AWS Quick Turns Chat Descriptions Into Workflows: Fintech Pomelo Saves 80 Hours a Month
A report that an analyst assembles from memory can't be repeated by colleagues, and during their vacation it has to be rebuilt from scratch. In AWS Quick, you describe such a process in chat, and the service builds a shared workflow: agents write drafts, and calculations run on fixed formulas. Fintech Pomelo says one such workflow saves it 80 hours a month, or about ten working days. It makes sense to start with a single recurring task, such as a report.
AI-processed from AWS Machine Learning Blog; edited by Hamidun News
A familiar picture: each month an analyst builds a report using their personal checklist, and when they go on vacation, a colleague has to start over, and the result looks different. AI assistants speed up individual people, but that experience rarely becomes shared. AWS offers a solution in the update to Quick, its AI workspace for enterprise tasks: you describe a process in chat, and the service turns it into a workflow (flow) that any employee can run.
As an example of the result, AWS points to Latin American fintech Pomelo, which issues cards and processes payments for its customers. According to the company, one such workflow saves it 80 hours a month, or roughly ten eight-hour working days. In addition, nearly four times as many employees now work with data there without any growth in headcount.
A process from chat instead of a manual builder
The main innovation is Quick Flows, which can now be built, run, and refined directly in a conversation. The user describes the process the way they would explain it to a new employee, and Quick assembles a workflow that combines agents, skills, tools, and data. Next to the chat, a diagram of the steps is displayed so you can check whether everything matches how the person actually works.
AWS demonstrates this with a hypothetical example involving financial manager Maya. Each month she prepares a report for the vice president on changes in costs and margins: she pulls results for each region, compares them with the plan in a spreadsheet, writes comments on the most notable deviations, and distributes the document. In the new workflow, the figures are taken from a dashboard, regions are processed in parallel, the comparison with the plan follows fixed formulas, and significant deviations are flagged against thresholds that Maya herself set.
An agent writes draft comments, and the final decision stays with a human. The system puts the finished Excel table and the email aside for review and tells the team in Slack that the draft is ready.
An important detail for engineers is "deterministic steps." These are ordinary verifiable calculations and checks that involve no language model: they produce the same result on every run. AWS recommends using them where creativity can't be allowed, for example in financial formulas. In addition, workflows can now run for longer and handle larger volumes of data.
While Maya is in a meeting, her colleague Jordan runs the same workflow from the desktop app, where he can use a local spreadsheet with adjustments. New hire Priya will get a report assembled in exactly the same way at her first month-end close.
The web app and a team of agents
The updated web app puts what needs attention on the home screen: a feed of emails and messages in priority order, plus a daily summary. You can ask Quick to draft replies and send them after review. Long tasks keep running after you close the tab, and results and questions from agents wait in the Mission Control section. Conversations, agents, and files sync across the web, desktop, and phone.
The second part of the release is multi-agent orchestration. You ask one agent a question, and it brings in several narrow specialists at once and merges their work into a single answer. An example from the release: why delivery costs rose in a region. A forecasting agent with planning data and a procurement agent with supplier contracts each answer for their part, and you can see each one's contribution. Each agent keeps its own instructions, knowledge, and context, so the team that knows the subject area best can maintain it.
Connectors, control, and what to do about it
Connectors to Meta ads and Amplitude have been added through their MCP servers. You can ask questions in natural language about ad campaigns and user behavior and combine the answers with other sources. Before an action that writes data, such as creating a campaign or changing a budget, Quick asks for permission, and access is limited to the rights of the specific person.
Administrators get three tools. Prompt guardrails (in preview) block prohibited topics and words, catch prompt-injection attacks, and hide selected types of personal data, and the policy can first be run in a detect-only mode. Permissions and limits are now set by identity-provider groups: for example, you can prevent the finance group from creating agents or cap the number of agent hours. Finally, you can choose which types of data sources employees are allowed to connect. These features are available on the Professional and Enterprise plans. The interface has also gained Japanese, German, Chinese, and Vietnamese.
Among the hard numbers in the release, there are only Pomelo's figures: 80 hours a month on one workflow and an almost fourfold increase in the number of people working with data. This is the customer's self-assessment in AWS's marketing text, with no independent verification. AWS explicitly calls the Maya example fictional.
The practical takeaway for teams considering enterprise agents: the value here lies in locking a process down, not in yet another chatbot. Pick one recurring task, such as a report or a reconciliation, move the calculations into fixed steps, and leave the text and analysis to agents with mandatory human review.
- Nov 21, 2026OpenAI launches ChatGPT Wallet: agents buy without ever seeing your card number
- Nov 21, 2026US Advertisers Spent $70.7 Million on ChatGPT Ads in October, Up 127% From April
- Nov 21, 2026ONNX Runtime vanishes from GitHub: Microsoft hides repositories after package hack
- Nov 21, 2026OpenAI Launches ChatGPT Wallet: Agents Pay by Card, Approved With Face ID
Key Questions Answered
What are the key details on A process from chat instead of a manual builder?
The main innovation is Quick Flows, which can now be built, run, and refined directly in a conversation. The user describes the process the way they would explain it to a new employee, and Quick assembles a workflow that combines agents, skills, tools, and data. Next to the chat, a diagram of the steps is displayed so you can check whether everything matches how the person actually works. AWS demonstrates this with a hypothetical example involving financial manager Maya.
What are the key details on The web app and a team of agents?
The updated web app puts what needs attention on the home screen: a feed of emails and messages in priority order, plus a daily summary. You can ask Quick to draft replies and send them after review. Long tasks keep running after you close the tab, and results and questions from agents wait in the Mission Control section. Conversations, agents, and files sync across the web, desktop, and phone. The second part of the release is multi-agent orchestration.
What are the key details on Connectors, control, and what to do about it?
Connectors to Meta ads and Amplitude have been added through their MCP servers. You can ask questions in natural language about ad campaigns and user behavior and combine the answers with other sources. Before an action that writes data, such as creating a campaign or changing a budget, Quick asks for permission, and access is limited to the rights of the specific person. Administrators get three tools.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.