return: $45,600. Payback period: about 11 weeks.

Three ROI factors most people miss

The direct labor math is the floor, not the ceiling. There are three additional value drivers that rarely appear in the initial business case but consistently show up in retrospective analysis.

Error-reduction value. Repetitive human tasks have error rates. For document classification, the typical human error rate on a high-volume, fatiguing task is 2–5%. On 1,000 documents per week, that's 20–50 errors, each of which requires a human to catch and fix. If each error costs 15 minutes to resolve, you're spending 5–12.5 person-hours per week on rework. A well-built agent runs at less than 1% error on high-confidence cases. The rework savings compound.

Throughput expansion. The most common thing we hear after an agent is deployed is: "We can now take on clients we were turning away." When a bottleneck task is automated, the team's capacity to handle volume increases without headcount. If the task was gating revenue, removing the constraint has value that's separate from labor savings — and often larger.

Consistency and compliance value. Humans apply different standards depending on time of day, workload, and familiarity with a document type. An agent applies the same ruleset every time, to every document. For regulated industries, this consistency has audit value that can reduce legal and compliance risk — a benefit that's real but hard to put in a spreadsheet until you need it.

How to run the calculation for your situation

Here's a simple template you can apply to any task you're evaluating:

• Step 1: Count the hours. Add up every person-hour per week that touches the task, including handoffs and review steps.
• Step 2: Price the hours. Multiply by fully-loaded hourly cost.
• Step 3: Estimate agent automation rate. For routine, high-volume tasks: 75–85% is realistic. For complex, exception-heavy tasks: 50–65%.
• Step 4: Calculate remaining human time after agent deployment. Apply the automation rate, add back review time for the unautomated portion.
• Step 5: Get a build quote and estimate ongoing costs. A good shop should be able to scope this in a 30-minute conversation.
• Step 6: Calculate payback. Net annual savings ÷ build cost = payback in years. Multiply by 12 for payback in months.

If payback is under 12 months, the economics are usually clear. If it's 12–24 months, you need to factor in the non-labor benefits to make the case. Over 24 months, either the scope is too large, the task volume is too low, or the wrong approach is being used.

The number that actually matters

In most of the deployments we've run, the first calculation operators make underestimates the ROI — because they're thinking about direct labor savings and ignoring throughput expansion. The accounting firm we wrote about in our first-week post estimated 55 hours of weekly savings. Six weeks after deployment, they reported taking on 40% more client volume with the same staff. That capacity expansion was worth more than the labor savings.

Build the base case with the conservative numbers. Then build a second scenario that accounts for the throughput you can unlock. The difference between the two is usually where the business case gets interesting.

If you want to run this calculation for a specific task in your business, book a 30-minute call. We'll help you scope the task, estimate the automation rate realistically, and give you a build cost estimate — so you have real numbers before you make a decision.