Reliable is not a feature you can tick on a spec sheet. It is a measurement, and it has four separate parts. A marketing team can test each one on their own site, with their own content, in under an hour.
The four tests are: does the agent answer from your own published material, does it say when it cannot answer, does it behave the same way on the same question a week later, and does it hand over cleanly when a human is needed.
One failure matters more than the others in marketing. The agent gives a confident answer about a price or a claim your company never made. That is the cost this article measures, and it is the reason the four tests exist.
Why "most reliable" is not a spec sheet claim
Every vendor calls its agent reliable. It is one of the most repeated words in the market, and it means something different in every product description. Some vendors mean the agent stays online. Some mean it answers quickly. Some mean it rarely says it does not know. None of those meanings overlap, and none of them tell you whether the agent will damage your brand on a live chat.
Reliable in marketing has a specific meaning. The agent answers questions accurately from your own published content, it admits what it does not know, it is consistent across time, and it passes complex or sensitive questions to a person without losing the context. Each of those four behaviours can be tested independently. No vendor's marketing copy tests them for you.
Reliability Has Four Separate Parts
The Cost of One Wrong Answer at Scale
The most costly failure is one your customer acts on before you know it happened. That is why the four tests exist.
Test 1: Does it answer from your own published material?
The test is straightforward. Find ten questions your team answers every week. Write them down in the exact phrasing a customer uses, not the phrasing in your documentation. Feed each one to the agent.
For each answer, check two things. First, does the answer match what your website says? Second, does the agent's answer contain anything your website does not say?
A reliable agent answers from your material and nothing else. If it adds a detail about your returns policy that is not on your site, that is a hallucination. If it gives an answer that contradicts your published terms, that is a confident error.
Run this test against your live agent using your actual content. The result tells you whether the agent is a reliable channel for your own information, or whether it is adding its own version of your brand.
Test 2: Does it say when it cannot answer?
This is the test most teams skip, and it is the one that matters most in marketing. Ask the agent things it cannot possibly know. Your pricing for a specific customer scenario. Whether a particular feature exists. What happened on an account three months ago. Whether a particular offer is still running.
A reliable agent says it does not have that information and routes the question to your team. An unreliable agent makes something up and says it confidently.
The difference matters because customers cannot tell the difference between a confident wrong answer and a correct one. They do not have access to your internal documentation. They take the agent at its word, and when that word is wrong, they remember it as your word.
Test 3: Does it behave the same way a week later?
Take the ten questions from Test 1. Ask them again, seven days later, in the same phrasing. Compare the answers.
A reliable agent gives the same answer to the same question about your published content. It may rephrase, but the substance should be identical. If the answer changed, one of two things went wrong. Either the underlying content changed and the agent was not updated, or the model is unstable on your material.
This test catches a failure mode that does not show up in a demo. An agent can perform perfectly on a scripted walkthrough and behave differently on a Tuesday when the model has been updated or when a different server is handling the request.
The practical check is simple: run the same ten questions twice and note every difference. If your team cannot explain why a difference appeared, that is a reliability gap.
Test 4: Does it hand over cleanly?
Pick the three most complex conversations your team handles. Ask the agent something that requires those conversations: a multi-part complaint, a question about a past order, a request that spans two products. Then stop responding as the customer and ask to speak to a person.
A reliable agent hands over with the full context of the conversation. It does not start again. It does not ask the customer to repeat what they already said. It passes the record intact, including what it answered and what it did not answer.
If the handover requires the customer to start over, or if it loses the agent's earlier answers, that is a reliability gap that costs your team time on every escalation.
The failure that costs the most in marketing
The four tests above catch four different failure modes. One of them causes more brand damage than the others combined. It is the confident answer about something your company never said.
A customer asks about a discount. The agent invents one. A customer asks whether a feature works a certain way. The agent describes a capability that does not exist. A customer asks whether you price-match. The agent says yes. The customer screenshots the chat, shares it with your team, and expects the price.
When that happens, you have a choice. Honour the invented offer and train the agent by rewarding its error. Or refuse the offer and explain that the agent lied. Neither option protects your brand.
This failure is more damaging than a slow response or a confusing menu structure. Those are friction. A hallucinated answer is a false commitment made in your name.
The cost is not only the customer interaction. It is the support ticket your team receives, the management time to investigate what happened, and the chance that the customer posts the screenshot somewhere public. In a marketing context, where your agent represents your brand to every visitor on your site, that risk is present on every conversation.
What the cost looks like in time
Every hallucinated answer that reaches a customer generates a support response. Your team has to read the original exchange, understand what was claimed, decide whether to honour it, and write back. That takes time regardless of how the team resolves it.
The arithmetic depends on your own volumes and staffing costs, which are not published figures. What the calculation shows is that the cost scales with message volume and with the proportion of conversations that go wrong. A more accurate agent does not just feel better to use. It reduces the support overhead that follows every confident error.
What to measure on your own site
Once you have run the four tests above, you have a baseline. Return to it every quarter. The measurements that matter are not abstract reliability scores. They are things you can count.
First, the containment rate: the proportion of conversations the agent resolves without escalation. Second, the escalation quality: whether the handed-over conversation contains enough context for your team to act without asking the customer to start again. Third, the hallucination rate: how many answers contain information that does not appear on your site. Fourth, the consistency score: whether the same ten questions get the same answers on two runs a week apart.
Record those four numbers. They tell you whether the agent is improving, whether a model update has changed its behaviour, and whether your content updates have reached the agent.
Your marketing stack generates these numbers across every channel your agent touches. The efficiency of that stack depends on the agent reducing ticket volume rather than creating it. A reliable agent lets your team focus on the conversations that actually need a person. An unreliable one adds work to every channel it touches, including the ones you did not expect it to affect.
How channel-specific risks change the stakes
The four tests apply to every marketing channel, but the cost of failure varies depending on where your customers find you. A wrong answer on a landing page reaches a smaller audience than a wrong answer in a channel your team uses to run a campaign.
Consider how information travels when the agent makes an error. The customer screenshots the chat. They post it on social media. They tag your account. Your marketing team sees it before your support team does, and the damage is already public. The longer that cycle takes to resolve, the more it affects how new visitors perceive your brand before they ever ask a question.
On a platform like Instagram, where DMs arrive directly and screenshots spread fast, the window between error and exposure is short. The agent does not need to be active on that platform for the error to travel there. It only needs to happen on your site, and for one customer to find their screenshot useful to others. A reliable agent on your site limits the source of those screenshots. An unreliable one generates them as a routine output of normal traffic. The same four tests apply wherever the agent answers, including an Asyntai chatbot for Instagram, because a direct message is as quotable as a chat window.
When you are running a campaign that drives traffic to a specific landing page, the agent handles a concentrated spike of first-time visitors. Those visitors have no prior experience with your brand and no context to question what the agent tells them. They arrived because of something your marketing team published, and they trust the agent because it is on your domain. That trust is an asset the unreliable agent converts into a liability.
How Asyntai handles the four tests
Asyntai answers from content you publish, not from training data that may include your competitors' material or outdated versions of your own pages. When it cannot answer, it says so and routes the conversation to your team. Its append-only audit log, on Enterprise, records what it answered and when, so you can check consistency across time without relying on a demo. When a handover is needed, the conversation record transfers intact.
Asyntai's output classifier runs separately from the prompt. It checks the answer after generation rather than instructing the model to behave, which means it catches confident errors before they reach the customer rather than hoping the model follows a rule.
The pricing is self-serve from $39 per month for 800 messages on the Starter Plan, up to $449 per month for 50,000 messages on the Pro Plan. The price per message decreases as volume increases, which means the cost of each accurate conversation falls as your marketing agent handles more of your inbound load.
Run the four tests against Asyntai on your own site. Start a free account with 50 messages and test with your own content before committing to a paid plan.
What changes the answer
The four tests are the same for every marketing team, but the results depend on three things that vary.
The first is content freshness. An agent that answers from your published material is only as reliable as that material. If your pricing page changes, the agent needs to be updated before it answers questions about the new prices. Outdated content is not a model failure. It is a process failure that the model will reproduce faithfully.
How your site publishes changes decides how quickly they reach the agent. A platform with a direct integration keeps the agent in step with the page. A platform without one leaves a manual step, and the gap between publishing a change and the agent seeing it is exactly where Test 3 finds a difference between Monday's answer and next Monday's.
The second is question specificity. General questions get general answers. If your customers ask about your brand in ways your documentation does not anticipate, the agent cannot invent reliable answers from material that does not exist. The investment that makes an agent reliable is writing the answers your customers need before they ask.
The third is escalation design. A clean handover requires that your team is reachable, that the escalation path is configured, and that the agent has permission to end a conversation and start a ticket. If the agent cannot hand over, it either stays silent or invents an answer rather than admit it has nowhere to go.
What to do next
Run the four tests this week. Start with the ten questions your team answers most often. Ask them of your current agent, or set up a free account and ask them of Asyntai with your content loaded. Write down every answer that differs from your site, every answer that contradicts your terms, and every question the agent answered confidently when it had no source material to draw from.
That document is your reliability baseline. Return to it after any model update, any content refresh, or any change to your product lineup. A reliable agent is not a property you buy once. It is a standard you maintain by checking.
Asyntai pricing published at asyntai.com/pricing, September 2026. Self-serve plans: Starter $39 for 800 messages, Standard $139 for 15,000 messages, Pro $449 for 50,000 messages. Free plan includes 50 messages. Enterprise is a signed order form and is not listed in the public pricing table.