Shopping assistant
Answers plain-language questions and recommends only products that exist in your catalog, at the right price and stock.
We publish: groundedness, price and stock correctness
Indexwright Commerce Agent Lab
Commerce agents, proven before they go live.
We build AI agents for e-commerce: shopping assistants, support and returns, catalog enrichment. Every agent is replayed against your real shopper questions and scored before a customer ever sees it.
Free snapshot · No login required · Your real shopper questions · Delivered in 2 business days
Prefer to talk it through? Book a 15-minute call →Is this for you?
- 20,000+ products
- Search or an agent that matters to revenue
- Shoppers who ask in plain words
- Real queries or support tickets to replay
- Meaningful monthly search or AI cost
- An engineering contact
A good demo is not a working agent.
Most agents look right on ten hand-picked questions. Real shoppers ask thousands of different ones, and one wrong price or invented product costs trust that is hard to win back. Industry roundups report support-agent hallucination rates of 15 to 27 percent and production resolution of 55 to 70 percent against 85 percent or more in vendor demos (source).
We close that gap by testing before launch, not after the first complaint.
Does it answer only from your catalog?
Does it quote the right price and stock?
Does it stay inside your refund policy?
What does each conversation cost?
Agents for every stage of the shopper journey.
Some agents talk to shoppers. Some work for your merchandising and operations team. Some prepare your store for the AI assistants that shop on other people's behalf. All of them sit on the same search layer and the same test harness.
Plate 01 · Commerce agent mapEleven agents, four launch offers
Support and returns
Handles order status, returns and refunds inside your policy, and hands off to a person when it is unsure.
We publish: resolution rate, policy compliance
Catalog enrichment
Fixes and fills titles, attributes and categories at scale, so search and every agent have clean data to work from.
We publish: attribute completeness, error rate
Agent readiness
Makes your products readable and buyable for the AI assistants shoppers already use, through structured data and standard endpoints.
We publish: share of catalog agent-readable
We try to break the agent first.
Anyone can show an agent demo. We run each one through the same loop: replay your real questions, check every answer against your catalog and policy, score it, run it beside your team, and only then launch it with a rollback ready. Every failure becomes a permanent test.
Plate 02 · The lab loopSix steps, one feedback path
How an agent answers, and where we check it.
This is the reference design behind every agent we build. The shopper never talks to a model directly. Questions pass input checks, the agent calls approved tools, and every answer is verified before it is sent.
Plate 03 · Reference architectureOne agent, end to end
Every answer is checked
An output check compares each reply with the catalog, live stock and your policy before the shopper sees it.
Search is one tool of four
The agent retrieves real products first, then writes. That is why a weak search layer caps how good any agent can be.
People stay in the loop
Low-confidence conversations go to a person with the full context, and every trace is kept for replay.
We publish what passed and what failed.
Each client gets scorecards in this format. A failure is the useful result: it shows what the test caught before a customer did.
Specimen A-0042
Pass
"waterproof jacket for hiking, under $150"
Agent: Two jackets fit: the Ridge Shell ($129, in stock) and the Trail Rain Jacket ($145, in stock). Both are rated waterproof.
- Answers from catalog
- Pass
- Price within budget
- Pass
- Stock correct
- Pass
- Cost per conversation
- Measured
Specimen C-0007
Pass
Catalog enrichment · 500-item audited sample
Before: "Blue Tee M" with no material, fit or care fields.
After: "Men's Short-Sleeve Crew Tee, Cotton, Regular Fit, Blue, M" with every field filled and checked.
- Attribute completeness
- Measured vs. human audit
- Invented attributes
- 0 allowed
- Human review
- Sampled
Specimen S-0014
Fail
"I bought this 45 days ago, can I still return it?"
Policy says
Returns accepted within 30 days of delivery.
Agent said
Yes, you can return it. I've started the return for you.
What the check caught
Policy violation and an action taken outside limits. Launch blocked.
Action: add the return window to the policy check → replay → score again
Illustrative specimens that show the scorecard format. Client results are published with permission, anonymized.
Every good agent starts with good search.
An agent that talks is easy to build. One that only sells what you stock needs your catalog turned into searchable meaning, a retrieval step that finds real products, and answers written only from what it finds. That is why we fix search first and add the agent on top.
- VectorizeProduct titles become vectors.
- RetrieveThe agent searches your real catalog.
- AnswerIt answers using only catalog data.
- RememberIt remembers the conversation.
Explore the Search Lab Where search cost comes from Seven levers to cut it How the alternatives compare Migrating without risk
Based on AWS's published reference architecture for an AI shopping agent. We still migrate and tune search on its own when that is the right answer.
Autonomy is earned, one level at a time.
You decide how much the agent may do on its own. It moves up a level only after it passes scored tests on your real traffic and a clean shadow run. Most teams launch at level 2 or 3.
Plate 04 · Autonomy ladderFour levels
SuggestsThe agent proposes a reply. A person writes and sends it.
DraftsThe agent writes the full reply or action. A person approves each one.
Acts within boundsHandles routine cases alone, such as order status. Refunds above a set limit wait for approval.
Acts on its ownCloses whole categories end to end, with sampled review and instant rollback.
Start with evidence.
Agent Snapshot
Free
A focused review of how an agent would read and answer from your store.
- 25–40 real shopper questions
- Grounding, price and stock checks
- Catalog readability score
- Search relevance and zero-result review
- 3-minute recorded walkthrough
Agent Diagnostic
$1,500–$3,000
A deeper technical and economic investigation of your search and the agents worth building.
- Query and ticket replay
- Architecture and data review
- Cost per conversation model
- Build or don't-build recommendation
Agent and Search Build
$5,000–$15,000
Build the agent, replace the engine, or both. If neither is justified, we find savings in what you already run.
- Replay, scoring and shadow run
- Acceptance criteria agreed up front
- Agents tested on real queries before launch
- Managed cutover with rollback
Managed Agents and Search
$500–$1,500 / mo
Ongoing monitoring, relevance testing, regression replay and optimization.
Agent monitoring: contact for pricing
Three founding partner places are open. We are establishing the Indexwright search and agent benchmark with a small number of teams. Founding partners get preferred pricing in exchange for permission to publish anonymized results. You get the investigation. We get the evidence.
Free Agent Snapshot
See how an agent would answer your shoppers.
Send us your store URL. We replay a sample of real shopper questions against your catalog as an agent would read it, and show you where answers would break.
Your snapshot includes
- 25–40 shopper questions — representative searches and plain-language asks across your catalog.
- Grounding checks — would the answer come from products you actually stock, at the right price?
- Catalog readability — are titles, attributes and structured data complete enough for an agent?
- Search review — relevance, typos, SKUs, zero-result queries and ranking.
- Agent-fit review — which queries need a conversation, not a keyword match.
You receive a 3-minute recorded walkthrough of the findings and the highest-impact opportunities. Delivered within 2 business days.
Questions
Both. We build the agent and test it on your real shopper questions before it goes live. We can also test an agent someone else built.
It can only answer from what the catalog search returns. An output check compares each answer to the catalog and your policies before the shopper sees it. If nothing matches, the agent says so.
Often no. Many search engines can store the vectors an agent needs. The Diagnostic confirms what fits yours.
Only as far as you allow. Agents launch at a set level of autonomy, from suggesting replies to acting within limits, and move up only after passing scored tests on your real traffic.
The Agent Snapshot does not require production access or API credentials. A Diagnostic or build may require technical access depending on the architecture.
No. Search is the foundation layer under every agent, and we work with whatever engine you run. We also still migrate search when the evidence supports it.
We'll tell you. Sometimes better search or cleaner catalog data is worth more than an agent.
Your agent shouldn't be a black box.
Find out what your shoppers would be told. Find out what it would cost per conversation. And find out whether an agent is ready to go live, before it meets a customer.
Get Your Free Agent Snapshot25–40 questions · Grounding checks · Search review · 3-minute walkthrough · 2 business days