By Miki Furman
Every AI support tool now arrives with a number. Resolves half your conversations. Cuts support costs by 40 percent. Pays for itself in a quarter. The number is usually true somewhere. The question is whether it is true for you, and the vendor’s deck cannot answer that because it does not know your inputs.
I run a support operation for a living, and I have sat through enough of these demos to have a routine. It has three steps. Print the formula. Name every input. Say which inputs are measurements and which are guesses. It takes twenty minutes, and it has never once produced the number on the slide.
Step one: print the formula
A cost claim is arithmetic, so write the arithmetic down. For a support tool, it is this:
Annual savings = contacts per year
x share the tool closes on its own
x your cost per human contact
– platform fees
– usage fees
– implementation time
– human review time
Nothing exotic. But once it is on paper, every word in the vendor’s pitch has to land in a specific slot, and most of the pitch does not fit anywhere.
Step two: name every input, and who owns it
Contacts per year is yours. Pull it from your help desk. If you do not know it, stop here and find out, because nothing downstream means anything without it.
Cost per human contact is yours too. Loaded hourly cost of an agent divided by contacts handled per hour. A team paying $26 an hour, loaded and handling eight contacts an hour, is at $3.25 a contact. That is my assumption for the example below, not a benchmark. Your payroll knows your real figure.
The share the tool closes on its own is the input the vendor owns, and it is where the claim lives or dies. Three questions turn a headline into a number you can use:
What is the denominator? “Resolves 50 percent of conversations” can mean half of everything, or half of the conversations the tool chose to take. Intercom reported that its Fin agent achieved a 50 percent resolution rate while handling just over 20 percent of all interactions across its customer base (Intercom, 2025). Both figures are honest. Only one of them belongs in your formula.
What counts as closed? A conversation the tool ended is not the same as a problem it solved. Ask for the re-contact rate: how many people came back through any channel within seven days. Contained is not resolved, and the gap between them is paid for by your human team.
Which customers is that number from? This is the median-versus-mean problem, and it is worth learning to spot because it is everywhere. A vendor’s headline rate is usually the average across its deployments, or its best ones. Averages are dragged upward by a few outstanding accounts. Ask for the median across customers in your volume band, and ask how many customers in that band cancelled in the past year. Gartner expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027 (Gartner, June 2025). A vendor who will not tell you its own cancellation rate is answering the question anyway.
For scale, service teams themselves say AI handles about 30 percent of cases today (Salesforce, State of Service, 2025). If a pitch is far above that, the burden is on the pitch.
Step three: run the example, then run yours
Take a team handling 10,000 contacts a month, 120,000 a year, at $3.25 a contact. The vendor claims a 50 percent resolution rate. Its own arithmetic:
120,000 x 0.50 x $3.25 = $195,000 a year saved.
Now, the same team runs a two-week pilot and measures using the three questions above. Say the tool closes 31 percent of all contacts with no human touch and no re-contact within seven days. That is a made-up pilot result for the example, and a decent one. The team’s arithmetic:
Gross saving: 120,000 x 0.31 x $3.25 = $120,900.
Then the subtractions the slide left out. Platform fee: $2,500 a month, $30,000 a year. Per-resolution fee of $0.99 on 37,200 resolutions, $36,828. Implementation: 80 hours of an engineer at $60, $4,800. Weekly review of transcripts and failures, three hours at $50 for a year, $7,800.
$120,900 – $30,000 – $36,828 – $4,800 – $7,800 = $41,472.
Forty-one thousand dollars a year. Still worth doing for many teams. Also, about a fifth of the number on the slide. Nothing about the tool changed between the two calculations. One input moved from the vendor’s average to your measurement, and four inputs appeared that were never on the slide at all.
Every figure in that example is labeled: the volume and the cost per contact are yours, the 50 percent is the vendor’s, the 31 percent is what a pilot would measure, and the fees are whatever your quote says. Swap in your own, and the arithmetic holds.
What the good vendors do
The good ones do not flinch at any of this. They hand over the denominator, the re-contact rate, and the median without being asked twice, and they will structure a pilot around your contacts rather than a demo around theirs. The tool is often genuinely useful. Handling routine cases with AI assistance freed agents an estimated four hours a week in one large survey (Salesforce, 2025), and a McKinsey study at a 5,000-agent company found generative AI lifted resolutions per hour by 14 percent (McKinsey, 2023). Those are real gains. They are just smaller, slower, and more conditional than a headline, and they show up only after the formula is printed.
If you want sourced figures to pressure-test a pitch with, I keep a page of AI customer service statistics at https://callforce.global/tools/ai-customer-service-statistics-2026/ . Every number on it names its original source and year, and anything I could not trace to a real source was left off. Use it to check the vendor’s numbers, then measure your own.
Byline: Miki Furman is Founder and CEO of Call Force Global.
