Can AI do data entry better than a person types?
People key data with about 0.29 errors in 100 fields; in an October 2026 vendor test, ChatGPT and Gemini silently returned nothing on a 20-page document. BLS projects data entry keyers down 26 percent by 2035. Read 2026-11-06.
Contents
Short answer, from the sources below: AI models can read a document and return the fields in it, and in the only 2026 test on this job, three of four tools got every field right on short documents. The failure that two tests report is not typos but a silent gap on a long document or a hard number. The best human figure found is an error in 0.29 percent of fields for one person keying, and no source here puts a chatbot and a person on the same documents. This page lists what each source measured, who ran it and what it leaves out.
What people get wrong when they key data
A systematic review of 93 papers (published 1978 to 2008, all from clinical research) pooled the error rate of each way of getting data into a system. The rates "ranging from 2 errors per 10,000 fields to 2,784 errors per 10,000 fields" differed by method: "the pooled error rates for optical scanning, single-data entry, and double-data entry methods were 0.74% (0.21, 1.60), 0.29% (0.24, 0.35) and 0.14% (0.08, 0.20), respectively."[1] Single-data entry is one person keying a form; double-data entry is two people keying it independently and a third settling the differences.[1] The same review found that pulling data out of medical records by hand had a pooled error rate of 6.57 percent.[1]
That gives a yardstick: a person keying structured forms is wrong in about 3 fields in 1,000, and two people checking each other halve it.[1] The studies are about clinical forms, not invoices or bank statements, and most are older than any current software.
Can ChatGPT do data entry?
Data entry means two different jobs, and the sources test them separately: reading a document and returning its fields, and typing those fields into another system.
Reading a document and returning fields
Parseur, a company that sells document-extraction software, tested ChatGPT, Claude, Gemini and its own product in October 2026 on seven bills of lading, 77 fields in all. On the six documents of one or two pages, "ChatGPT, Claude and Parseur extracted all 66 fields correctly. Gemini made two errors."[2] On the seventh, a 20-page document listing 47 containers, ChatGPT (GPT-6) and Gemini (3.6 Flash) both returned null for every container number, though the document lists 47 containers.[2] Parseur writes that "Every other field in both answers is correct, which makes the miss easy to overlook."[2] Across all seven documents Parseur scored 78 (77 fields plus one bonus point), Claude 77, ChatGPT 76 and Gemini 74.[2] Parseur says it "checked every value against the printed document."[2]
Two things about that test matter for a buyer. The vendor's own product came first, so the ranking is a vendor's. And the problem it reports is not a wrong value but a missing one with nothing to say it is missing: "Neither ChatGPT nor Gemini signals in its output that data is missing."[2] Parseur also reports that "the same prompt produced four different JSON (JavaScript Object Notation) structures" across the chat apps, so each would need its own mapping before the data reached a business system.[2] How to test an agent for wrong or missing values is covered on testing AI agents and hallucination rates.
An academic test of the same job used ChatGPT-4 and Claude 3 Opus to pull numbers out of 105 published clinical trial reports and compared them with checked datasets. "For binary outcomes, ChatGPT-4 and Claude 3 Opus showed high accuracy in group size extraction (Pacc: 91%-94%) and moderate accuracy for event counts (Pacc: 57%-71%)", while "mean and standard deviation extraction was poor (Pacc: 24%-56%)."[3] The authors call it a proof-of-concept study.[3] The models are older than the ones Parseur tested, and the task, finding numbers inside long scientific reports, is harder than copying the total from an invoice.
Typing into another system
OpenAI's help page for ChatGPT agent, read on 2026-11-06, begins: "ChatGPT agent is no longer available. Use ChatGPT Work for longer, multi-step tasks and finished deliverables."[4] The page still lists the old limits ("Plus: 40 messages/month", "Pro: 400 messages/month").[4] Its replacement for web tasks is described on a second page: "Cloud browser gives ChatGPT Work its own browser on a separate computer in the cloud. It can read web pages, click buttons, enter information into forms, and carry out steps on supported public and signed-in websites."[5] The same page sets limits on that: "Some websites use security measures that restrict access from automated browser agents", and to the question whether the cloud browser can finish every task, the answer is "No. Support varies by website and action."[5] It pauses for "input, sign-in, or confirmation", and asks for confirmation before "actions that could be hard to reverse or create a financial, legal, account, or other real-world commitment".[5] Neither OpenAI page gives an accuracy figure or a count of mistakes.[4][5]
Automating entry without an AI model
Microsoft's Power Automate records or scripts the clicks and keystrokes a person makes. Its documentation says desktop flows "enable you to automate all repetitive desktop processes", and names "an employee of a large enterprise who automates data entry on an ERP system" as one of the people it is designed for.[6] The pricing page, in U.S. dollars and read on 2026-11-06, lists Power Automate Premium at $15.00 per user a month paid yearly, which includes desktop flows "in attended mode"; Power Automate Process at $150.00 per bot a month for desktop flows "in unattended mode"; and Power Automate Hosted Process at $215.00 per bot a month, with a Microsoft-hosted virtual machine.[7] The pages give no error rate for these flows.[6][7]
What this means for the job
The Bureau of Labor Statistics projection, relayed by O*NET, puts data entry keyers at 131,800 jobs in 2025 and 98,200 in 2035, "-26%", with 7,700 projected annual openings from growth and replacement.[8]
Not yet proven
- A chatbot and a person on the same invoices. The human error rates come from clinical forms published up to 2008. No published test runs a current model and a person on the same bank statements or bills.
- Accuracy outside one vendor's documents. The only 2026 test is Parseur's, on bills of lading, run by a company that sells a competing tool. No independent test on invoices, receipts or statements is published.
- How often a long document loses data. The 20-page result is one document. The test reports no rate for how often, or at what length, a chatbot drops a field.
- Accuracy of cloud-browser entry. OpenAI's pages describe what it can do and where it stops, and report no measurement of entered values.
- Error rates for scripted entry. Microsoft's documentation and pricing give no figure for desktop flows.
- Why keyers are falling. The projection is a number from a table; the O*NET page states no reason for it.
If you run a bookkeeping practice and have compared an extraction tool with your own staff on the same documents, I would like to compare notes: mackenzie@mackenziebowes.com.
Sources
- Error Rates of Data Processing Methods in Clinical Research: A Systematic Review and Meta-Analysis of Published Literature. PubMed Central, read 2026-11-06. pmc.
ncbi. nlm. nih. gov/ articles/ PMC10775420/ - Parseur, LLM Document Extraction Benchmark 2026 (tested October 2026), a vendor's own test. Read in Rotunda 2026-11-06. parseur.
com/ blog/ llm- document- extraction- benchmark - Assessing data extraction in randomized clinical trials with large language models, BMC Medical Research Methodology, published 2026-01-14. Read 2026-11-06. link.
springer. com/ article/ 10. 1186/ s12874- 025- 02729- 5 - OpenAI Help Center, ChatGPT agent. Read 2026-11-06. help.
openai. com/ en/ articles/ 11752874- chatgpt- agent - OpenAI Help Center, Using cloud browser in ChatGPT. Read 2026-11-06. help.
openai. com/ en/ articles/ 20001280- using- cloud- browser- in- chatgpt - Microsoft Learn, Introduction to desktop flows - Power Automate. Read 2026-11-06. learn.
microsoft. com/ en- us/ power- automate/ desktop- flows/ introduction - Microsoft, Power Automate Pricing. Read 2026-11-06. www.
microsoft. com/ en- us/ power- platform/ products/ power- automate/ pricing - O*NET OnLine, National Employment Trends: 43-9021.00 Data Entry Keyers (Bureau of Labor Statistics 2025-2035 employment projections). Read 2026-11-06. www.
onetonline. org/ link/ localtrends/ 43- 9021. 00
Terms on this page
- AI agent: A system where a language model decides for itself how to do a task and which tools to use.