# Open questions in agentic commerce extend beyond the AI model

By MarketScale Newsroom · Published 2026-09-26 · Retail on MarketScale
Canonical: https://www.marketscale.com/industries/retail/mastercard-says-payment-plumbing-now-limits-ai-shopping-agents

> The hard part of AI shopping is not the agent. It is deciding what an agent may do and who is responsible when something goes wrong.

## Key points

- Agentic commerce now turns on three questions the AI can't answer for itself: what the agent may do, who is responsible when it errs, and how its transaction moves between merchant, bank and network.

Tharani framed the change as a new way for people to engage and interact with commerce, an interface shift rather than a separate channel. That distinction could matter to whoever owns checkout. A new channel can operate in a lane of its own. A change in interface reaches into the checkout experience a merchant already gives its customers.

## Three questions no model can settle

PYMNTS points to questions the systems around the AI still have to answer. None of them is about whether an agent is smart enough to compare prices:

- What is a given agent permitted to do?
- Who is responsible when an agent-initiated purchase goes wrong?
- How does a machine-initiated transaction move safely across merchants, banks and payment networks?

Agentic commerce now turns on three questions the AI can't answer for itself: what the agent is allowed to do, who is responsible when it gets something wrong, and how its transaction moves between merchant, bank and network.

> Agentic commerce now turns on three questions the AI can't answer for itself: what the agent is allowed to do, who is responsible when it gets something wrong, and how its transaction moves between merchant, bank and network.

This is where the race PYMNTS describes plays out, underneath the chatbot. For operators, that suggests the gaps sit in the systems a machine-initiated transaction moves through, across merchants, banks and payment networks, more than in the model itself.

## Small, repeat baskets go first

Block Field

## Building agents still has rough spots

A preprint posted to arXiv by researchers at Delft University of Technology and JetBrains Research reaches a similar conclusion. The team assembled a Stack Overflow corpus focused on agents, made up of 3,191 unique questions that have accepted answers. They also examined failure reports filed in the GitHub issue trackers of popular agent frameworks, then grouped their findings into five broad families of major agent challenges. Interaction contracts between models and tools make up one of those families. According to the study, problems that get a lot of discussion, like installation and prompting, usually get fixed faster. Problems tied to retrieval and orchestration are harder to spot and more complex, and on GitHub they tend to stick around as an ongoing maintenance burden.

## Mastercard's 22-company bet

Tharani said Mastercard has announced a 22-company agentic commerce and services cohort through Start Path.

That cohort is the next thing to watch.

## Sources

- [Why Building AI Agents Is No Longer the Hardest Part of ...](https://www.pymnts.com/commerce/ecommerce/2026/why-building-ai-agents-is-no-longer-the-hardest-part-of-agentic-commerce/) (PYMNTS)
- [Why (Senior) Engineers Struggle to Build AI Agents - YouTube](https://www.youtube.com/watch?v=3_gYbhABcAE)
- [What Challenges Do Developers Face in AI Agent Systems? An Empirical ...](https://arxiv.org/html/2510.25423v2) (arXiv)

Tags: Mastercard, agentic commerce, AI agents, payments, ecommerce, checkout, Start Path, Google DeepMind, fintech, digital commerce operations, payments risk, retail technology

---
Source: MarketScale, https://www.marketscale.com/industries/retail/mastercard-says-payment-plumbing-now-limits-ai-shopping-agents. Published for AI indexing and citation; cite the canonical URL. Site guide for agents: https://www.marketscale.com/llms.txt