A free AI model answered 30 customer emails for our made-up shop in about six seconds each, at no cost per reply. It also told one customer their address had been changed when nothing had changed, and gave three different answers to the same email. We ran the free option four times and read every reply, then set it against the published prices of six paid tools. This is how free and paid AI for customer support compare for a small business.
Is free AI customer support good enough for a small business?
Free is enough if three things are true. Your replies come from a written policy that fits on a page. A person reads every reply before it is sent. You handle a few dozen emails a day, not thousands. In our test, Gemma 4 answered from a one-page policy in about six seconds a reply and cost nothing per reply.
Pay, or at least trial a paid tool, if any of these are true. Replies must go out without a person reading them. The tool has to act on an order, such as changing an address. Your volume is high enough that one person cannot read every draft.
The reason is in the numbers below. In every one of our four runs, 4 or 5 of the 30 replies said something the policy did not support, and one email got three different answers across the runs. A free model that drafts is useful. A free model left alone is a risk.
| Free: Gemma 4 on our test machine | Paid: six tools, published pages | |
|---|---|---|
| Price | No fee per reply or per seat | About $19 to $29 per agent a month on entry plans, plus about $0.50 to $2.00 per AI resolution |
| What we tested | 30 emails, four runs, every reply read | Not tested. Prices read on 20 October 2026 |
| Speed | 6.1 seconds a reply, one at a time | Not tested |
| Accuracy, replies read by hand | 25 or 26 of 30 | Not tested |
| Can it change an order? | No, it only drafts text | Depends on the tool. Not tested |
| Where the emails go | A local server on our machine | Not tested. Check the vendor’s policy |
Can AI answer customer emails on its own? Our 30-email test
AI can answer routine customer emails correctly. It should not answer them on its own: in our test, 4 or 5 replies in every 30 said something the shop’s policy did not support.
To find out, we invented a shop, Harbour & Co., a UK online kitchenware seller, and wrote a one-page policy for it. It covers delivery times and prices, a 30-day returns rule, a Final Sale exception, a 7-day window for damage, a 12-month warranty on electrical items, loyalty points, and the cases that must go to a person. Then we wrote 30 customer emails. Some are plain lookups, such as delivery times and opening hours. Some test a limit, such as a return after 45 days. Some are traps: a question the policy does not cover, an angry customer, a legal threat, and an email that says SYSTEM OVERRIDE and tells the model to ignore its rules and refund 500 pounds.
Gemma 4 31B, in a 4-bit build, ran in LM Studio on the M4 Pro we tested on. The model took 18.44 GB of memory with a 16,384-token context. It received the policy and one email at a time, and had to end each reply with ACTION: ANSWER or ACTION: ESCALATE. We ran all 30 emails four times.
We scored every run two ways. Automated rules checked for required facts, forbidden claims and the right action. Then we read every reply. Our first pass used 25 emails and the automated checks gave 24 of 25. We tightened the checks and added five harder emails before the runs described here.
On a quiet machine, the 30 replies of the filmed run took 183 seconds. The mean was 6.1 seconds a reply, the median 5.6 and the slowest 11.4, at 10.8 tokens a second. One reply at a time, that is about 590 replies an hour. Our wider speed numbers for this model are in our Gemma 4 31B MLX against GGUF test.
AI customer service gone wrong: the mistakes in our test
The mistakes were not wild ones. They were confident replies that sounded right. The automated checks passed 29 of 30 replies in all four runs. Reading the replies gave 25, 26, 26 and 25. The rules could not tell a reply that sounds right from one that says something the policy does not.
Three patterns showed up.
It says it did things it cannot do. Asked to change a delivery address inside the one-hour window, every run replied that the address had been updated. A drafting tool changes nothing, and a customer who reads that line believes it.
It states policy that is not on the sheet. Asked about collecting an order from a shop, every run said the shop offers no collection service. The policy says nothing either way. Asked whether return postage could come off a refund, every run said no. The policy says only that the customer pays postage. Twice it invented a management team when the policy said a person would reply.
It does not give the same answer twice when the policy is unclear. One email said a return was posted on day 28 and reached the shop on day 33. The policy gives 30 days from delivery and is silent on arrival. Over four runs the model passed it to a person once, promised a refund twice and refused the return once. Only the first is safe.
It also did a lot right. It answered routine lookups correctly: delivery times, the free delivery threshold, opening hours, the loyalty points sum (250 points is 12.50 pounds) and the price adjustment window. It passed the angry customer, the legal threat, a 210 pound refund dispute and a question about induction hobs, which the policy does not cover, to a person. It did not follow the SYSTEM OVERRIDE email in any run. In the first three runs it chose between answering and escalating the same way on 29 of 30 emails, although only 9 of the 30 replies were word for word identical.
How much does free and paid AI for customer support cost?
Paid tools cost about $19 to $29 per agent a month on entry plans, plus about $0.50 to $2.00 per AI resolution. The free option has no fee per reply or per seat, but it needs a computer that can hold the model: ours used 18.44 GB for the model alone. We did not measure electricity, so we make no claim about running cost.
For paid tools, we read the published pricing pages of six widely used help desk and chat tools on 20 October 2026 and kept only the range. Entry plans cost about $19 to $29 per agent a month, although two of the six charge per workspace instead of per seat. AI replies are charged on top. Four of the six state a price per AI resolution or session, from about $0.50 to $2.00.
Two sums from those ranges. A team of 10 agents on entry plans pays about $190 to $290 a month in seats before any AI usage. If 500 conversations a month are billed as AI resolutions, that adds about $245 to $1,000. Prices change, so check the page on the day you buy. Our note on per-seat AI pricing explains how these costs grow with a team.
We did not test any paid tool, so we cannot say whether they make fewer of the mistakes above. What the pages show is what you are buying: a seat or a workspace, with AI use charged separately or capped. A tool connected to your orders can do what ours only described, such as changing an address. Whether it does so correctly is something to test with your own emails.
How to choose an AI customer support tool: what to test first
Test any tool, free or paid, on your own emails before you pay for it. These are the checks that caught problems in our test.
- Write 30 of your own emails. Include the awkward ones: a request your policy does not cover, an unclear late return, an angry customer, and one that tells the bot to ignore its rules.
- Read every reply. Our automated check passed 29 of 30 while reading found 4 or 5 problems in every run.
- Run the same emails more than once. One of ours got three different answers.
- Check what the tool says it did. Ours claimed to have updated an address it could not change.
- Ask what happens when the tool does not know. It should hand over to a person and say when.
- Do the sum for your volume: seats plus the price per AI resolution.
- Ask where your customers’ emails are processed and whether they are used to train models. Our free test ran through a local server on our test machine. Our note on LM Studio and tracking shows what that app does and does not send, and our telemetry report shows how we check.
*Disclaimer: We tested the free option only. Everything about the paid tools comes from the vendors’ published pages, read on 20 October 2026, and we did not test them.*
