What an assistant does and does not do
The most common confusion is a store owner expecting a full employee, then feeling let down. An assistant that works from your store data is not a salesperson and not a negotiator. It is a first layer that answers questions whose answers you have on record, and leaves everything else to a person who knows the store.
- It does: answer with a listed price, availability, options, opening hours, and a written policy
- It does: reply when nobody is around, instead of leaving the customer waiting for hours
- It does not: decide a discount, approve an exception, or promise a date you have not set
- It does not: know what is not recorded. If your data is stale, its answer is just as stale
- It does not: sense anger in a message, or know this is the third time this customer has complained
Questions that are safe to automate
A safe question has one correct answer that exists in your data, and a mistake in it costs little. Start with this group, because it gives the team time without meaningful risk.
- The listed price of a specific item, provided the price in your data is the real price
- Whether an item exists in a given size or color, if stock is kept in units and not as a general number
- Opening hours, address, and contact method
- The exchange and return policy exactly as written, with no improvisation and no exceptions
- Available payment methods and whether cash on delivery is offered
- The stated delivery time as a window, never as a promise of a specific date
Risky questions that need a person
- Complaints: "the order came late" or "the item is damaged". An assistant may argue or apologize in a way that fuels anger
- Exceptions: "can I return it after the period?" That is an owner's decision, not a rule
- Negotiation: any request for a discount, a bundle price, or a free item
- Individual cases: a missing order, a transfer that did not show up, an unavailable size where they want a dated alternative
- Questions you never recorded: "will the fabric survive washing?" If the fact is not documented, do not let it be invented
- Any conversation containing anger, a threat to leave, or a mention of a past bad experience
A practical rule of thumb: if the answer would be the same for any customer, it is a candidate for automation. If it depends on this particular customer or on your judgment, it belongs to a team member. The rule is not perfect, but it prevents most costly mistakes.
Before switching on: prepare your store data
The assistant mirrors your data, so there is no way to make it more accurate than your shelf. Fix the data first, otherwise you will deliver old errors to customers faster and in a more confident tone, which is worse than not replying at all.
- Prices: one item with two prices in two places is the first thing to settle
- Options: each size and color gets its own unit, price, and quantity, not one total for the item
- Policies: write exchange, return, and delivery as short sentences that can be quoted word for word
- Discontinued items: delete them or mark them clearly as unavailable, so the assistant does not answer about them confidently
- Common questions: collect the last fifty real questions from your chats and write an approved answer for each
A four week safe rollout plan
Do not switch the assistant on across all conversations in one day. Week one is not about the assistant at all, it is about your data. Going step by step gives you a chance to catch an error before a customer does.
- Week one, cleaning only: fix prices, options, and policies, and write twenty approved answers. Turn nothing on yet.
- Week two, price and availability only: the assistant handles just this group, and you review every conversation at the end of the day.
- Week three, peak hours: let it reply during busy hours or outside opening hours, while sensitive conversations stay with the team.
- Week four, measurement: compare how many questions ended unanswered before and after, and how many answers the team had to correct.
The decision at the end of the month is not "did I like it?" but "did fewer questions go unanswered, and did the answers stay correct?" If the answer is no, the problem lies in the data or in how narrowly you enabled it, not in the idea itself.
When it should hand over to a person
The real value of an assistant shows in how fast it hands over, not in how much it answers. An assistant that answers everything and never hands over is more dangerous than one that says "I will ask the store owner and get back to you" at the right moment.
- On any complaint or mention of late, damaged, or wrong
- On a request for a discount, an exception, or a delivery date pinned to a day
- On a question with no answer in the data, however simple it looks
- On any sign that the customer bought before and has history with the store
- On a request to speak to a person, with no attempt to keep them with the assistant
The handover should carry context: who the customer is, what they asked, what was already answered, and who picked it up. A handover without context forces the customer to repeat themselves, which is worse than waiting a little longer.
How to audit the answers
Auditing is not reading everything forever. It is a structured sample at the start, then a smaller sample as confidence grows. The goal is not to prove the assistant is good, but to find the error before a customer does.
- First week sample: every conversation, with no exceptions, because volume is small and errors are likely
- Afterwards: ten conversations a week, from different days and different question types
- Record three fields per audit: the question, the answer, and correct, wrong, or incomplete
- Any error repeating twice must become a change in the data or in the allowed question scope
- Also review conversations that ended in silence, that is where unnoticed errors hide
An example of a simple audit sheet: date, question in brief, what the assistant answered, the verdict, and what you did. After a month you will have a list of recurring errors, which is worth more than any general impression.
What to do when the answer was wrong
An error is not the end of the experiment, but how you handle it decides whether the customer stays. The order that matters: correct it with the customer first, then fix the cause, then switch the assistant off if the cause is still unclear.
An example: a customer asked the price of a bag, and the assistant quoted the old price because the price had been changed in the shop but not updated in the data it reads from.
- Correct it with the customer: "The correct price is [price], and I apologize for the confusion. Would you like me to complete the order at the correct price?" with no long explanation.
- Fix the cause: update the price in the data, and check other items whose price changed in the same period.
- Review the scope: if the error came from a question type the assistant should not answer, block it from that type instead of relying on its attention.
- Log the incident: item, question, answer, correction. A second occurrence means a decision, not bad luck.
And do not hide the error from the customer or excuse it with "the system said so". The customer dealt with your store, not your system, and a short clear apology costs less than an argument.
Common mistakes when rolling it out
- Switching on before cleaning the data: the fastest way to spread old errors in a confident tone
- Enabling it on every question on day one: it hides where an error came from when one appears
- Leaving it with no handover path: an angry customer in a closed loop leaves quietly
- Not telling the customer a reply is automated: honesty sets the right expectation and makes a team takeover normal
- Judging the experiment from one conversation: a decision this size needs at least two weeks and one number to compare
- Relying on it for sensitive follow-up: whoever complained or asked for a refund needs a human name to deal with
If this happens, do this
- The assistant quoted a wrong price: correct it with the customer in the same conversation, then update the data, then find every item whose price changed recently.
- The assistant promised a delivery date: send one clear correction, and block that question type until you write fixed windows.
- The customer wrote "I want to talk to someone": hand over immediately with no questions, and pass the conversation context to whoever takes it.
- The customer did not realize the reply was automated and asked for something impossible: apologize in one sentence, say who will follow up, and do not make them explain twice.
- The same error repeated on one item: remove that item from the assistant's scope instead of switching everything off or letting the error continue.
- The assistant stopped or failed: do not leave conversations open. Turn it off and send a short message to waiting customers that replies will be delayed.
The analysis assistant is a different question
Besides the customer assistant there is another kind: one that reads your store numbers and explains them. This kind is useful even if you never enable the first, because it does not talk to customers, it talks to you.
The risk here is different: not a wrong answer to a customer, but a wrong conclusion you build a decision on. So always ask it for three things: the figure it relies on, the period it compared, and what the data cannot say. If it says "sales fell because a competitor cut prices", ask which of your own data proves that. If none exists, the cause is unknown, and that is a valid answer too.
The general rule for both kinds is the same: keep the decision with you. A tool speeds up reading and replying, but it does not know the store, its history with a particular customer, or the consequences of a decision.
How FIHRIST helps
In FIHRIST, a unified inbox brings together WhatsApp, Messenger, Instagram, email, and web chat; each conversation links to the customer, product, and order; each can have an owner and a status; and follow-up reminders can be switched off. Stock is recorded as units for sizes, colors, and options, each unit with its own price and quantity, each order deducts the quantity sold, and you can print a QR code to identify items and compare item movement across periods. Permissions are set per member. The customer assistant answers from recorded store data only when enabled, you can keep it off and follow up manually, and your team can follow any conversation and step in, with the reminder that any AI can make mistakes. The analysis assistant explains only what the data supports. The interface runs in a browser, in Arabic and English.
Frequently asked questions
Can an assistant answer price questions without errors?
It can be very accurate if the price in your store data is always the real price. Errors usually come from two sources: a price that changed and was not updated, or an item priced in two places. So the practical rule is to update prices in one place and recheck after any change. There is no full guarantee, which is why human takeover stays available.
How do I tell a customer the reply is automated without losing them?
One sentence in the first reply is enough, such as "I am the store assistant, and I can connect you to a team member at any time". What matters is honesty and an easy door out to a person. In practice, honesty reduces anger when an error happens, because the customer expects someone to follow up.
What is the minimum data needed to start?
You do not need everything, but you need at least three things: item names with their options, their real prices, and the exchange and delivery policy written down. Start with the twenty most asked about items and four common questions. A small trial on a narrow scope gives you a clearer signal than waiting for complete data.
Does an assistant help a small store with no team?
It helps more in that case, provided its scope is narrow. A one person store loses many questions because the owner is asleep or busy, and an assistant covers that gap if it returns decisions to the owner. If it is left to cover everything with no review, it multiplies the risk rather than reducing it.
What if the assistant keeps giving a long apology in every conversation?
That usually means the instructions it works from are too broad, so it tries to cover every possibility. Set the style with short limits: one apology sentence, then one practical step, then let the customer continue. Review ten conversations afterwards to confirm the length is no longer a problem.