How to Run a Fair One-Month Trial of AI Support on the Free Tier

Most AI support trials fail before the AI writes a single reply, because the trial is rigged to pass. You feed the demo three softball questions you already know the answers to, the AI nails them, and you come away with a warm feeling and zero real information. A month later you are paying for a tool you never tested, or you walked away from one that would have worked. A fair trial does the opposite. It points the AI at your real inbox, your real documentation, and your real last month of questions, and tells you something you did not already know. Here is how to run that trial on a free tier, in about thirty days, and three numbers to watch so the result is honest.

What a fair trial is actually testing

The question you want answered is not whether the AI is impressive. Any modern support AI can write a fluent, confident paragraph, and that tells you nothing. The real question is narrower: if this had been running on my inbox last month, how much of my support work would it have done correctly, and how much would it have handed back to me. A trial that answers that earns its month; one that only confirms the AI can form sentences wasted it.

So the thing you are measuring is not quality in the abstract. It is fit: between the AI and your specific documentation, your customers, and the mix of questions you actually get. An AI that is excellent on a software product with tight docs can be mediocre on a service business full of one-off situations, and the only way to find out which case you are in is to run it on your own material. The trial is a fit test, not a beauty contest.

Set the trial up so it cannot cheat

A trial lies in three predictable ways, all easy to avoid once you name them. The first is testing on invented questions. If you type the questions you think customers ask, you will type the ones you already answer well. Real customers are messier. They bury the question in a paragraph of context, or ask two things at once. Point the trial at genuine incoming mail, not a script you wrote.

The second way a trial lies is by hiding the replies from you. If the AI sends on its own from day one, you only find out it was wrong when a customer writes back annoyed, and by then the evidence is tangled up with a real support failure. Run the whole trial in draft-and-approve mode: the AI writes, the reply parks as a draft, and you read every one before anything goes out. You cannot measure what you never saw.

The third way is feeding the AI only the clean parts of your documentation and leaving out the vague policy page you have been meaning to rewrite. That makes the AI look better than it will be in production. Load the real docs, including the messy ones. A thin doc that produces a weak answer during the trial is not the trial failing. It is the trial doing its job, showing you a document to fix.

What the free tier actually gives you, and what it does not

A free tier exists so you can run exactly this test without a card on file. On Trigli, free means 50 emails, 25 chats, and 10 tickets a month, with room for 3 documents. It is deliberately small, and for a trial mostly enough. Fifty emails is a real cross-section of a small inbox: enough variety to see the common questions several times each, plus a couple of the weird ones that only show up occasionally. You are not handling a month of production volume; you are gathering enough examples to decide, and fifty will do that for most small operations.

Be honest about the limits too, because they shape what the trial can tell you. Three documents is tight. If your support knowledge lives across a dozen policy pages, the free tier tests a compressed version of your knowledge base, so a handoff during the trial might be a doc-count limit rather than a real gap. The free tier also does not include the reply-style scanner that learns from your past sent emails, which is a Growth-plan feature, so free-tier drafts are the AI working from your docs alone rather than from your docs plus your learned tone. That shapes how the drafts sound, less whether they are correct. Read the free-tier result as a floor: if it is good enough to decide yes, a paid plan with learning and a fuller document set only goes up from there. The full breakdown is on the pricing page.

The three numbers to watch

You do not need a dashboard for a one-month trial. You need a notes file and three numbers, tracked the same way every week. None of them require anything beyond reading your own drafts honestly.

Volume handled

Of the support contacts that came in during the trial, what share did the AI produce a usable answer for. A draft you sent with no changes counts. A chat the AI resolved on its own counts. A message where you had to write the reply from scratch does not. This is the headline number, closest to the real question: how much work would this have taken off my plate. Track it as a simple fraction each week. If the AI produced something usable for 40 of your 50 trial emails, that is 80 percent volume handled, which tells you more than any feature list.

One caution: volume handled is not volume attempted. An AI that confidently answers everything, including the questions it should have refused, posts a high number and a pile of hidden errors. Count a reply as handled only if it was correct and you would have sent it, which is what draft mode lets you check.

Edit rate

Of the drafts the AI wrote, how many did you send unchanged versus how many you had to fix first. A low edit rate means the AI has good documentation for that category and sounds close enough out of the box. A high edit rate splits two ways. If you are fixing facts, the underlying document is wrong or ambiguous, and that is fixable. If you are fixing tone, the draft is correct but does not sound like you yet, which on the free tier is expected because the learning feature is off. Separating the two is the most useful thing you can do during the trial: one is a documentation problem, the other mostly resolves on a paid plan.

Escalation rate

How often the AI declined to answer and routed the message to you instead, rather than guessing. People read this one backwards. A high escalation rate looks like the AI failing, but an AI that hands off the questions it cannot answer from your docs is doing exactly what you want. The failure mode is the opposite: an AI that never escalates and instead invents an answer to a question your documentation does not cover. During the trial, read your escalations and ask one thing of each: was this genuinely something a human should take, or was the answer sitting in a doc the AI should have found. The first kind is the system working. The second is a documentation problem worth noting.

These three numbers interact, and the interaction is where the real picture lives. High volume handled with a low edit rate and a sane escalation rate is a clear yes. High volume handled with a high edit rate means the tool fits but your docs need work first. Low volume handled with a high escalation rate usually means your knowledge base is too thin for the AI to stand on, which the three-document free-tier limit can exaggerate. For the ongoing version of this measurement once you are past the trial, the fuller guide is at /blog/measure-ai-customer-support-quality-small-team.

The traps that make a month lie to you

Even a well-set-up trial can mislead you. The most common mistake is quitting early. You run it hard for three days, the novelty wears off, and the last three weeks are a trickle of half-attention. A month of real reading is the test. Three good days and a lapse is a demo with extra steps.

The second trap is moving the goalposts in your head. You decide beforehand that 70 percent volume handled would be a yes, the trial comes in at 68, and you talk yourself into a no because one draft annoyed you on a Tuesday. Or the reverse: you wanted it to work, so you count a draft you heavily rewrote as handled. Write down your threshold before the trial starts, while you are neutral, and hold yourself to it.

The third trap is judging the AI on the questions it should never answer. If you spend the month fixated on how it handled an angry customer demanding a refund outside your policy, you are testing the wrong thing, because that message should go to a human on any plan. The questions worth judging are the lookups: the facts that live in your docs and used to require you only because they were in your head. I wrote about which families to keep away from the AI entirely at /blog/support-questions-never-let-ai-answer. Keep those out of the scorecard.

Reading the result: the go/no-go

At the end of the month you are making one of three calls, not two: yes, no, and the one most common for a small business, maybe-with-work. Knowing the third exists keeps you from throwing away a tool that would have been fine.

A yes looks like this: volume handled was at or above the threshold you set, your edits were small and mostly about tone rather than facts, and the escalations were genuinely the questions a human should take. That is a tool that fits as it stands. Move to a paid plan and start the real adoption arc, automating the safe categories once you have watched them perform, which the week-by-week guide at /blog/draft-first-ai-email-first-90-days walks through.

A no looks like low volume handled that did not improve across the month, escalations that were really the AI failing to find answers that did exist in your docs, or drafts confidently wrong about facts. If that is your month, believe it. A tool that could not handle your questions with your docs in front of it will not improve because you started paying.

The maybe-with-work case is the one to slow down on. Volume handled was mediocre, but when you read the misses, most trace back to a document that was vague, missing, or contradicted another one. That is not the AI failing to fit. That is your knowledge base not being ready for any AI, including the expensive ones. The honest move is to spend a few hours fixing the worst two or three source documents and run a second short trial. The quality of AI replies is mostly the quality of the documents behind them, and a trial is often the first time anyone reads those documents critically. How the AI pulls answers from your docs, and why structure matters more than volume, is at /blog/how-ai-reads-your-docs-chunking-retrieval.

When one free month is not enough to decide

A free-tier month is the right test for a small inbox making a first decision. It is the wrong test in a few situations. If you handle hundreds of support contacts a month, fifty trial emails is too small a sample, and you will want a paid month to see real volume. If what you most need to judge is whether the AI can learn your voice, the free tier cannot show you that, because the reply-style learning is a Growth feature. And if the decision depends on how a few teammates share the queue, the free tier is a single-seat test. Even then the free tier does a job: it tells you cheaply whether the tool is worth a paid month of deeper evaluation. Use it as the first gate, not the only one. Most small operations never need the second, because fifty honest examples settle the question.

How Trigli handles this, and what it does not do

Trigli is built to be tested this way. Draft-and-approve is the default, so a trial shows you every reply before it goes out. The AI answers from the documents you upload and nothing else, with a confidence floor that routes uncertain replies to a human instead of guessing, which is what makes the escalation number meaningful rather than noise. You can connect Gmail or Microsoft 365, so the trial runs on your real inbox and the drafts land where you already work. The free tier is 50 emails, 25 chats, and 10 tickets a month with 3 documents, no card required.

What the free tier leaves out is worth repeating: no reply-style learning from past sent emails, so free-tier drafts are the AI working from your docs alone, and three documents, so a thin-doc handoff may be the limit rather than a real gap. More broadly, Trigli is not a full enterprise help desk. There is no phone or SMS channel, no per-agent productivity analytics, and no native store integration, so moving money stays a human step. If you are a ten-person CX org that needs those things, the honest recommendation is a heavier tool. If you are a small team on Gmail or Outlook deciding whether AI support fits at all, a careful free-tier month is exactly the test to run, and Trigli is built to pass or fail it honestly.

The short version

A fair trial is a fit test, not a demo. Point the AI at your real inbox, keep it in draft mode so you see everything, and load your actual documentation including the messy parts. Watch three numbers: how much volume it handled correctly, how often you had to edit the drafts and whether those edits were facts or tone, and how often it escalated instead of guessing. Pick your pass threshold before you start and hold to it. Read a mediocre result as a possible documentation problem rather than an automatic no, fix the worst docs, and run it again. Thirty honest days will tell you more than the demo ever could.