Conversational artificial intelligence should be judged by what its conversations finish, not by how human they sound. That is the position this article takes, and it answers one question: how should a firm with 10 to 250 staff measure and design conversational AI so it books meetings, qualifies enquiries and resolves requests, rather than producing pleasant chats that go nowhere? The answer changes how you write the agent, which channels you put it on and what you look at each week.
What conversational AI is, briefly
Conversational AI is software that holds a dialogue in natural language, by text or by voice, and acts on it. It combines three things: understanding what someone means, deciding what to say or do next, and producing a reply. Modern systems use large language models for the first and third, and a set of rules, knowledge and connected actions for the second.
On the levelpartner.ai platform this takes the form of AI agents for chat, SMS and WhatsApp, and AI voice agents for phone calls. They share the same knowledge, the same CRM and the same inbox, so a conversation that starts on the website and continues by text is one record.
What most writing on the topic gets wrong
The common framing measures conversational AI by how natural it feels. Vendors demonstrate agents that make small talk, use humour and handle long, wandering conversations. Reviews praise how hard it is to tell the agent from a person.
For a business, that is the wrong yardstick. A visitor who has a pleasant five-minute conversation and then leaves without booking, giving contact details or getting an answer has cost you money: the AI usage and the lead. Meanwhile a blunt, three-message exchange that ends in a booked call is a success.
The second common error borrows from large contact centres, where the goal is "containment": keeping customers away from human agents. For a firm selling professional services, a hand-off to a person is often the best outcome. Measuring containment rewards the agent for keeping prospects away from your salespeople.
Measure completion instead
Define, for each agent, the two or three outcomes that count as a finished conversation. For a sales-facing agent these are usually: a meeting booked, a qualified lead handed to a person, or a clear "not a fit" with a polite close. For a client-facing agent: question resolved, request logged, or escalated with full details.
Then measure the share of conversations that reach one of those outcomes. Everything else is a conversation that stalled. Reading the stalled ones each week tells you exactly where the agent needs work.
Imagine a chat agent holding about 200 conversations a month, averaging nine messages each, and a team pleased with how natural it sounds. Now count completions. Say 34 end in a booked call, 21 in a hand-off and 12 in a clean close: 67 of 200. If most of the rest stall after the visitor asks about price, the fix is a direct price answer and a booking offer at that point. Should that lift completions to about 110 a month, conversations would also get shorter, perhaps six messages on average, because they reach a finish sooner.
Design for the finish, not the flow
Once completion is the measure, design follows from it.
Give every conversation a destination. The agent's instructions should name the outcomes and the moment to offer each. "After the visitor confirms their team size and timescale, offer a call."
Ask fewer, better questions. Each qualifying question should change what happens next. If an answer does not affect routing or booking, drop the question.
Answer the hard question early. Price, eligibility and timescale are where conversations stall. A direct answer, even "we quote after a call, and here is what drives the price", keeps the conversation moving.
Make the hand-off a good outcome. Name the person, give a realistic time and confirm the channel. A clean hand-off is a completion.
The trade-off
Outcome-led agents are more directive. They steer toward a booking or a hand-off, and some visitors will find that less friendly than an agent that chats freely. You will lose a few people who wanted to browse in conversation.
You will also need to accept shorter conversations as a good sign, which runs against instinct. A long conversation feels like engagement. Usually it means the agent did not answer the question that mattered.
For firms whose growth has stalled, the trade is worth it. Your bottleneck is converting interest into conversations with the right person, and a directive agent does exactly that.
Choosing channels
Conversational AI is not one channel. Each one suits a different moment.
- Website chat catches visitors while they are reading, when intent is highest.
- SMS suits fast follow-up after a form or missed call, with short messages.
- WhatsApp suits clients who already message you that way, common in mortgage and property work.
- Voice answers calls out of hours or when lines are busy, where a missed call would be a lost lead.
Start with the channel where enquiries go unanswered today. For most firms that is either out-of-hours web chat or missed phone calls.
Suppose the firm checks a fortnight of activity and finds around 30 web chats unanswered overnight and about 45 calls sent to voicemail during viewings, with voicemails rarely returned the same day. Calls are the bigger and slower leak, so the sensible order would be an AI voice agent for overflow calls first, booking valuations and viewings directly, then web chat a month later.
Setting up a completion-led agent
- Go to AI Agents > Conversation AI and click Agents List.
- Click the agent, then Bot Goals.
- In Information Collection, select fields such as Name, Email and Phone.
- Add custom questions for each detail a finished conversation needs.
- Open Update Contact Field for each question and pick its field.
- Turn on Skip if Already Filled for details the contact already gave.
- Turn on Appointment Booking once the calendar is configured.
- Select Trigger Workflow if a completion should start an automation.
- Set up Email Notification Settings to alert the right person.
- Turn on Enable Conversation Summary so hand-offs carry context.
- Click Save and send sample messages in the test panel.
Reading stalled conversations
The weekly review is where the method pays off. Stalled conversations fall into a handful of patterns, and each has a different fix.
If the visitor stopped after a price or eligibility question, the knowledge base lacks a direct answer. Write one. If they stopped after a qualifying question, that question was too broad or too personal too early; reword it or move it later. If they stopped after a booking offer, the offer came before they trusted you, or the calendar showed no useful times. If they simply drifted off topic, the greeting did not make the agent's purpose clear.
Fix the most common pattern first. One change a week, checked against the next week's completions, beats a dozen changes made at once.
When not to use conversational AI
If your enquiry volume is low enough that a person can reply to every message within minutes, an agent adds little. If your conversations are mostly complex negotiations or regulated advice, keep people at the front and use AI for summaries and follow-up drafts behind the scenes instead.
How it differs across the four industries
For agencies, the main outcome is a discovery call with a qualified budget. For consultants, it is often a clear "not a fit" as much as a booking, because senior time is scarce. For mortgage brokers, it is a complete case summary handed to an adviser, with no advice given by the agent. For real estate, it is a booked valuation or viewing with property and timescale captured.
The method is the same in each: define what finished means, design toward it and read the conversations that did not get there.
Frequently asked questions
Is conversational AI the same as a chatbot?
A chatbot is one form of it, usually text on a website. Conversational AI also covers SMS, WhatsApp and voice, and the better systems share knowledge across them. The chatbot conversational AI article covers the chatbot form in detail.
Should the agent pretend to be human?
No. Tell people they are talking to an AI assistant. Completion rates do not depend on disguise; they depend on useful answers and clear next steps.
What does a voice agent add?
It answers calls that would otherwise go to voicemail and can book directly. See AI receptionist for the voice setup and conversational AI for sales for sales-specific design.
Where can I see good conversation design?
The conversational chatbot examples article shows openings, qualifying questions and hand-offs that move conversations to an outcome.