A Small Service That Connects Internal Information to Approved Action

A question is now followed by action

On January 29, 2026, OpenAI introduced a data-analysis tool used inside the company. It read table structures, past analysis records, and internal materials such as Slack and Google Docs to answer questions. At that point, it was an internal OpenAI tool, not a product sold to external customers.

On September 10, 2026, it released similar capabilities as a public product called Data agent in ChatGPT Work. Employees can ask questions in everyday language, search and analyse company information, and create interactive dashboards. Within permitted connections, they can also share analysis results or pass them to the next task.

The important change is that it inherits a company’s existing viewing permissions. According to OpenAI, AI cannot see collections of information, specific records, or individual items that the user could not already access. For this control to work, though, the original information must already be separated by employee. If everyone in a company can see every document, the AI inherits those broad permissions too.

Reading information and changing it are also treated differently. OpenAI’s enterprise guidance says the system is set to ask the user again by default before taking actions such as sending an email or editing a record. OpenAI also says that work information accessed through enterprise plans is not used by default to train AI models.

OpenAI said that almost all of its product organisation and more than two-thirds of its sales and go-to-market organisation use the tool for data analysis. This is the company’s own statement, so the depth of use and results need separate verification. OpenAI also warns that it cannot fully eliminate the risk that instructions hidden in documents could cause unexpected information access or actions.

If you run a 12-person window-installation business

Imagine the owner of a window-installation business with 12 employees. For one construction job, the owner copies a customer address, window specifications, promised date, and amount into a quote, schedule, materials order, and settlement sheet four times. Site photos sit in an employee group chat, while change requests arrive through calls or text messages.

Every Monday morning, the owner asks about the status of each job and checks again whether materials have been ordered. If the quoted amount differs from the actual order amount, the owner asks staff at which stage it changed. Even when a customer agreed to extra work, the owner may delay billing or absorb the loss if the record cannot be found.

This company does not need a huge AI system that takes over every task. It needs a small service that reads quotes, schedules, order records, and site photos, then gathers only: “jobs scheduled for installation this week with no materials order,” “jobs where the order amount exceeds the quote,” and “extra work without a record of customer consent.” Instead of searching through everything, the owner only reviews the few exceptions.

Add approval-request creation, and the work shrinks another step. The service groups the text messages and photos that support an extra charge, then drafts a revised quote for the customer and an approval request for the owner. Before the owner approves, it neither sends anything nor changes an amount.

After approval, only defined follow-up actions continue. For example, the system can reflect the approved change amount in the settlement sheet, send a materials-order confirmation request to the person responsible, and record the processing time and approver. Because mistakes can change money or schedules, each company should set its own boundaries for automatic action.

Some work still belongs in human hands. Measuring actual dimensions in an old building, interpreting an ambiguous customer request, and deciding who is responsible for a defect all require field experience and relationships. AI can gather evidence and flag missing items, but it does not take on that responsibility.

This is where the hard-to-copy advantage becomes visible. Other companies can build screens and document-summary features, but they cannot easily copy historical records that show which photos support extra work, each supplier’s delivery habits, the owner’s trusted approval criteria, or sales channels based on introductions from people in the trade. More durable advantages remain in organised data, commercial relationships, trust in outcomes, and distribution rather than technology alone.

Places already using human approval

Riyad Capital in Saudi Arabia used UiPath to reduce the process of reading and checking customer documents for investment-account opening and regulatory requests. When information in a document was unclear or a decision was required under regulations, the case was passed to a human review screen. UiPath charges enterprise subscription fees based on execution tools, user permissions, and document-processing scope, while fixed contract prices are not public.

According to the customer case published by UiPath, customer-request processing volume increased by about 20 times, while the monthly manual verification time required for roughly 76,500 regulatory requests fell from 1,024 hours to about 73 hours. Because these figures were published by the supplier and customer, they do not mean other organisations will achieve the same results.

In the United States, Salesforce applied Agentforce HR Service to employee questions about benefits, expense reimbursement, leave, and compensation. It reads company policies and HR information to prepare answers, asks managers to approve qualifying expense requests, and hands sensitive or complex issues to HR staff. Per-user monthly pricing depends on the number of users and selected product tier, and is billed annually.

Salesforce said its benefits assistant handled more than 1,000 conversations in its first seven business days and reduced responses that had taken days to seconds. Its manager-facing compensation assistant handled about 3,000 enquiries, while HR case volume fell 50% from the previous year. These are also figures Salesforce published from its own internal operations.

Small things you could build from here

A construction-change approval inbox. This service groups quotes, site photos, and customer messages to show the evidence for extra work and the changed amount, then updates the settlement sheet after approval. It is for window-installation, wallpapering, or bathroom-renovation businesses with around 12 employees. It can start with one approval flow by connecting only documents the company already has, making it practical to build now.

The first screen shows three lists: “Awaiting approval,” “Insufficient evidence,” and “Unprocessed after approval.” A growing archive of past construction photos and records of recognised extra charges, plus relationships with materials suppliers and installation teams, becomes a hard-to-copy asset.

A shared purchasing check for three shops. This service compares sales volume, waste records, and supplier orders for each shop, then shows the owner only orders that are much larger than usual. It is for operators of three neighbourhood side-dish shops that have newly started delivery sales. It can combine anomaly detection and approval-request creation in one flow, making it possible to start now.

The first screen should show not a recommended quantity by item, but “why this order differs from normal.” As records build around each shop’s promotions, weather-related sales changes, supplier lead times, and the owner’s decisions, the service becomes harder to copy than a basic ordering app.

A grant-programme evidence checker. This service connects spending records, receipts, contracts, and activity records to find missing evidence. After confirmation from the person responsible and the organisation’s representative, it drafts a report. It is for the accounting staff member at a six-person nonprofit delivering a support programme run by a local government or foundation. It can keep information checks and approval history in one place, making it feasible to try now.

The first screen shows “No evidence,” “Amount mismatch,” and “Awaiting representative approval,” rather than the remaining budget for each programme. Rejection reasons from each funding body, the accounting staff member’s correction experience, and referral relationships with tax and accounting specialists become the service’s defensibility.

A defective-parts shipment hold. This service reads customer complaints, inspection results, and production records to gather risks in the same production batch, then requests shipment-hold approval from the quality manager. It is for a quality manager at an industrial-parts manufacturer with around 20 employees. AI handles record linking, while a person decides whether to stop actual shipments, making the division of responsibility clear.

The first screen shows potentially defective production batches, linked customer orders, and actions that will not run before approval. As company-specific defect categories, customer-specific response commitments, and evidence trusted by the quality manager accumulate, it becomes distinct from a general-purpose document-summary tool.

What to check today

Call one business owner you know and ask them to explain one approval task from last week, from start to finish. If they copy the same information into three or more places and processing stops at least once a week while waiting for an approver, it may be worth building something small. Define the first product not as the entire workflow, but as one screen that gathers evidence before approval.

Why this matters where you are

Look for a recurring approval task in your own market where information is scattered across documents, messages, and photos. The records, approval rules, and permitted automatic actions will differ by organisation and local context. You can begin by identifying the exceptions a person repeatedly checks, then make the evidence visible before any action is taken.