What customers actually want
Not “an AI experience”. Three things:
A correct answer. Quickly. From someone who already knows their situation.
AI can help with all three. It can also damage all three if it is used to put distance between the customer and a person who can help.
The difference is entirely in how it is set up.
Where it genuinely helps
Four uses, in order of how safely they can be adopted.
1. Giving the agent context
When a ticket arrives, the agent needs to know who this is. Recent orders, open invoices, past tickets, delivery history, whether they are already unhappy about something.
In Odoo this is all one database, so a summary can be assembled instantly.
Why this is the best place to start. It makes your team better without changing anything the customer experiences. There is no downside risk.
2. Drafting the reply
The agent gets a suggested response based on the question and the customer’s record. They edit and send.
The saving is real on routine questions — order status, availability, standard policy questions — which are most of most inboxes.
The rule: the agent reads every word before it goes. Generated text is fluent, and fluent is not the same as correct.
3. Routing and prioritising
Reading an incoming ticket and sending it to the right team, with a sensible priority.
Better than a keyword rule, because it handles how people actually write rather than the words you thought to anticipate.
The check. Agents should be able to reassign easily, and someone should look at what gets misrouted.
4. Suggesting knowledge base articles
To the agent, and optionally to the customer before they submit a ticket.
The best deflection there is — a customer who finds the answer themselves in thirty seconds is happier than one who waits an hour for the same answer.
FIGURE 1: FOUR USES, EASIEST TO ADOPT FIRST
Context for the agent
- Summarise the customer’s history. No customer-facing risk at all.
Drafted replies
- Agent edits and sends. Real time saved on routine questions.
Routing and priority
- Better than keyword rules. Agents must be able to override.
Article suggestions
- The best deflection — customers prefer finding it themselves.
Chatbots, honestly
The most oversold application, and the one most likely to damage a relationship.
What they do well
Simple factual questions with a definite answer. Where is my order. What are your opening hours. What is your returns policy.
Routing. Asking two questions and passing the conversation to the right team with context attached.
Out of hours. Something is better than nothing at eleven at night, provided it is honest about what it can do.
What they do badly
Anything unusual. The moment a question is outside the pattern, a chatbot produces a confident irrelevant answer, which is worse than no answer.
Anything where the customer is already annoyed. A frustrated customer being asked to rephrase their question is a customer you are losing.
Anything complex. Multi-part problems need a person.
FIGURE 2: CHATBOTS THAT HELP AND CHATBOTS THAT COST YOU
Set up well
- Handles a defined set of questions
- Says plainly when it cannot help
- Hands over to a person in one step
- Passes the conversation history across
Set up badly
- Claims to handle everything
- Loops when it does not understand
- Hides the route to a human
- Makes the customer repeat themselves
The rule that matters
A customer must be able to reach a person easily, and the bot must admit when it cannot help.
A bot that says “I can’t help with that, let me pass you to someone” after one failed attempt is a good bot. A bot that asks the customer to rephrase three times is worse than no bot, because the customer now has a problem and an opinion about your company.
Measuring whether it worked
Four numbers. Take a baseline before you switch anything on.
First response time. Should improve, and it is the number customers feel most.
Resolution time. Should improve. If it does not, drafting is saving typing but not solving problems.
Reopened tickets. Watch this closely. If it rises, replies are getting faster and less correct, which is a net loss.
Satisfaction. Response rates are low, and the trend still tells you something.
The one that matters most is reopened tickets. Fast wrong answers look excellent on a response time report and cost you customers.
FIGURE 3: WHAT TO MEASURE, AND THE TRAP
First response time
- Improves fastest, and it is what customers notice most.
Reopened tickets
- The honest test. If it rises, you are answering faster and worse.
Satisfaction trend
- Low response rate, but the direction is informative.
What to avoid
Five specific mistakes.
Automatic replies with no review. A confident wrong answer sent to a customer is worse than a slow correct one.
Hiding the human. Customers notice, and they resent it more than they resent waiting.
Pretending it is a person. If a bot is answering, say so. Being caught out costs more trust than the automation saves.
Leading with a chatbot. Start with agent-facing tools where there is no customer risk. Move outward only when those are working.
Ignoring reopened tickets. The metric that tells you whether speed came at the cost of quality.
A sensible order
Weeks 1 to 4. Context summaries for agents. Measure your baseline properly. No customer sees any change.
Weeks 5 to 8. Drafted replies, with agents editing everything. Watch reopened tickets.
Weeks 9 to 12. Routing and article suggestions. Check what gets misrouted.
Only then. Consider a chatbot, for a narrow defined set of questions, with a one-step route to a person.
Reverse that order and you will meet the failures first, in front of customers.
What Odoo brings to this
Odoo Helpdesk sits on the same database as sales, inventory and accounting, and Odoo now includes AI across the system.
The practical result: when a ticket arrives, everything about that customer is already there. Their orders, their invoices, their deliveries, their previous tickets. Nothing needs integrating and nothing needs syncing.
An AI support tool bolted onto a separate helpdesk can only see support history. That difference shows up in the quality of every suggestion it makes.
The short version
AI in support is genuinely valuable behind the agent — summarising, drafting, routing, suggesting.
It is risky in front of the customer unless narrowly scoped and honest about its limits.
Start behind the agent. Measure reopened tickets. Keep the path to a person short.
Get those three right and support gets faster without getting worse.
Support taking too long, or answers getting inconsistent?
Get in touch. We start with agent-facing tools and a proper baseline, so you can tell whether it actually worked.
