Choosing an AI assistant for your DTC (D2C) brand: 5 things the demo won't show you
Every AI assistant demos well. That is the problem.
Book five demos this week and you will watch five near-identical performances: a customer asks a question, the agent answers beautifully, everyone nods. Nothing you saw in those thirty minutes tells you which one will still be earning its keep in six months.
The demo is a scripted conversation on a clean catalogue with a friendly question. Your customers are none of those things. So the decision has to be made on things the demo does not show you.
Here are the five things to check instead — and one decision to make before you check any of them.
Before the five: decide what job you are hiring it for
Most of these tools are good at one job and passable at the other two. Pick your job before you shortlist, or you will end up buying whichever vendor pitched best.
Deflection. You are drowning in tickets and want fewer of them reaching a human. Helpdesk-first tools were built for this.
Selling. Your traffic is fine and your conversion is not. You want something that helps an unsure visitor decide. Commerce-first tools were built for this.
Retention. Your first order is healthy and your second is not. You want someone answering the questions that happen after the box arrives.
If you genuinely need all three, say so out loud in the first call and watch how the vendor reacts. Most will say yes. Ask which one they were built for first, and what the other two looked like on day one.
1. Where it actually lives
Most AI assistants live on your website. Your customer does not.
She sees a reel, replies in Instagram DMs, asks a follow-up on WhatsApp two days later, then opens your site on her laptop to buy. If the assistant only works on the storefront, it is absent for the first two-thirds of that journey.
Ask specifically:
Does it work in Instagram DMs, or only Instagram comments?
Is WhatsApp native, or a separate product with separate pricing?
Who owns the WhatsApp Business API number, and what happens to it if you leave?
That last one catches people out. Getting a number migrated off a vendor is an unpleasant surprise to discover during a switch rather than before one.
2. Whether it remembers her
There are two kinds of memory, and vendors blur them.
Session memory means it remembers what was said five messages ago, within one conversation. Almost every tool worth considering has this now. It is table stakes, not a feature.
Customer memory means it recognises the same person across sessions and across channels. She told you about her sensitive skin in an Instagram DM in March. In June she asks about a new serum on your website, and the agent already knows. That is rarer, and it is what separates an assistant from a very articulate FAQ.
Two questions that cut through the marketing:
How do you identify the same customer on Instagram and on my website? (Look for email or phone matching, not cookies.)
If I open a WhatsApp chat with a question and then visit the site an hour later, what does the agent know?
Ask them to show it live, with their own demo account, on two channels. The demo that cannot do this will suddenly become a conversation about the roadmap.
3. Whether it can act, or only answer
An assistant that answers "you can track your order in the confirmation email" has not helped anyone. One that pulls the tracking status and tells her the parcel is out for delivery has.
The gap between those two is integrations. Check which of these it can do on its own, without a human picking up the thread:
Action | Why it matters |
Read live catalogue and stock | Stops it recommending things you cannot ship |
Fetch order status | Removes the single most common support question |
Start a return or exchange | Turns a refund request into a size swap |
Request or post a review | Retention work nobody has time for |
Apply a discount or build a cart | Lets a conversation end in a purchase, not a link |
Ask which of these are live today versus on the roadmap. Ask for the integration list in writing. A tool that reads your catalogue but cannot see stock will confidently sell a sold-out size.
4. What stops it saying something stupid
This is the fear every founder has and most are too polite to lead with. It is a reasonable fear. An agent wearing your brand's name can promise a delivery date you cannot meet, or tell someone a supplement is safe during pregnancy.
What to probe:
Where does its knowledge come from? Your product pages and approved claims, or the open internet? The second one is how you end up with a confidently invented ingredient.
What happens at the edge of what it knows? Good agents say they do not know and hand over. Bad ones improvise.
How does escalation work? Does it hand to a human mid-conversation with context attached, or dump the customer into an email form?
Can you see what it said? Full transcripts, searchable, with a way to correct a bad answer so it does not repeat.
The brand voice question matters too, but it is easier to fix than accuracy. Tone can be tuned in a week. A tool that fabricates cannot be tuned out of it.
5. How it charges, and who keeps the data
There are three pricing models in this market, and they reward very different behaviour.
Per resolution or per ticket. Cheap at low volume, and it quietly penalises you for growing. Model it at three times your current volume before signing.
Flat platform fee. Predictable, usually with a floor high enough to price out smaller brands. A fee of several hundred dollars a month plus per-conversation costs is common at the enterprise end.
Per seat or per store. Simplest to forecast, often the worst fit if conversation volume is your real variable.
Whatever the model, get the all-in number for your actual volume in writing, including the onboarding fee that appears in month one.
Then ask the question almost nobody asks: who owns the conversation data? Every chat is a customer telling you what they want, in their own words. That is the most valuable research your brand will ever collect. Find out whether you can export it, whether it trains a shared model, and what you keep if you leave.
A two-week evaluation you can actually run
You do not need a procurement process. You need two weeks and a bit of discipline.
Pull twenty real conversations from your inbox, DMs and WhatsApp. Pick the messy ones: the vague question, the angry one, the person asking about a product you discontinued, the one who switches topic halfway.
Write down what a good answer looks like for each. Do this before any demo, while you are still objective.
Shortlist three tools, no more. One from each camp if you are unsure which job you are hiring for.
Run your twenty conversations through each demo. Not their script, yours. Send the questions verbatim, typos included.
Score only two things: did it understand what the customer actually wanted, and did it do something useful about it. Ignore how charming the reply was.
Pilot the winner on one channel for 30 days. Your lowest-stakes channel, not your highest-volume one.
Measure against your own baseline, not the vendor's case studies: response time, conversion on assisted conversations, tickets reaching a human, and repeat purchase rate on customers who chatted.
Step four is where most shortlists collapse from three to one. Real customer messages are untidy in ways scripted demos never are.
Who should not buy one yet
Some brands are better off waiting, and the honest vendors will tell you so.
Very low conversation volume. Under roughly twenty inbound messages a week, you are the AI agent, and you are better at it.
A catalogue in disarray. If your product data is thin, inconsistent or out of date, the agent inherits all of it. Fix the catalogue first; it improves your site either way.
A single, simple product. If there is genuinely nothing to choose between, there is not much of a conversation to have. A good FAQ may be the right answer.
No clear owner internally. These tools need someone reviewing transcripts weekly for the first month. Without that, quality drifts and nobody notices until a customer complains publicly.
None of these are permanent. They are just reasons the money is better spent elsewhere this quarter.
The test that actually matters
Strip away the feature comparison and one question is left: does it behave like the best salesperson you ever had?
That person did not recite the catalogue. They asked what you were trying to do. They remembered you from last time. They told you honestly when something was not right for you, and you trusted them more for it. They were still useful after you had paid.
That is the bar. Every tool on your shortlist should be measured against a person, not against each other.
Most of them are not there yet. Pick the one that is closest, and be honest with yourself about which of the five you are willing to compromise on. There is always at least one.
Comments