BFCM Support Triage: What to Automate Before Traffic Multiplies

Every DTC brand knows the Black Friday support wave is coming, and almost every one waits until it’s breaking over them to decide what to do about it. By then the only lever left is overtime. The teams that come out of the weekend with their CSAT intact made one decision in October instead: not whether to automate support, but what to automate first — because when volume multiplies, you cannot hand an agent everything at once, and the order you pick decides whether the surge costs you money or quietly makes you some.

This is a triage problem, not a technology problem. The question isn’t “can AI handle support” — it’s which conversations you route to an agent before traffic hits, which you keep in front of a human on purpose, and how you rank the queue in between. Get the ranking right and your people spend the weekend on the tickets that actually need a person. Get it wrong and you’ve automated the wrong third of your volume while the revenue-critical conversations sit in a two-hour queue.

The BFCM math: volume multiplies, your team doesn’t

The structural problem is simple. Holiday retail spend keeps climbing — U.S. holiday sales have run into the trillion-dollar-plus range in recent seasons (per Deloitte’s holiday retail forecasts) — and for a promotion-driven DTC brand that spend compresses into a handful of days. Order volume over the Cyber Weekend can land several times above a normal week (Klaviyo’s peak-season benchmarks, vendor-published), and support tickets track orders. Your headcount, meanwhile, is fixed. You can add a few seasonal reps, but you can’t 8x a trained team in a weekend, and the customers don’t wait — a large majority tell researchers they expect a fast response and lose patience quickly when they don’t get one (Zendesk CX Trends, vendor-published).

Note the last number. Independent analysis of AI support deployments lands well short of the marketing headlines once you subtract the escalations and the tickets that come back — realistic net deflection sits closer to 20–35% (consistent with McKinsey’s work on GenAI in operations). That’s still enormous at 8x volume. But it means you have to spend that automation budget on the right conversations. Which is the whole point of triage.

Rank every ticket type on three axes

Before BFCM, pull your last 90 days of tickets and score each category on three things. This is the entire method:

• Frequency.How much of your volume is this? Automating a category that’s 2% of tickets buys you nothing under surge. Automating the top category buys you the weekend.
• Resolution complexity. Can it be finished with data the agent can reach — order record, tracking, returns policy, sizing chart — or does it need judgment, an exception, or an apology that means something? Low complexity is safe to fully automate; high complexity is where autonomy gets you in trouble.
• Revenue proximity.How close is this conversation to a dollar changing hands? A pre-purchase sizing question is a sale in progress. A “where’s my order” is a refund and a lost repeat customer if it’s ignored. Proximity decides what you protect first.

A ticket type that’s high-frequency, low-complexity, and revenue-proximate is the first thing you automate. A low-frequency, high-complexity, emotionally charged ticket is the last — and maybe never. Everything else falls in between. Here’s how the common DTC categories usually score.

Automate first: WISMO and pre-sale questions

Order-status questions — “where is my order”, or WISMO — are routinely the single largest ticket category for e-commerce brands, on the order of a third of volume and higher during peak (Zendesk CX benchmarks, vendor-published). They are also almost pure lookup: the answer lives in the order record and the carrier tracking. This is the ideal first automation — highest frequency, lowest complexity, and every unanswered one is a customer heading toward a refund request or a chargeback. An agent that resolves WISMO in-thread over SMS, with the live tracking pulled in, takes the biggest single slice off your queue on day one.

Pre-sale sizing and product questions are the other “automate first,” for a different reason: they’re the closest thing to a sale you have. During a flash sale, a shopper asking “will the medium fit a 40-inch chest” is holding a full cart. A two-hour reply means a lost order, not just a grumpy customer. These are usually answerable from the product data and the sizing chart, which makes them safe to fully automate — and the revenue upside makes them urgent.

Automate next: returns, exchanges, and promo issues

Returns and exchanges are high-frequency and mostly rules-based, but they carry a retention decision, so they’re the second wave, not the first. An in-policy exchange (“wrong size, same item”) is fully automatable and, done well, saves the sale — an exchange keeps the revenue a refund gives back. The judgment cases — out-of-policy, “it broke after two weeks” — should route to a human with the full thread attached. The agent’s job there is to triage and hand off cleanly, not to make the call.

Promo-code and discount problems are spiky rather than constant — they cluster the instant a code fails at checkout — but each one is a cart seconds from abandonment. They’re low-complexity and high revenue proximity, which is why they belong in the automated tier: the agent confirms the code, explains the exclusion, or applies the fix before the shopper closes the tab.

What to leave to humans — on purpose

Triage cuts both ways. Naming what you will notautomate is what keeps the deployment trustworthy, and it’s where most “AI-first” peak-season plans quietly fail. What people hate isn’t AI — it’s being trapped by it with no way to reach a person when the situation actually needs one.

The point of automating the routine third isn’t to remove humans — it’s to free them for the conversations where a human is the whole product.

1. Anything with a real complaint attached. Damaged goods, wrong item, an order that missed a birthday. The agent should recognize the emotion, triage it, and warm-transfer with context — fast — not try to talk the customer down.
2. Disputes, chargebacks, and money-back arguments. These are judgment and policy exceptions. Let the agent gather the facts, then put a person on the decision.
3. VIPs and churn-risk accounts. Your best customers and the ones about to leave are exactly who a human should reach. Flag them and route them out of the automated tier by rule.

The mistake isn’t drawing this line — it’s not drawing it, and letting the agent loop a furious customer three times before it gives up. Design the escape hatch first: a human is reachable the moment the customer asks, and the moment the agent detects it’s out of its depth.

Instrument it before the volume, not during

None of this ranking works if you can’t see your own ticket mix. The honest window to instrument closes in the first days of November — after that you’re changing the engine while the car is moving. Before the surge:

• Pull 90 days of tickets and tag them by category, so your top three are facts, not guesses.
• Wire the agent to the systems the top categories need — order data, tracking, returns policy, product and sizing data — and test it against real past tickets.
• Write the escalation rules explicitly: which categories, which keywords, which customer tiers always reach a human, and how fast.
• Set the scoreboard to resolution and revenue protected, not raw deflection — a “contained” ticket that didn’t solve anything is a refund you haven’t received yet.