It is Monday morning. You open the sales report and revenue is down 14% on last month. You know the number. What you do not know is why. So you spend the next two hours opening spreadsheets, comparing weeks, asking the shop supervisor whether anything changed, and eventually deciding it was probably the rain.
A data agent is software built to do those two hours for you, and to do them before you ask. In September 2026 the idea stopped being a research demo and started appearing inside tools business owners already pay for. Here is what a data agent actually is, how it differs from the dashboard you already ignore, and whether a Hong Kong SME has any use for one yet.
What Is a Data Agent?
A data agent is an AI system that connects to your business data, investigates a question on its own, and returns a finished explanation. Unlike a chatbot, it plans several steps, queries the data itself, and reports what changed and why, without anyone writing a database query or building a chart.
The word that matters in that definition is investigates. A search tool retrieves. A chatbot answers what you asked. A data agent decides what to check next based on what it just found, the way a junior analyst would.
Ask it "why did revenue drop last month" and it does not hand back a number. It breaks the question apart: was it fewer customers, or the same customers spending less? Which branch? Which product? Which week? It runs each check, discards the dead ends, and comes back with a short answer and the evidence behind it.
How Is a Data Agent Different From a Dashboard?
A dashboard displays numbers you chose in advance and waits for you to interpret them. A data agent starts from a business question, works out which numbers are relevant, and explains the cause. The dashboard tells you revenue fell 14%. The agent tells you it fell because Tuesday lunch covers dropped at one branch.
This is the difference between reporting and diagnosis, and it is why so many dashboards go unopened after month three.
Most small businesses do not lack data. They lack the twenty minutes of disciplined digging that turns a number into a decision. A dashboard hands you the number and stops. Somebody still has to do the digging, and in a 15-person company that somebody is usually the owner, at 11pm.
Three things a dashboard cannot do
- － It cannot ask a follow-up question of itself. A chart showing a drop will never go and check whether the drop was concentrated in one product line.
- － It cannot join data that lives in two systems. Your POS and your booking system each have half the answer; the dashboard shows whichever half it was built on.
- － It cannot tell you that nothing interesting happened. A dashboard renders the same twelve tiles whether the month was normal or a disaster.
How Does a Data Agent Actually Work?
A data agent works in four steps: it connects to approved data sources, decomposes your question into checks, runs those checks itself, then writes a plain-language answer with the supporting figures. Access permissions are set by whoever administers the data, so the agent only sees what that person allows.
Take a concrete example. OpenAI's Data agent, released inside ChatGPT Work on 10 September 2026, connects to sources including Google BigQuery, Snowflake, Databricks, Amazon Redshift, MongoDB, ClickHouse and Datadog, as well as files sitting in Google Drive and SharePoint. It can also build and read dashboards inside Power BI, Tableau, Sigma, Omni, Oracle BI and ThoughtSpot.
That connector list tells you something important about who these tools were built for. Those are warehouses and BI platforms owned by companies with a data team. A 12-person trading firm in Kwun Tong does not have a Snowflake account.
The four steps, in order
- － Connect. Someone with admin rights approves which systems the agent may read. This is a governance decision, not a technical one.
- － Decompose. The agent turns "why did margin fall" into a list of testable sub-questions.
- － Investigate. It queries the data for each sub-question, following the evidence rather than a fixed script.
- － Explain. It writes the finding in sentences, attaches the figures, and usually offers a chart you can share.
What Can a Hong Kong SME Use a Data Agent For?
The realistic uses for a smaller Hong Kong business are narrow but real: explaining a swing in monthly sales, spotting which customers stopped ordering, comparing branch performance, and checking whether a promotion actually paid for itself. All four are questions owners ask monthly and answer by guessing.
Data analysis is also where Hong Kong SMEs are furthest behind. In a Dah Sing Bank survey of more than 340 local SMEs conducted in May 2026, marketing and content led AI use at 56% and customer service at 42%, but data analysis trailed at 30%. Roughly 23% of the firms surveyed had adopted AI at all, with a further 32% planning to within one to two years.
Put plainly: most local SMEs have used AI to write a Facebook post, and very few have used it to understand their own numbers.
Four questions worth handing to a data agent
- － "Which customers bought last quarter but not this one?" A retention question that almost nobody runs manually, and the one most likely to be worth money.
- － "Did the September discount raise volume enough to cover the margin we gave away?" Requires joining sales and cost data, which is exactly where manual analysis stalls.
- － "Which two hours of the week are we overstaffed?" A rostering question sitting inside POS timestamps that nobody has time to open.
- － "Which supplier's prices moved most this year?" Buried in invoices, and the reason quiet cost creep goes unnoticed.
What Do People Get Wrong About Data Agents?
The three most common misconceptions are that a data agent replaces an analyst, that it works on messy data, and that it removes the need to check anything. None is true. It compresses the mechanical part of analysis; it does not supply business judgement, and it cannot fix records that were never kept properly.
Misconception 1: it replaces the person who reads the numbers. It replaces the querying, not the deciding. Someone still has to know that the drop in July is seasonal because the schools are out, and that the same drop in November would be an emergency.
Misconception 2: it works on whatever data you have. If the same customer is recorded three different ways across your systems, the agent will confidently report three customers. Agents inherit the quality of the records underneath them, and small businesses usually discover their data is messier than they assumed on day one.
Misconception 3: the answer needs no checking. These systems still make mistakes, and a fluent, well-formatted wrong answer is more dangerous than an obviously wrong one. This is the same failure mode behind AI inventing facts about your business. Treat the first month of output the way you would treat a new hire's first month: useful, and verified.
Misconception 4: it is a security-free upgrade. Pointing an AI system at your customer and financial records is a data-governance decision. Who approved the connection, what it can read, and where the data is processed all become questions you should be able to answer before you switch it on.
Do You Have Enough Data for a Data Agent to Be Useful?
You need three things, not volume: records kept consistently for at least twelve months, a single reliable identifier for each customer or product, and data that lives somewhere a system can read. A shop with two years of clean POS exports is better positioned than a firm with ten years of inconsistent spreadsheets.
Twelve months matters because without a full year you cannot separate a real decline from a seasonal one. A consistent identifier matters because every useful question involves grouping. And machine-readable storage matters because a photograph of an invoice is not data, however many of them you have.
If you fail all three, the honest next step is not an AI tool. It is spending a quarter fixing how records are captured, which is unglamorous and will improve every decision you make afterwards, with or without an agent.
Is It Worth Paying Attention to Yet?
For most Hong Kong SMEs the answer in late 2026 is: learn the concept now, buy later. The current generation of data agents is priced and connected for companies with data warehouses. The concept, though, is arriving inside ordinary business software fast, and the businesses that already keep clean records will be the ones able to use it on day one.
The gap is not access to AI. Adoption figures show most Hong Kong SMEs are within a year or two of using it in some form, with an HKT Business survey in 2026 putting SMEs at 49% implemented, piloting or planning, against 79% of large enterprises. The gap is readiness: whether the numbers underneath are in a state worth investigating.
So the useful work this quarter is not choosing a tool. It is making sure that when you finally point something clever at your business, it finds a clean set of books rather than a shoebox.
Technology only helps when someone stays beside you long enough to make it fit how you actually work. We understand AI. UD stands with you.
Frequently Asked Questions
Is a data agent the same as an AI agent?
A data agent is one type of AI agent, specialised in analysis. Other AI agents handle customer replies, scheduling or document processing. What they share is autonomy across several steps rather than a single answer.
Can a data agent read my Excel files?
Some can read spreadsheets stored in Google Drive or SharePoint. Consistency matters more than format: one clean sheet with stable column names is more usable than twenty files with different headings.
Does using one mean sending my data to an AI company?
It depends on the product. Some process data inside infrastructure the customer controls, others send it to the vendor. Ask where processing happens and what is retained before you connect anything.
How is this different from asking ChatGPT to analyse a spreadsheet?
Uploading one file is a single task. A data agent holds a standing connection to live systems, runs multi-step investigations across them, and can be scheduled to look without being asked.
Understanding the concept is step one. Working out which of your numbers are ready, and which need fixing first, is step two. UD has spent 28 years helping Hong Kong businesses get there, and we will walk you through every step, from a plain assessment of your current data to a system your team will actually open.
Reviewed by the UD AI team, Hong Kong. Published 15 September 2026.
