Here is a story that repeats itself. A company buys an AI tool expecting it to solve half of customer service, then finds a few weeks later that it barely returns generic replies. The model is rarely the problem. The problem is that it was switched on with no system around it: no organized knowledge source to answer from, no permission to actually carry out requests, and no clear plan for when a human should step in. It all comes back to one question. Did you buy a bot that replies, or an agent that acts?
A successful rollout moves through seven connected steps. You start by picking your most repetitive cases, then prepare your knowledge source, build your workflows, connect your channels and systems, design human handoff, test with the phrasings customers really use, and finally launch, measure, and improve. This article is part of our wider guide to , and here we go deep on the execution.
One idea is worth keeping in mind throughout: success does not come from the size of the project, but from getting your first case right before you scale.
What you need before you start
An AI agent does not invent solutions on its own. It builds on what you give it, so it pays to prepare four things before you open any tool:
Current policies and content, so its answers come from your source and not from the open internet.
Access to your systems through the API, because the line between an agent that promises and one that executes is exactly this connection.
A list of the questions your team answers most, so you know where to begin.
An owner for the project, and the metrics you will judge it by.
Miss any of these and the results will disappoint you no matter how strong the tool is. Prepare them first.
Step one: Start with your most repetitive cases
Trying to automate everything at once is the fastest way to fail. Pick a single case that clears three filters: high volume, low complexity, and a clear action you can carry out inside your systems.
Go back through your conversation logs and pull out the five questions that eat your teamâs time, the âwhere is my order?â, âhow do I cancel?â, âwhat is the return policy?â kind. Rank them by volume, then set aside anything that needs complex human judgment for now. Begin with the most frequent case that has the clearest action, because a quick, visible win buys you the trust of your team and your leadership before you expand. And you can see the difference here between a : one deflects, the other resolves. One case finished end to end beats ten half-built ones.
Step two: Prepare your knowledge source
The agent needs clear data to work from, and that is your knowledge source. Rather than training a model from scratch on everything, you connect it straight to your own sources through retrieval-augmented generation (RAG). As long as it stays tied to your source, it is far less likely to invent an answer that does not exist.
You have plenty of sources to draw on: a web crawler that reads your site and re-syncs on a schedule, file uploads in PDF, DOCX, and XLSX, connections to Notion and Google Drive, and product catalogs pulled from Shopify, WooCommerce, Salla, and Zid.
What makes the real difference over time is trained answers. When one of your agents corrects a reply once, that correction is stored and ranked above the content scraped from your site, so the agent gets sharper with every fix instead of repeating the same mistake. Your knowledge source stops being a static file and starts improving from your actual conversations.
A practical tip: break your knowledge into short, focused articles rather than one enormous document, because accurate retrieval needs sources with clear edges. And since that same source feeds your self-service channels, build a unified on top of it that answers customers before they even open a chat. One rule stays non-negotiable: update the content before you connect it, because a stale source produces wrong answers that sound confident.
Step three: Build your workflows
Once the agent knows where to answer from, it needs to know how to behave. That is the job of your workflows, the path it follows in each case.
In a serious setup, you build that path visually from ready-made nodes: a trigger that starts the journey, a step that collects input and validates it, a conditional branch based on the answer, a webhook call that carries out an action, a âlet the AI answerâ node for open questions, and a handoff node to a human when needed. With draft and publish versioning you test the flow before it goes live, and you run the same logic on the web and on WhatsApp, Instagram, Facebook, and email.
Take changing a delivery slot as an example. The agent asks for the order number or infers it from context, checks the order status and whether an alternative slot is free, then makes the change and confirms. A clear flow is what lets it act steadily instead of stalling at the first case outside the script.
Step four: Connect your channels and systems
An agent that cannot reach your customer where they are is only half a solution. Start with the channels your customers use every day, and sits at the top of that list for much of the world.
Then comes the part that turns the agent from ârepliesâ into âexecutesâ: connecting the systems that hold your data. The agent can look up real orders in Shopify, WooCommerce, and Salla and act on them, with the note that product sync on Zid is strong while order lookup there is enabled after verification. Beyond that, you can define Actions in plain words that send a request to any webhook, such as creating a lead in your Odoo system, and connect custom tools through MCP.
The piece many teams overlook is what happens at the moment of escalation. Every conversation that moves to a human arrives pre-sorted: a summary of what happened, a read on the customerâs sentiment, a frustration score from 1 to 10, an automatic priority, and automatic routing to the right team. The agent starts from an advanced position instead of a blank slate. For the full picture of what delivers, the product page lays it out.
Step five: Design human handoff
The goal is not for AI to replace your team, but to free their time for the cases that genuinely need them. That is why any serious system needs a clear path for human handoff.
Decide in advance when the agent should step back. A high frustration score can trigger escalation on its own, and so can a repeated complaint, a request that needs a management decision or an exception outside policy, or a customer plainly asking to speak to a person. When the transfer happens, your teammate moves into the conversation with one click, carrying the full summary and the actions already taken, and can leave internal notes for colleagues and mention them without the customer seeing any of it. Add response timers calculated against your real working hours so no conversation slips through the cracks. The payoff is that the customer never re-explains their problem from scratch, which is the moment that drains their trust fastest.
Step six: Test with real phrasings, across languages
Testing with formal language alone is a trap that costs you after launch. Your customer does not write âI would like to cancel my order.â They write âcancel it,â or âforget it, I donât want this,â and sometimes they mix two languages in a single sentence.
Build a test list of real phrasings for each case, across every language you serve, and run it before you launch. Every phrase the agent misreads is not a failure but material for improvement: turn it into a trained answer that corrects the behavior next time. Real-world testing is the only way to reach a level of understanding that ends conversations in a resolution rather than frustration.
Step seven: Launch, then monitor and improve
Do not open the door to all your customers at once. Start with one case or one channel, watch the results closely, then expand step by step.
Track what actually matters: the share of conversations closed without a human, the split between AI and human resolution, average response and resolution times, and your most frequent intents and topics. And there is a goldmine that often goes ignored, the search insights that reveal the questions your customers searched for and did not find an answer to. That is a ready-made list of what your knowledge source is missing. Gather these signals on a regular cycle, and improve your knowledge and workflows on that basis. Without measurement, you are guessing, not managing.
A worked example from start to finish
Let us tie the steps together in one scene. A customer messages you on WhatsApp: âHi, I want to change the delivery address on my order.â The agent understands the intent and infers the order from the earlier conversation without asking for its number. It looks the order up in your store and checks its stage. If it has not shipped yet, it offers to change the address, makes the change in the store system for real, and sends a confirmation. If it has already entered shipping, it does not promise what it cannot deliver. It escalates to a human with a summary that spells out the customerâs request and the shipment status, so the conversation finishes in seconds instead of minutes. That is the practical difference between a bot that replies and an agent that understands, executes, and knows its limits.
Common mistakes many teams make
Launching with an incomplete knowledge source: it produces wrong answers that sound confident. Start with one case on a current source.
An agent that replies but does not act: it causes frustration and repeat messages. Connect the systems so it executes.
Ignoring the escalation path: it costs you customer trust. Design a clear handoff with a summary.
Testing in formal language only: it fails at the first casual phrase. Test with your customersâ real phrasings.
Launching to everyone at once: it creates wide errors that are hard to contain. Start with one channel and expand on data.
Conclusion
This is not a complex undertaking, it is seven sequential steps: organized knowledge that improves through corrections, clear workflows, a connection that truly executes, thoughtful escalation with a summary, realistic testing across dialects and languages, then continuous measurement that feeds improvement. Start with one case you can get right, and expand based on what the numbers tell you rather than guesswork.
Try it yourself
Start with a single case alongside your current system, and try Mando on the free plan with no commitment. The page shows what fits the size of your business.
Frequently asked questions
How long does implementing conversational AI take?
It depends on how ready your knowledge source is. If your policies and data are organized, you can launch one case quickly, then expand gradually.
Do I need to replace my current helpdesk or CRM?
No. It is better to trial the agent on one channel or case alongside your current system, then expand once you are confident in the results.
How much data do I need to start?
Current policies and FAQs for the case you start with are enough. Freshness and accuracy matter more than volume, and the agent improves later with every correction.
Will it work with my existing channels like WhatsApp?
Yes. It can connect to WhatsApp and other channels, and to order and payment systems, through the API, Webhooks, and MCP.
Can it actually execute requests, or only reply?
It executes permitted actions such as looking up an order and changing a slot or an address, and it escalates exceptional or sensitive cases to a human.
How do I know it is working?
Watch your containment rate, the accuracy of answers against your policies, your escalation rate and its reasons, and the search insights for questions with no answer. Those metrics point you to what needs improving.