The value of an AI chatbot can be estimated correctly from a single source: your company's data. The generic percentages in online articles (“cuts costs by X%”) say nothing about your case, because volume, question types and the cost of an hour differ from one company to another. The method below uses only numbers you can measure yourself, in two weeks, before you invest.
What to measure before any decision
For two weeks, the team logs every request arriving on the channels the assistant would take over (phone, email, WhatsApp, form). For each one: the topic, the channel, how long it took to resolve and whether it ended in an action (booking, quote, sale) or just information. This log gives you the four numbers you need:
- N — the number of requests per month (extrapolated from the two weeks, with an eye on seasonality).
- T — the average time to resolve a request, in minutes, per topic.
- P — the share of requests the assistant could resolve completely, without a human. Estimate it topic by topic, not overall: “hours and address” — almost all; “complaints” — almost none.
- C — the cost of one hour of work for the person who answers today, including taxes.
The formula for time recovered
Monthly time recovered is the sum, across topics, of N × P × T, expressed in hours. From it, subtract the time the team will still spend: reviewing handed-over conversations, correcting sources and handling exceptions. Multiply what remains by C and you get the monthly value of the time recovered.
Two warnings. First: recovered time becomes money only if those hours are used for something else useful or avoid a new hire. Second: calculate simple and difficult topics separately — an average that mixes “how much does it cost” with “why doesn't the product work” hides the fact that the first topic can be automated and the second can't.
Value that isn't time
Often the main value doesn't come from hours saved, but from requests that are no longer lost:
- Out-of-hours requests that today get an answer the next day or not at all. The log shows how many there are and how many became customers when they got a timely answer.
- Incomplete requests that generate a second and third exchange of messages. An assistant that collects the necessary details up front shortens the path to a quote.
- Response time to quote requests, where the customer asks several suppliers in parallel.
These effects are measured after launch, by comparing the same indicators over similar periods. Before launch you can only estimate them as assumptions — and they should be written down as such.
The full cost, not just the implementation
Include every component, over 12 months:
- Implementation: question analysis, source preparation, the build, integrations, testing.
- The AI model: paid by the volume of text processed; it grows with the number of conversations and the size of the sources consulted for each answer.
- The channel: there is no channel cost on the website; on the WhatsApp Business Platform, Meta charges per delivered template message, while replies within the 24-hour window opened by the customer are free.
- Operation: hosting, monitoring, updating sources, periodic review of conversations.
- The team's time: someone is responsible for the assistant — reading handed-over conversations and correcting sources.
What a pilot that actually tells you something looks like
A good pilot has clear limits and an acceptance threshold set before it starts:
- A single channel and 3–5 topics from the top of the frequency list.
- A test question set with the expected answers, run before launch and on every change.
- The indicators tracked: the share of conversations resolved without handover, the share of handovers with full context, requests taken outside business hours, complaints related to the assistant.
- The acceptance threshold: for example, “at least X of the chosen topics are answered correctly on the test set, and the team confirms the handovers are useful”. You set the value of X, based on how risky the topic is.
If the pilot passes the threshold, you expand topics and channels. If it doesn't, you've learned cheaply what's missing — usually better sources, not a bigger model.
Common calculation mistakes
- Counting all requests as “automatable” — use the share P per topic.
- Ignoring the supervision time after launch.
- Comparing against a “no assistant” scenario in which lost requests cost nothing.
- Using percentages from studies done in other industries, at other volumes and in other countries.
We apply the same method in every chatbots & virtual assistants project; if you'd like to see how it fills in with your data, you can start with a free audit.