Most AI support ROI numbers you see are built on someone else's averages, and they fall apart the moment finance asks where the inputs came from. A defensible ROI uses five numbers you measure in your own support operation, subtracts the cost of wrong answers, and says plainly whether the saving is cash or capacity. This post gives you that method and one worked illustration you can replace with your own figures.

The five inputs you need before any spreadsheet

You can calculate a real ROI with five inputs. Each one has to come from your own data or a price sheet, not from a vendor deck.

  • Monthly conversation volume on the channels you plan to automate. Count conversations, not messages. If you run chat and WhatsApp, count both; if half your volume is email and the AI will not touch email, leave email out.
  • The share of those conversations the AI answers correctly without a human. This is the input everyone guesses, and it is the one you should measure in a pilot (more on that below).
  • Your cost per human-handled conversation. Take the fully loaded monthly cost of the people who answer these conversations (salary, benefits, tools, management time you can attribute) and divide by the conversations they handle in a month.
  • The platform price. For LayBuild this is a flat monthly plan. For per-resolution vendors it is a unit price multiplied by what they count as a resolution, which is a separate audit (see the deflection vs resolution post).
  • LLM token cost. This depends on the model and provider in use, how long your answers are, and how much retrieved context goes into each prompt. If you bring your own provider key, read it off the provider's usage dashboard during the pilot rather than estimating from a per-token price table.

There is a sixth input people leave out: the cost of a wrong answer. We will add it explicitly, because leaving it out is how ROI models end up too optimistic.

Why "share answered correctly" has to be measured, not assumed

Vendors report deflection or resolution rates, and those numbers measure different things (a customer who gives up counts as deflected). What you need for ROI is narrower: the share of conversations where the AI's answer was correct and the customer did not need a human afterwards.

The practical way to get it is a labelled sample. Pull a random set of AI conversations from the pilot, say a few hundred, and have someone who knows the product mark each one as correct, wrong, refused, or handed off. Two things make this easier in LayBuild. When retrieval finds nothing, or the answer shares too little vocabulary with the retrieved sources, the agent sends a fixed reply ("I do not have specific information about that in the knowledge base...") instead of guessing, so refusals are easy to count. And the dashboard shows the CSAT average and the share of ratings at 4 or 5, which is a useful cross-check but not a substitute for labelling, since most customers never rate.

Keep the "wrong" count. You need it for the next step.

The formula

Monthly net value is the human cost you avoid, minus everything you add.

text
avoided_cost   = volume x share_correct x cost_per_human_conversation
added_cost     = platform_price + llm_cost + review_time_cost + (volume x share_wrong x cost_per_wrong_answer)
net_monthly    = avoided_cost - added_cost

Review time is the hours someone spends each month reading transcripts, fixing knowledge-base gaps, and checking what the agent learned on its own. It is not optional. Cost per wrong answer is the follow-up conversation plus whatever the wrong answer caused (a refund, a repeat contact, a lost renewal). You will have to estimate that one, so state the estimate and test the result against it.

A worked illustration with LayBuild's plan prices

This is an illustration. Every number except the plan price is a made-up input, chosen to show the arithmetic. Replace them with yours.

Suppose you handle 3,000 chat and WhatsApp conversations a month. That is above the Starter plan's 1,500 conversations, so you would be on Pro at ₹3,999 a month (₹3,199 a month if billed annually). Suppose your pilot sample showed the agent answering 30% of conversations correctly without a human and 3% wrongly. Suppose your cost per human-handled conversation works out to ₹120, you pay ₹1,500 a month in LLM tokens, someone spends 8 hours a month on review at ₹500 an hour, and you estimate a wrong answer costs ₹300.

LineCalculationMonthly (₹)
Avoided human cost3,000 x 0.30 x 1201,08,000
Platform (Pro, monthly billing)flat3,999
LLM tokensfrom provider dashboard1,500
Review time8 x 5004,000
Wrong answers3,000 x 0.03 x 30027,000
Net1,08,000 minus 36,49971,501

Two things stand out. The platform price is the smallest cost line after tokens. And the wrong-answer line is larger than the platform, LLM and review lines combined. If your wrong-answer rate or its cost is higher than you think, that line moves the result far more than switching vendors would.

The break-even check finance will ask for

Fixed monthly costs in the illustration (platform, tokens, review) add up to ₹9,499. At ₹120 per avoided conversation, you need 9,499 / 120 = 79.2, so about 80 conversations a month answered correctly just to cover them. That is under 3% of 3,000. Break-even on the platform itself is rarely the problem.

The harder break-even is on quality. Each wrong answer in the illustration costs ₹300 and each correct one saves ₹120, so every wrong answer cancels out two and a half correct ones. Run the sensitivity: net before wrong answers is 1,08,000 minus 9,499, which is 98,501. Divide that by 3,000 x 300 and you get 0.109, so at 30% correct the result stays positive until about 10.9% of conversations get a wrong answer. Put that threshold in front of whoever signs off. It turns a vague worry about hallucinations into a number your pilot either clears or does not.

Cash savings or capacity: say which one you mean

The avoided cost in the table is real money only if your staffing changes. If the same people stay on the same hours and simply handle fewer conversations, you have bought capacity: slower hiring as volume grows, shorter queues, time for harder tickets. That is worth something, but it is not ₹71,501 of cash a month, and presenting it as cash is the fastest way to lose a finance lead's trust.

Write the business case in whichever form is true for you. "We avoid hiring one additional agent next quarter" is a defensible claim. "We save ₹8.6 lakh a year" is only defensible if a budget line goes down by that amount.

What this method leaves out on purpose

We did not include industry benchmarks for resolution rates or cost per ticket, because they describe other companies' products, customers and documentation. Your share correctly answered depends mostly on how well your knowledge base covers the questions you actually get, which is why a pilot on your own traffic is the only input worth using.

We also did not model the effect on CSAT or churn in money terms. You can, but those estimates stack assumptions quickly. Track CSAT in the pilot and report it next to the ROI rather than folding it in.

Finally, LayBuild auto-publishes answers it judges learnable into the knowledge base without human approval, and retracts them if the customer rates the conversation 2 stars or lower. That can raise the correct-answer share over time, and it can also spread a wrong answer. Budget review time for it, which is why the illustration includes a review line.

Next steps