Contact center artificial intelligence works best in an established firm when it drafts and your people decide. This article answers one question: how do you run AI as an assistant to the staff who already answer clients, so replies get faster without anyone losing control of what is sent? On the levelpartner.ai platform that means a Conversation AI bot in Suggestive mode, trained on a knowledge base you approve, drafting replies that a person reviews and sends. Only once the drafts prove themselves, channel by channel, does any of it move to Auto-Pilot.
Why contact center artificial intelligence should assist first
There are two ways to put AI into customer conversations. One puts a bot in front of the customer and takes people out of the loop. The other keeps people in front and gives them a drafting partner underneath. The second is slower to show a saving and much safer for revenue, because in an agency, consultancy, brokerage or property firm the client relationship is the thing being paid for.
Assist-first also produces the evidence you need for any later decision. Every draft a person sends unchanged is a vote for automating that question type. Every draft they rewrite tells you where the knowledge base is thin or where judgement is needed. After a few weeks you know, from your own data, which conversations a bot could answer alone.
What Suggestive mode does
A Conversation AI bot has three modes. Off means it does nothing. Suggestive means it writes a proposed reply inside the conversation and waits; a member of staff reads it, edits it if needed and sends it. Auto-Pilot means the bot replies on its own. New bots can start in Off, so check the mode before you assume anything is running.
Suggestive mode is the contact centre assistant. The reply appears where your team already works, in Conversations, next to the full history of email, SMS, WhatsApp and chat for that contact. Nothing goes to the client without a human pressing send.
Before you start
- Every client channel your team uses is connected to Conversations, so drafts appear in one inbox.
- A named owner for the knowledge base who approves what the bot is allowed to say.
- Your current answers to the twenty questions clients ask most, written down and checked.
- The web pages that describe your services, fees and process, with anything outdated removed.
- Agreement on which channel goes first. Web chat or SMS is usually the lowest risk.
How to set up Conversation AI in Suggestive mode
The guided setup is the quickest route to a working assistant. These steps follow the platform's own order.
- Go to AI Agents, open Conversation AI and click Create Bot.
- Select Guided Form Setup.
- Click Create New Bot.
- Choose the bot type: General Q&A for an assistant that answers questions, or Appointment Booking if it should also offer times.
- Select an existing Brand Voice or create one that matches how your team writes.
- Configure the Bot Settings.
- Configure the Bot Goals.
- Save the bot.
- Set the bot's mode to Suggestive.
How to train the assistant on your knowledge base
A drafting assistant is only as good as the material behind it. Train it on the pages and answers you have approved, then test before anyone sees a draft.
- Go to AI Agents, open Conversation AI and open the Agent List.
- Select the bot you created.
- Open the Bot Training tab.
- Choose a crawling mode: Exact URL, All URLs with this Path or All URLs in this Domain.
- Enter the full URL, including https://.
- Click Get Data.
- Review the pages found and select only those that answer client questions.
- Click Train Bot and wait until each selected URL shows Trained.
- Click Add Q & A, enter a question in the words a client would use, enter the approved answer and save it.
- Enter a question in Test your Bot and confirm the answer matches your approved content.
Once drafts are flowing, the Response Info option beside supported AI messages in Conversations shows which knowledge base sources fed a reply. Use it whenever a draft looks wrong, so you fix the source rather than the symptom.
Suppose the agency receives around 900 client and prospect messages a week across email, web chat, SMS and WhatsApp. It trains one General Q&A bot on its service pages and 40 approved Q&A entries, then runs it in Suggestive mode on web chat only. Say that in week two, account managers send 55 of every 100 drafts with light or no edits. If the knowledge base owner then adds answers for the questions most often rewritten, 75 of every 100 by week eight is a reasonable target. The pattern to watch for: drafts about invoices and meeting times go out unchanged almost every time, while drafts about scope are rewritten almost every time. That split tells you which topics could move to Auto-Pilot and which never should.
When to move a channel to Auto-Pilot
Move one channel at a time, and only when the evidence is clear. The test is simple: over several consecutive weeks, staff send the drafts for that channel's common question types without changes, and nobody has had to correct a sent reply. If scope, pricing or complaints arrive on the same channel, keep a human handover in place before switching.
Many firms never move some channels at all, and that is a reasonable outcome. A consultancy may run web chat on Auto-Pilot for scheduling and invoice questions while keeping email in Suggestive mode for good, because email is where clients raise the conversations that matter. Phone is a separate decision: Voice AI agents answer calls rather than draft replies, and the AI call centre guide covers that route.
What most writing on this topic gets wrong
Most articles about contact centre AI assume a large room of agents on headsets, and they measure success by deflection: how many customers never reached a person. At 10 to 250 staff that metric points the wrong way. Your queue is an inbox that two account managers check between client work, and keeping clients away from them is not a goal.
The better measures are about your people. Track first response time by channel, conversations handled per person each week, and the share of drafts sent unchanged by question type. If those improve while retention holds, the assistant is working. For the design of a queue where AI answers first, see conversational AI for contact centres, which takes the opposite starting point.
The trade-off
Assist-first is slower to show savings. Headcount does not change in the first quarter, and the gains appear as faster replies and fewer dropped conversations. It also asks more of staff: they must read drafts critically rather than click send, and report wrong answers so the knowledge base improves. A team that treats drafts as gospel spreads mistakes faster than before.
Training the bot on the whole domain in one pass. Old blog posts, retired price pages and job adverts all become material for drafts. Crawl only the pages a client should rely on, and add the rest as reviewed Q&A entries.
Troubleshooting
No drafts appear in Conversations. The bot is still Off, or it is not active on the channel the message arrived on. Check the mode first, then the channel settings in Bot Settings.
Drafts quote outdated fees or policies. A crawled page still holds the old text. Refresh or remove that URL in Bot Training and add a Q&A entry with the current answer.
Drafts are accurate but sound wrong. The Brand Voice does not match how your team writes. Edit it using three or four real replies your best account manager has sent.
The same question gets different drafts. Two sources conflict. Search the trained URLs and Q&A entries for the topic and delete the one you no longer stand behind.
Frequently asked questions
Does contact center artificial intelligence replace our team?
Not in Suggestive mode. The bot writes a proposed reply and a person decides whether to send it. Most firms use the time saved for client work rather than reducing staff. The AI customer support article covers what changes once a channel is automated.
How long should we stay in Suggestive mode?
Long enough to see a stable pattern by question type, usually several weeks per channel. Move a channel to Auto-Pilot only when its drafts are sent unchanged week after week.
Can the assistant work across email, SMS and chat at once?
Yes, the drafts appear in Conversations for the channels the bot is active on. Start with one channel so you can read the results properly, then add the others. Omnichannel conversational AI explains how to plan channel coverage.
Who should own the knowledge base?
One named person with the authority to say what is true about your services and fees. Without an owner, conflicting answers creep in and draft quality falls.