Product enquiry agent
We design, build and deploy a WhatsApp AI agent that answers product questions the moment they land — is it in stock, what are the specs, how much, which model fits — straight from your own catalogue and live stock, so buyers get an answer in seconds instead of drifting to a competitor while your staff are busy.
At a glance
Most product questions arrive on WhatsApp and most buyers won't wait — an instant, accurate answer is the difference between a sale and a lost lead, day and night.
A routing front-end into a few read-only handlers on a single model; the real work is wiring WhatsApp, your catalogue/stock and your spec content — no vision, no autonomous money movement, no heavy compliance.
What changes when it works
Time per application
Answering an enquiry by hand
~3–8 min (often much later)
With the agent
~5 sec, instant
Time saved
≈ 3–8 minutes saved per enquiry — about 40–80 hours a month at 800 enquiries. But the real win is revenue: an instant answer converts far better than one that lands an hour later, and at ~RM 0.08 in AI cost, a single recovered sale a day pays for it many times over.
Indicative; actual time and lift depend on catalogue size and how many enquiries you get after hours.
Our agentic development process
Every use case follows the same seven stages — from framing the problem to production.
Our process, built on Anthropic's agent guidance and the Agent GPA eval framework: Building Effective Agents · Agent GPA
Stage 1 · Frame
The business problem
A Malaysian retailer or online seller gets a constant stream of 'ada stock?', 'berapa harga?', 'ni boleh guna untuk…?' on WhatsApp. Every one needs a staff member to stop, check the catalogue and the stock, work out which product they mean, and type a reply — while three more messages pile up. At peak it's a bottleneck; after hours it's silence, and by morning the buyer has ordered from someone who answered first. The questions are repetitive, but the answers have to be exactly right — a wrong price or a phantom 'in stock' costs you money and trust.
Stage 2 · Map
The manual workflow today
- 01Read the incoming WhatsApp question
- 02Work out which product they actually mean
- 03Check the catalogue for specs and the current price
- 04Check stock — is it actually available?
- 05Type back the answer, and handle the follow-up
Where it breaks: The bottleneck is that every simple question still needs a person to look things up and reply — so answers are slow at peak, absent after hours, and drift out of date whenever price or stock changes.
Stage 3 · Design
The agentic workflow
We build it as a routing workflow: each message is first classified — is this an availability check, a spec question, a price question, a comparison, or something for a human? — then dispatched to the specialised handler for that intent, each answering from your live catalogue and stock, and escalating cleanly to your team when it's a real buying conversation or beyond scope.
The router re-classifies as the conversation shifts — a spec question can turn into a price then a buy — dispatching each turn to the right handler.
Watch a case flow through the agent — it does the reading and matching; a human still makes the decision.
See it in action
Customer message
Instant answer sent
- Air Fryer 5.5L (Black) — in stock, 12 units at your PJ warehouse
- Price: RM 249 (was RM 299) — promo ends Sunday
- COD available in KL · or delivery 1–2 working days
Every figure comes from your live catalogue & stock, never guessed · ~RM 0.05–0.10 in AI cost · a hot lead flagged for a human to close.
Step by step
- 1
Classify the intent
The agent reads the message — in Malay, English, Chinese or the mix people actually type — and works out what's really being asked: is it available, what are the specs, what's the price, how do two products compare, or is this a human matter like a complaint or a bulk deal.
- 2
Identify the product
It resolves the message to a real item in your catalogue — matching model numbers, nicknames, colours and shorthand — and asks a short clarifying question only when it genuinely can't tell which product is meant.
- 3
Route to the right handler
Based on the intent, it dispatches to the specialised handler — an availability check, a spec lookup, a price/promo answer, or a side-by-side comparison — each tuned for that kind of question rather than one generic reply.
- 4
Answer from live data
It answers from your live catalogue and current stock — the real price, the real specs, whether it's actually in stock and where — so it never quotes an out-of-date price or promises stock that isn't there.
- 5
Escalate the buying moment
When the buyer is ready to order, wants a bulk price, or asks something out of scope, it hands a clean summary of the conversation to your team to close — the enquiry becomes a warm lead, not a dead end.
- 6
Hand to your team
Your team steps in only for the conversations that are worth their time — ready-to-buy leads and genuine edge cases — with the context already gathered.
Stage 4 · Build
How we build it with Claude
Because it answers only from your live catalogue and current stock, it can't quote yesterday's price or promise stock you've already sold — the answer is right or it doesn't send it. It never invents a spec or a discount it can't cite, it never takes payment on its own, and every enquiry and answer is logged, so you can see exactly what customers ask and what the agent told them.
Integrations
- WhatsApp Business API / EzyChat
- Your product catalogue & spec content
- Live stock / inventory system
- Optional: your CRM or lead inbox for hot leads
Under the hood
- Pattern: a routing workflow on the Claude Agent SDK — a lightweight classifier dispatches each turn to one of a few specialised handlers (availability / specs / price / compare / escalate), each exposed via an in-process MCP server.
- Context engineering: your catalogue schema, spec sheets and promo rules are retrieved per product and prompt-cached; live stock is read at answer-time, never from a stale copy, so the reply reflects what's actually on the shelf right now.
- Tools: read-only catalogue, spec and stock lookups plus a comparison helper — all read-only, so there is no path by which an enquiry can change inventory, pricing or an order; the only write is creating a lead record.
- Grounding guardrail: answers are constrained to what the tools return — the agent may not state a price, spec or availability it can't cite from live data, which closes off the hallucinated-spec / phantom-stock failure mode.
- Untrusted input: customer messages are treated as untrusted (prompt-injection resistant) so a crafted message can't make it leak internal data or invent a discount; names and phone numbers are handled PDPA-aligned.
- Escalation & recovery: low classifier confidence, a buying signal, or an out-of-scope question routes to a human queue with the conversation summarised — every uncertain path ends at a person, never a confident wrong answer.
- Observability & audit: every classification, tool call and answer is logged and traceable, giving you a searchable record of what customers ask and exactly what the agent replied — useful for both compliance and merchandising.
- Eval harness: a golden set of real historical enquiry chats — messy 'rojak', mixed-language, ambiguous-product and trick questions — with deterministic checks (right product, right price, right stock state) plus LLM-as-judge on tone and routing, regression-tested on every change before it touches live chats.
Stage 5 · Architect
Single model or multi-agent?
Routing an enquiry and answering it shares one short context, so a single well-instrumented model with a cheap classifier front-end is the right call — fast, cheap per enquiry, and easy to audit — rather than a multi-agent design.
We reserve multi-agent designs (≈15× the tokens) for genuinely parallel, breadth-first work — not for a single question-and-answer exchange, where a Haiku classifier into a Sonnet handler is both cheaper and faster.
Which model does what
Classify the intent & match the product
Cheap and fast for the bulk of routing and item-matching, which is most of the traffic.
Answer specs, price & availability clearly
Reliable, well-phrased answers on messy mixed-language questions, with the nuance to compare products.
Handle a tricky comparison or unusual question
Takes the harder cases the classifier flags — multi-product comparisons, edge-case fit questions.
What it costs — an estimate
AI usage — per enquiry
a short classify-and-answer exchange at Claude token rates
~RM 0.05–0.10
AI usage — at 3,000 enquiries/month
scales with enquiry volume; the Haiku classifier keeps the bulk cheap
~RM 150–300 / month
Build (one-off)
connecting WhatsApp, your catalogue and your stock system is a few weeks; we quote fixed after a short discovery
scoped per integration
Ongoing
small next to the after-hours enquiries it catches and the sales it recovers
monitoring + support
Indicative only, shown in Ringgit at roughly RM 4.70 to the USD; per-token rates follow Anthropic's published pricing (confirm current figures). Actual AI usage depends on enquiry length and monthly volume.
Stage 6 · Evaluate
How we measure success
- Deterministic checks first — right product, right current price, right stock state
- Self-verification — the answer must cite live catalogue & stock before it's sent
- Business KPIs — enquiries answered after hours, answer accuracy, and hot leads passed to your team
We score the agent with the Agent GPA framework — Goal, Plan, Action, a current standard for agent reliability — on top of the human baseline. (reference)
Setting the baseline
First we measure your current WhatsApp enquiry handling on a sample — how fast a question gets answered, how many are wrong (stale price, wrong stock), and how many after-hours enquiries are missed. That human baseline is what every agent metric is measured against, so the lift is provable rather than assumed.
What we test — Goal · Plan · Action
Answer accuracy vs a human
The reply matches the correct product, price and stock state, on held-out chats
Correct intent routing
Availability / specs / price / compare / escalate classified correctly
Hot leads escalated
Ready-to-buy and out-of-scope conversations handed to a human — none dropped
Product match accuracy
The message mapped to the right catalogue item
Price & stock correctness
Every quoted price and availability matches live data at answer-time
No invented specs or stock
Never states a spec, price or availability it can't cite from your data
The go-live gate
The agent doesn't answer a live customer until it beats the human baseline and clears these GPA targets on a golden set of real historical enquiry chats — including the messy and trick ones. Every change is regression-tested against the same set before it ships.
Stage 7 · Deliver
How we'd deliver it
We start with a proof-of-concept on your real WhatsApp enquiry chats, pilot it on one channel or catalogue with staff oversight, then take it live. A focused single-catalogue rollout is a matter of weeks.
Free consultation
Answering the same product questions all day on WhatsApp?
Tell us your catalogue size and enquiry volume and we'll tell you honestly whether a product enquiry agent is worth building — and what it would take.
Related use cases
Frequently asked questions
- How does it avoid quoting a wrong price or fake stock?
- It answers only from your live catalogue and stock through connected tools, and is constrained so it can't state a price, spec or availability it can't cite from that data. If it isn't sure, it asks or escalates rather than guessing.
- Does it handle Malay, English and Chinese — and mixed messages?
- Yes. It's built for how Malaysians actually type — Bahasa Malaysia, English, Chinese and 'rojak' mixes — and answers clearly in the same conversation.
- Does it take orders or payment?
- No — this agent answers questions and hands ready-to-buy leads to your team. If you also want it to capture the order end to end, that's our order processing agent, which pairs neatly with this one.
- What does it connect to?
- WhatsApp Business (or EzyChat), your product catalogue and spec content, and your live stock or inventory system — plus optionally your CRM so hot leads land where your sales team works. We connect it during the build.
Examples are based on real, anonymised engagements; details are generalised. Anchor Sprint is a member of the Anthropic Claude Partner Network — a deployment and rollout partner, not a reseller. This is general information, not legal or compliance advice.