Stop Calling Workflows ‘Agents’: The Speech-to-Speech Future of E-Commerce.

That’s systems that can:

•Hold a goal (“complete this order within policy”).

•Re-plan when payments fail, stock changes, or customers push back.

•Choose tools dynamically (payment rails, loyalty points, shipping options).

•Operate over speech-to-speech, chat, or email — not force customers into yet another app.

•Finish the job without human micro-orchestration.

Most of what’s on the market today?
Level 1 assistants dressed up as Level 3 agents.

E-commerce buyers should start demanding real metrics:

•Unattended Completion Rate (UCR)

•Obstacle Recovery Rate (ORR)

•Mean Time to Human (MTTH)

•Policy Breach Rate (PBR)

Because the paradigm shift isn’t “AI in the checkout.” It’s speech-to-speech agents that listen, act, adapt, and transact across your stack without hand-holding.
Stop rebranding workflows. Build agents. Or sell what you have, proudly, as what it really is.

For e-commerce, that means it can:

1. Hold a goal (e.g., “Get the customer through checkout with the best payment option within their preferences.”)

2. Form and revise a plan over multiple steps.

3. Select and compose tools dynamically (payment rails, loyalty APIs, shipping, fraud checks, speech interfaces) without a human telling it which, when, or how.

4. Act in external systems, observe results, and adapt when reality changes.

5. Handle obstacles (card declined, OOO shipping provider, customer pushback) by re-planning or negotiating alternatives.

6. Operate via existing channels — in this context, speech-to-speech, chat, or customer’s chosen interface, not forcing them into a bespoke app.

7. Respect policy and compliance via guardrails (KYC/AML, refund windows, PCI DSS) and auditable logs.

8. Finish the job (or stop safely) without human micro-orchestration.

If any of these are missing, we’re not in ‘agent’ territory.

❌ Not if/then checkout flows. Pre-baked branches are brittle. Agents plan, act, observe, and re-plan.

❌ Not just voice commands in a custom app. If the user still drives every step, it’s assistive, not autonomous.

❌ Not a copilot that suggests upsells you must approve. That’s assistive AI, not autonomy.

❌ Not integration sprawl. “We connect to 50 APIs” ≠ autonomy.

❌ Not locked behind a vendor dashboard. Real agents meet customers where they already are: voice, chat, email, or device-native.

1. Workflow Wrappers
Checkout flows with a few AI prompts glued in. Looks slick in a demo, breaks on the first edge case (customer wants to split tender across BNPL + loyalty points).

2. Integration Theatre
“We’re agentic, we integrate with 73 gateways.” But the system still needs the shopper to pick manually. That’s remote control, not autonomy.

3. Wizard Cosplay
A five-step conversational UI that asks you everything the agent should infer (e.g., “Which shipping do you prefer?”). If the human drives the path, it’s not autonomous.

4. Play-Acting Copilots
Voice assistants that suggest products but can’t handle objections, negotiate shipping, or re-route a failed payment. That’s voice search, not an agent.

5. LLM-as-Form-Filler
Auto-completes checkout fields, but can’t adapt when the customer changes delivery address mid-flow.

Level 0 – Automated Workflow: Pre-built checkout funnels. Reliable, brittle, cheap.

Level 1 – Assistive AI: Drafts product suggestions, classifies queries. Human/customer drives process.

Level 2 – Supervised Agent: Plans and acts across tools (e.g., selects payment rail, checks loyalty balance) but requests approval for exceptions.

Level 3 – Constrained Autonomy: Operates unattended within policy/risk bounds (e.g., retries failed payments via alternative rails, escalates only if risk triggers).

Most “agentic” commerce products today are Level 1 masquerading as Level 3.

•Goal & Plan Loop: Explicit planner that updates based on outcomes (not prompts alone).

•Tool Autonomy: Dynamic selection of rails, wallets, shipping, or loyalty.

•Obstacle Recovery: Detects blockers (card decline, fraud flag, OOO shipping) and replans or escalates.

•Policy Guardrails: Approval thresholds, refund rules, tax constraints enforced at run-time.

•Auditability: Immutable logs of what happened and why.

•Channel-Native Operation: Works over speech-to-speech, chat, or device OS APIs.

•Stop Conditions: Risk triggers, timeouts, retry ceilings.

•If a human must click “next” at every step, it’s not agentic.

A real speech-to-speech agent will:

•Parse the customer’s voice order: “Order two pairs of size 9 sneakers for pickup this Friday.”

•Cross-check stock, confirm pickup availability, log rationale.

•Negotiate payment method: if card fails, offer BNPL, loyalty points, or PayID.

•If shipping delay arises, re-plan: offer express courier, suggest alternate pickup, or reschedule.

•Complete payment, update CRM/order tracker, send receipt.

•Close the loop with audit trail and notification in the customer’s channel of choice (voice + email).

•A workflow wrapper will: capture the order, hit a card gateway, and stall on decline.

•Unattended Completion Rate (UCR): % of orders completed without human touch.

•Obstacle Recovery Rate (ORR): % of failures resolved autonomously.

•Mean Time to Human (MTTH): How long the system runs before intervention.

•Policy Breach Rate (PBR): Out-of-bounds attempts per 1,000 orders.

Then run a black-box test: give the agent a real product catalog, payment stack, and customer queries. No demo script. See what survives.

•Speech-to-speech lets customers say “Buy me a hoodie, size M, same address as last time” and the agent just does it.

•Agents don’t collapse on ambiguity — if the customer says “Actually make it two”, they re-plan.

•They operate across payment rails — voice triggers PayID, loyalty redemption, or crypto settlement without the shopper touching a button.

If your ‘agent’ only works inside your custom voice app, it’s not an agent. It’s an app with speech recognition.

Calling everything ‘agentic’ misleads retailers, investors, and customers. They expect autonomy but get scripted workflows. That discredits the field.

Language shapes budgets. If vendors blur ‘assistive’, ‘automated’, and ‘agentic’, businesses will buy the wrong thing, measure the wrong outcomes, and conclude “AI can’t do this.” It can — if we build the right systems.

Be honest about the level. If you’re L1/L2, say so. Copilots and smart workflows have huge value.

Start with bounded domains. High-volume, low-risk use cases (cart recovery, loyalty redemption, subscription renewals).

Engineer guardrails properly. Policy engines, safe payment schemas, anomaly monitors. Not “approve every step.”

Publish metrics. If you can’t show UCR/ORR/MTTH/PBR, don’t claim autonomy.

Put channels first. Voice, chat, SMS. Dashboards later.

The future of commerce isn’t “software with AI inside.”

It’s agents that listen, speak, and act across your stack to get outcomes. Different bones. Different responsibilities.

We don’t ‘use’ them — we task them, constrain them, and audit them.

Stop rebranding voice workflows. Build real agents. Or sell what you have — proudly — as what it is.