# eCommerce AI agent RFP — fill-in template

Operating artifact for the FactualMinds post [AI agent RFP template for eCommerce](https://www.factualminds.com/blog/ai-agent-rfp-template-ecommerce-2026/).

This is a structure a buyer fills in. It is not a scored vendor matrix, not a price, and not a client engagement.

Replace every `[bracket]`. Delete sections that do not apply. Do not leave "TBD" on security, writes, or acceptance.

## 1. Company context

- Legal name, storefronts, and regions: `[…]`
- Platforms (Shopify, Adobe Commerce, BigCommerce, custom): `[…]`
- Order volume and support volume you will share under NDA: `[…]`
- Named business owner and named technical owner: `[…]`
- Readiness score if you have run one (out of 30): `[…]`

## 2. Business objectives

State the job in operations language, not model language.

- Workflow in scope for this RFP (one): `[where-is-my-order | quote draft | reorder brief | …]`
- Workflows explicitly out of scope: `[refunds, price changes, purchase orders, …]`
- Success in 90 days means: `[a measured lookup with a handoff, not a revenue promise]`
- What a wrong answer costs: `[wrong tracking email | wrong price on a quote | …]`

## 3. Use cases

For each use case:

| Field | Your answer |
| --- | --- |
| Trigger | |
| Systems read | |
| Systems written | |
| Human approval | |
| Must escalate when | |

Week-one default if you have not decided: reads only. Writes wait.

## 4. Required integrations

Name the API, not the logo.

- Storefront Admin or REST surface and API version you will pin: `[…]`
- OMS, WMS, ERP, CRM, helpdesk: `[…]`
- Events you will subscribe to (webhooks) versus polls: `[…]`
- Sandbox availability and who provisions credentials: `[…]`

## 5. Security and compliance

- Data classes the agent may see (order id, email, payment last-4, full address): `[…]`
- Data classes forbidden: `[…]`
- Where inference runs (your AWS account, vendor tenant, both): `[…]`
- Retention of prompts, tool traces, and customer text: `[…]`
- Frameworks in scope (PCI DSS, SOC 2, HIPAA — only if they actually apply): `[…]`
- Requirement: prompt text is not an authorization boundary. Writes need a policy engine the bidder must name.

## 6. AI model requirements

- Who chooses the model, and can you swap it without a rewrite: `[…]`
- Region pinning: `[…]`
- Offline eval set the bidder must pass before production: `[count of golden tasks]`
- Forbidden: a bidder who will not name the model or the eval set.

## 7. Evaluation

- Golden tasks you will supply (minimum 10 lookups and 3 must-escalate cases): `[…]`
- Pass bar (example shape, not a result): exact match on order id and carrier status; escalate on the three must-escalate cases: `[…]`
- Regression: every prompt or tool change re-runs the set before release.

## 8. Observability

- Trace per turn: model, tools called, policy decision, latency, token or platform cost.
- Alarm when a write is attempted on a tool that should be absent.
- You can export traces. A vendor-only dashboard is not sufficient.

## 9. Implementation

- Environments: sandbox, then production.
- Identity: separate credentials for the agent, not a shared admin token.
- Idempotency on any write you later allow.
- Failure behavior when a downstream API times out (stop, do not guess).

## 10. SLA and support

- Hours, severity definitions, and who is paged: `[…]`
- What the bidder operates after go-live versus what your team operates: `[…]`
- Exit: you receive IaC, tool schemas, eval set, and runbooks. A prompt export alone is not exit.

## 11. Pricing

Ask for a split. Do not accept one blended "per conversation" number.

- Platform and model (consumption)
- Integration build (fixed scope)
- Evaluation and observability setup
- Human review time (your staff, not the bidder's)
- Monthly run cost at your stated session volume
- What is excluded

Published reference point, not a quote: a first-party FactualMinds benchmark in July 2026 put an AgentCore support silhouette near **$791 per month at 50,000 sessions** (platform plus a small model). That figure excludes integration, people, and your store's APIs. Bidders must show their own worksheet.

## 12. Ownership

- Who owns prompts, tool code, policies, and eval data after the engagement: `[you]`
- Subprocessors and model providers: `[…]`
- Training on your prompts or orders: default **no**

## 13. Acceptance criteria

The pilot is accepted only if all of these are true:

1. The named workflow answers from the record, with a citation the reviewer can open.
2. Forbidden writes are absent from the tool list, not merely discouraged in a prompt.
3. The golden set passes the bar you wrote in section 7.
4. A trace exists for a sampled turn, including a denied write.
5. A person still approves anything that moves money, inventory quantity, or a customer price.
6. You can name the monthly run-cost worksheet at your volume.

## 14. Response format

| Section | Page limit |
| --- | --- |
| Approach | 4 |
| Architecture diagram | 1 |
| Tool allow-list and deny-list | 2 |
| Security | 3 |
| Eval plan | 2 |
| Pricing worksheet | 2 |
| What you will not do | 1 |