The most useful chatbot best practices after launch are to keep business facts current, make the AI's role clear, handle uncertainty honestly and review what happens after an answer. Give each correction a specific home: the source, the behavior or the skill. Then check that the change helps the customer finish the job they came to do.
Ultimo Bots is an AI agent that answers visitors using your business knowledge and uses configured skills to collect requests or perform supported app actions.
Chatbot best practices start with a clear customer job
An agent that has been running for a month needs an owner who can explain what a useful conversation accomplishes. For a rental company, that might mean explaining delivery conditions and capturing requests for an individual quote. Use that purpose to decide which conversations deserve attention and which improvements matter.
A busy conversation is not necessarily a successful one. A visitor may send several messages because the first answer missed a condition. Someone else may leave after one accurate reply because they now have everything they need. Read the exchange before deciding what its length means.
Keep the original job visible as the business changes. If you add upstairs delivery, the agent needs the approved conditions. If staff start handling a new kind of request, agree what information they need before asking the agent to collect it. Changes in the service should trigger changes in its supporting content.
This gives ongoing maintenance a manageable boundary. You are checking whether the agent still explains your current service and provides the promised next step. Expand its responsibilities when you can identify the source, action and person who will support the additional work.
What are the best practices for using AI chatbots?
Start with identity, scope and a useful response. Visitors should understand that they are speaking with AI, what help is available and what happens when their question needs a person. Review those expectations whenever someone changes the agent's introduction or the service it represents.
Ultimo Bots displays an AI notice in its website widget and chat link. Treat that notice as part of the customer experience: keep the surrounding language clear about the business and the agent's role. A friendly name can coexist with an explicit AI identity.
For EU-facing use, the European Commission's Article 50 guidance explains that providers must design systems for direct interaction so people know they are interacting with AI, unless this is obvious. Under Article 50(1) and (5), the notice must be clear and distinguishable, appear at the first interaction at the latest and meet applicable accessibility requirements. The Commission says this rule has applied since August 2, 2026. The built-in website notice supports this disclosure task; your wider legal obligations depend on your use. As of October 2026; practical guidance, not legal advice.
For day-to-day service, make the opening promise specific enough to keep. “Ask about delivery or request a quote discussion” gives a visitor a clear starting point. Revisit it when the business changes so the invitation stays aligned with the help actually available.
Give each important answer an accountable source
Choose who approves each important business fact. Delivery conditions may belong to operations, prices to the owner and cancellation terms to the person responsible for contracts. One person can cover all three. What matters is that an editor knows whose answer is authoritative when two documents disagree.
Suppose your rental brochure says delivery includes carrying items inside, while the current service page says curbside delivery. Uploading both does not resolve the disagreement. Ask the service owner to approve the current policy, correct the outdated material and update the agent's knowledge base. Keep the exception beside the main condition so neither has to be inferred.
Ultimo Bots instructs the agent to look up company facts and ground its answer in the returned information. You still need to judge whether the answer preserved the condition that matters. “Delivery is available” may be true while leaving a visitor with the wrong expectation about where the items will be placed.
Prioritize sources that affect decisions: prices, eligibility, opening times, service areas and cancellation conditions. Mark temporary offers with their real end conditions, and assign someone to remove obsolete information. A smaller, coherent set of approved facts is easier for your team to maintain than a collection nobody has read recently.
Correct the source before adding another instruction
When an answer disappoints you, first identify what needs to change. A wrong price and an overly long explanation are different problems. Adding a general instruction to “be more accurate and concise” leaves the price conflict unresolved and makes the intended improvement harder to observe.
| What you observed | Where to investigate | A focused correction |
|---|---|---|
| The answer used an outdated condition | Business knowledge | Replace or correct the obsolete source |
| The answer was accurate but buried the condition | Behavior | Ask for the condition near the start |
| The visitor requested an action that was not completed | Skill and its configured action | Review the required details and intended result |
| The request arrived but nobody followed up | Team responsibility | Assign a person to respond |
In Ultimo Bots, behaviors guide the agent's responses, while skills define tasks. An app connection makes the app available; a skill must include its app action for the agent to use it. Check the relevant part before changing unrelated settings.
For a response rule, open the AI Studio, choose Add behavior, then Enter manually. A focused instruction is: “For delivery questions, put the applicable delivery condition before extra detail. Keep the answer brief and offer a quote discussion when the visitor needs an exception.”
That rule is short enough for the manual behavior field. It addresses presentation while leaving the actual delivery conditions in the knowledge source. After saving it, read replies to ordinary delivery questions and exception requests. Preserve useful detail while removing wording that gets in the customer's way.
Make uncertainty lead to a useful next step
An unanswered question can reveal a missing fact or a decision the business deliberately reserves for staff. Resolve that distinction before adding content. If the policy exists, make it available. If approval depends on circumstances, explain who decides and give the visitor a route to that decision.
Here is an illustrative conversation for the invented Alder Event Rentals. Its approved Q&A says the standard delivery fee covers curbside drop-off; upstairs delivery requires a separate quote and staff confirmation. These are example facts, not a customer case study.
Visitor: Does standard delivery include bringing tables upstairs?
AI: Standard delivery covers curbside drop-off. Upstairs delivery needs a separate quote and confirmation from the team.
Visitor: It's only one flight. Can you promise they'll do it?
AI: The team needs to confirm upstairs delivery. Would you like to send them a request to discuss it?
The useful result is that the condition survived the follow-up. The agent has offered to collect a request; nothing has been sent yet, and upstairs delivery has not been approved. Read your own conversations for that same distinction between explaining, offering and completing.
The built-in Lead Generation skill collects a name and email by default, confirms the visitor's details and forwards the request. Once a visitor chooses that route, check the saved request and conversation in Leads. The colleague responding should understand both the question and the decision still needed.
Keep collection proportionate to that next step. Add another required field only when the responder needs it. As you apply chatbot security best practices, direct sensitive account or payment matters to an appropriate approved process and avoid collecting unrelated private details merely because a conversation makes it easy to ask.
Keep personal help available when it is staffed
On the website and chat link, Ultimo Bots supports live chat takeover while a teammate is online. Make online status an operational responsibility. The person offering live help should be able to notice a request and join, and should switch offline when they leave that responsibility.
Keep the two waiting situations clear. With nobody online, the AI continues answering and can collect a follow-up request through the lead skill. If a visitor requests takeover while someone is online but nobody joins within the configured wait time, the configured offline message appears. Review that wording against the response your team can actually provide.
When reviewing a conversation that needed a person, ask what blocked progress. Was a policy missing? Did the visitor need an exception? Did staff have the information required to respond? Those lead to different improvements. Sending every difficult question to the same place is only useful if someone there can resolve it.
Agree how colleagues return useful findings to the agent owner. A one-off exception can remain a staff decision. A recurring, approved explanation may deserve a knowledge update. The responder should help distinguish the two before the next visitor receives a newly generalized answer.
What are some key metrics to evaluate chatbot performance?
Begin with a manual conversation review that you can repeat. Select a defined period and a manageable sample from the portal Inbox. Read complete exchanges, including those that appear uneventful. If you also inspect every complaint or exception, keep those separately identified so they do not silently change what the sample represents.
In your own review notes, record whether the factual answer was supported, whether the intended next step happened and whether staff follow-up was needed. These are review categories you assign yourself. Give “unclear” its own place rather than treating a visitor's silence as proof of resolution.
Count each category against a stated denominator. For example, report how many reviewed conversations contained a supported answer out of the conversations where you could check the source. For requests, examine the relevant saved result. Do not combine an answered opening-hours question with a quote request as though both have the same completion event.
NIST's AI reliability guidance describes continued testing or monitoring after deployment and notes that some failures cause greater harm. Apply that principle when choosing the next correction. An incorrect cancellation condition deserves attention before a merely awkward greeting, even if the greeting appears more often.
Use repeated observations to guide work, not to promise a performance increase. A small sample can reveal a recurring problem without establishing its frequency across every visitor. Compare similar periods and customer tasks, and ask the people receiving requests whether the change made their next step clearer.
Make source upkeep part of normal business updates
Attach the agent to the business's existing update process. When someone approves new hours, prices or delivery conditions, that person should also notify whoever maintains the relevant knowledge. Waiting for a customer to expose a stale answer makes routine editorial work unnecessarily reactive.
For agents using supported website sources, Ultimo Bots offers optional automatic rereading: monthly on Boost and up to daily on Ultimo. The owner must save the refresh settings. Smart has no automatic website refresh. Treat the schedule as a way to revisit selected website content; your team still approves what the business publishes and checks the resulting answers.
For files and Q&A, include the relevant knowledge update in the same business change. Ultimo Bots supports editing Q&A entries. Verify that processing has finished before checking the revised answer, and look for older material that still expresses the previous condition.
Budget for both capacity and upkeep. As of October 2026, direct portal plans have these USD prices per agent:
| Plan | Monthly billing | Per month, billed yearly | AI replies per month |
|---|---|---|---|
| Smart | $29 | $19 | 500 |
| Boost | $59 | $39 | 2,000 |
| Ultimo | $149 | $99 | 10,000 |
Paid plans start with a free seven-day trial, with a payment card collected at checkout. Use the pricing page to compare allowances with expected use, and reserve staff time for source approval and follow-up alongside the subscription cost.
Try your agent free for 7 days
Agree who has the right to change an answer
Finish with a small responsibility agreement. Write it wherever your team already keeps operating instructions, and name actual people. This is an editorial arrangement you maintain yourself. A solo owner can hold every responsibility; a larger team benefits from knowing where each decision belongs.
| Responsibility | Decision they own | What they pass on |
|---|---|---|
| Business fact owner | The approved policy or service condition | Current wording and its effective date |
| Agent editor | The source, behavior or skill change | The observed result after the change |
| Customer responder | The individual request or exception | Recurring questions that need an approved answer |
For Alder, operations decides whether upstairs delivery is offered and under which conditions. The agent editor updates the approved Q&A and checks how the condition appears in replies. The colleague handling quote requests resolves individual cases and reports recurring confusion. Each person can finish their part without silently deciding someone else's policy.
Agree when a correction needs immediate attention and when it can wait for the regular review. Changed charges or an incorrect service promise should reach the responsible person promptly. Cosmetic wording can join the next editorial session. Keep a backup person for absences so a simple correction does not depend on one inbox being watched.
This is where AI chatbot best practices become routine business care. The next improvement has an owner, a factual basis and a customer outcome someone can check. Choose the most consequential unresolved question from your recent conversations and put it in the hands of the person who can approve its answer.