AI Pricing Intelligence Agent: Should You Change This Product's Price? (2026)

This post is should you change this product’s price? — one SKU (or parent), named inputs, a recommendation: Maintain / Increase / Decrease / Promote / Bundle / Clear. It is not personalized offers (post 16: who gets a discount class). It is not dynamic merchandising (post 17: what to feature today). Live price is a write. Week-one default: Recommend only (autonomy spectrum).

On June 17, 2026, AgentCore Harness reached general availability (What’s New). A managed loop can recommend a price change. It does not make live PDP mutation week-one work. After July 30, 2026, do not start a repricer on Agents Classic.

AWS lifecycle notice (June 30, 2026) — Amazon Bedrock Agents Classic is in maintenance for new customers after July 30, 2026. Net-new pricing agents should use Bedrock AgentCore. Full matrix: lifecycle roundup.

It is not a client engagement. We do not publish a price-lift, a GM$ from repricing, or a competitor index. Thresholds you add to the artifact are yours. We do not invent a competitor scrape.

First-party signals we reuse (not eCommerce outcomes) — Gateway server-side tools cut median tool round-trip ~180 ms → ~95 ms on a B2B CRM assistant (12 tools, ~8k turns/day) — Gateway post. Platform TCO silhouette: support-style AgentCore at 50K sessions/mo ~$791/mo platform + model (decision guide). Model your mix on the AgentCore pricing calculator. Treat ~$791/mo as a platform cost floor to plan against, not as margin a price change will produce.

Reproduce this — Copy pricing-recommendation-governance.md. Fill Owner and lawful signal sources. If you cannot name a legal market feed, leave that row empty. Series folder: ecommerce-ai-agents-series/. Ship gates: monday-checklist.md.

This is series part 52. Contribution inputs come from margin intelligence. If cost is unknown, you may not claim a “profitable price.” Excess markdown order (stop-buy first) stays in post 23 — Clear here still needs that discipline plus a pricing owner.

Opinionated take: recommend a named action, then a human (and Cedar) writes the catalog. Trade-off: list price is still slow. The alternative is an agent with updatePrice and an unlicensed scrape in Memory.

FactualMinds is an AWS Select Tier Consulting Partner. We help merchants sequence agents — we do not sell a guaranteed repricer.

List price is not an offer and not a homepage

Merch wants conversion. Finance wants contribution. Legal wants MAP. Marketplace teams want channel parity. Shoppers want a bargain in chat. If you let one specialist do all of that, you get a coupon on a featured stockout.

Why this matters / who breaks without it: finance breaks when Decrease ignores cost. Brand breaks when Increase fights a live campaign. Legal breaks when MAP SKUs move without the MAP tool. Paid media breaks when Clear fires while ads still bid. The agent is useful when demand, ATP, margin, promo, and (optionally) lawful market signals live in different systems.

If the promo engine already schedules a price with a human calendar, keep that rule. The agent recommends what changed (cover collapsed, cost changed, promo ended) — it does not become a second engine.

Baymard 70.22% is unexpected cost at checkout, not permission to cut list price. Do not put it on a pricing-ROI slide.

How stores handle this today

A seasonal price list. A MAP spreadsheet. A marketplace repricer on Amazon that nobody wired to DTC. Shopify compare-at. A merch Slack “take 10% off.” Those are rules or habits. Keep MAP and engine calendars. Kill hallway percents.

There is no native Shopify AgentCore connector. Admin price is one OpenAPI target. Magento, custom catalogs, and ERP price books look the same at Gateway.

Inputs (lawful, named)

From the artifact:

If cost is missing, you may still recommend Maintain or hold-for-cost. You may not label Decrease “margin-safe.” That is post 49 again: unknown is not a GM%.

Recommendations and who approves

Veto:

• Agent invents a % off in chat
• Competitor price from an unlicensed scrape
• Execute on the storefront without Cedar + HITL
• Change MAP-protected SKUs without the MAP tool

Promote is not post 16’s pct_5 shopper class. Promote is “this SKU should enter a named engine campaign.” The engine is still the system of record. Bundle is not post 17 homepage publish — merch still publishes.

flowchart TD
  sku[This SKU]
  inputs[Cost ATP demand promo lawful market]
  rec[Recommend Maintain Increase Decrease Promote Bundle Clear]
  veto[Veto scrape invented percent MAP miss]
  owner[Pricing merch finance HITL]
  cedar[Cedar on updatePrice]
  sku --> inputs
  inputs --> rec
  rec --> veto
  veto --> owner
  owner --> cedar

SKU → named inputs → recommendation → veto → owner → Cedar. Detection of MAP and live promo should be tools. Using the model as a scraper is how you buy a lawsuit and a bad price.

Autonomy and money

Per autonomy spectrum: Observe and Analyze on reads; Recommend for week one; Draft a price-book row if your PIM allows; Request Approval before any catalog write; Execute only when Cedar is in ENFORCE, evals include must-not-write cases, and HITL exists for over-threshold SKUs. Fully Automated list price is not a 2026 week-one setting for this series.

HITL timeout fails closed. Do not auto-publish because merch is in a meeting (HITL).

Production shape

Bedrock is the model. Harness (GA June 17, 2026) or Runtime is the host. Gateway + Cedar default-deny updatePrice, updateCompareAt, createDiscount. Strands does not give you Policy. Browser off — Browser-on-competitor-sites is the scrape you just vetoed. Next.js can be the pricing approval UI — not the runtime.

Absolute latency will be PIM and ATP, not Gateway. ~180 ms → ~95 ms is the CRM canary. A morning recommend job is cheap; a chatty “make it cheaper” copilot on storefront is how you mint percents. Model the mix on the AgentCore pricing calculator. If volume would not notice ~$791/mo at 50K sessions, you are staffing a demo.

Context: Python 3.12+ for stubs; Harness or Runtime; LOG_ONLY then ENFORCE.

# From pricing-recommendation-governance.md — fill owners for your store
Maintain     → no write
Increase     → pricing owner; MAP tool if protected
Decrease     → pricing owner; MAP check; cost unknown ⇒ cannot claim margin-safe
Promote      → engine issues code; agent does not mint
Bundle       → merch owner
Clear        → finance + inventory; stop-buy discipline from excess post

Week-one evals: unlicensed market URL in tools must fail; invented % in output fails; MAP SKU without MAP tool cannot Decrease; updatePrice on this token DENY; cost-null cannot emit “profitable to cut.”

What broke

What broke — Sample harness with Browser enabled “to check competitors.” The model quoted a marketplace price from a page it was not licensed to collect, recommended −15% on a MAP SKU, and called updatePrice because the prompt said “be competitive.” Detection: Observability showed Browser + updatePrice in one turn; MAP tool was never called; Policy LOG_ONLY. Fix: Browser off; market signals only from a named lawful feed or omit; Cedar DENY updatePrice; MAP tool mandatory for protected SKUs; Recommend only. Lesson: competitive is not a scrape. Prompt text is not authorization.

A second counter-case: merging pricing with offers. Shopper asked for 20%; the specialist cut list price instead of returning human_approval on an offer class. Fix: hop cap; offers cannot see updatePrice; pricing cannot see issueDiscount.

How this relates to margin, offers, merch, and autonomy

• Margin intelligence: whether contribution inputs exist. Pricing consumes them; it does not invent COGS.
• Personalized offers: who gets a class. Not list price.
• Dynamic merchandising: what to feature. Not updatePrice.
• Autonomy: per-action Recommend vs Execute. Live price is not Fully Automated because WISMO notify worked.

What This Post Doesn’t Cover

• Offer classes (no_discount / bundle / pct_5) — post 16
• Homepage and collection publish — post 17
• Excess action order (stop-buy before markdown) — post 23
• A measured A/B lift or competitor index from a FactualMinds engagement
• Building or running an unlicensed scraper
• Marketplace buy-box algorithms as first-party tools
• Payment capture / x402
• A native Shopify AgentCore price app (it does not exist)
• Classic Agents cutover

What to Do This Week

1. Name the pricing owner. If you cannot, do not attach updatePrice — ever.
2. Clone pricing-recommendation-governance.md. List lawful inputs only. Leave market signals empty if unsure.
3. Confirm margin contract for any SKU you want to call margin-safe.
4. Read tools: cost, ATP asOf, demand, promo flags, MAP. Browser off. No scrape.
5. Harness (GA June 17, 2026) or Runtime; Gateway; Policy LOG_ONLY. Prove DENY on updatePrice.
6. Goldens: MAP without tool fails; cost unknown cannot claim profitable Decrease; invented % fails; Promote does not mint a code.
7. HITL: session id + trace. Timeout fails closed. Autonomy row = Recommend.
8. Price the job on the AgentCore pricing calculator. Need governance on list price before any copilot? Contact us. Amazon Bedrock, Generative AI on AWS, AWS for retail / eCommerce.

Full ship list: monday-checklist.md.

If You Only Do One Thing

Keep live price at Recommend, put Cedar on updatePrice, and omit any market signal you cannot name as lawful. Demand and ATP you own. Scrapes you do not.

Need a pricing recommender that does not write the catalog? Contact FactualMinds or start from margin intelligence.

FAQ

When should we NOT let an AI agent change live price?

Week one: never Execute. Skip even the brief when you have no costing, no ATP asOf, no named pricing owner, and no MAP tool for protected SKUs. Also skip it if the only market signal is an unlicensed competitor scrape. If a scheduled price rule already runs in the promo engine, leave that rule. Skip Agents Classic after July 30, 2026.

What could go wrong if the agent invents a percent off in chat?

You stack on a live sale, cut below contribution, break MAP, and train shoppers to bargain. Promote means the promo engine issues the code. The agent does not mint a string. Decrease is a list or scheduled change a pricing owner approves — not a hallway coupon.

When should we NOT use competitor prices as an input?

When you cannot name a lawful source and a policy owner. Do not invent a scrape. Do not paste a screenshot into Memory. Optional market signals are only those you are allowed to use. Demand from your own sales is required; competitor price is not.

What could go wrong if we merge this with personalized offers or merchandising?

Post 16 classifies who may get an offer class. Post 17 recommends what to feature today. This post asks whether this SKU list price should move. A merged specialist will coupon a MAP SKU, feature a stockout, and write the PDP price in one turn. Separate tools, separate autonomy rows, separate owners. offers. merchandising.

Should pricing intelligence run on Harness or Runtime?

Harness fits a scheduled recommend job with read tools and a JSON schema. Use Runtime plus Strands if pricing, merch, and offers specialists must not share updatePrice. Bedrock is the model. Strands does not provide Gateway, Identity, or Cedar. Live price stays Recommend or Request Approval on the autonomy spectrum until Policy is boring.

Is there a native Shopify AgentCore connector for Shopify Prices?

No. Price, compare-at, and marketplace price APIs are your OpenAPI or MCP targets. Do not wait for a native connector. Do not put payment capture in pricing tools. HITL on every live price write.