On this page
Can an AI agent do data entry?
Yes, for repeatable records with clear rules, with a person answering exceptions and sampling the output. An AI agent for data entry reads an email, form or document, picks out the fields, checks them against your rules and types the record into your system through its screens, the way a clerk would. That last part matters: the agent needs no API, so it can work in an old internal app or a vendor portal.
The named production evidence is vendor-published. Microsoft reports that Graebel, a relocation company with about 1,500 employees, runs an agent that reads relocation request emails, validates them and enters them into its Global Connect platform, which could not support an API integration; exceptions and low-confidence cases go to people. A Microsoft customer story says Asurion's agent corrects invoice data that fails capture, so the invoice can be reprocessed into its finance system. An Andreessen Horowitz article based on interviews with teams running these agents describes, among others, an agency that populates an applicant tracking platform as soon as a candidate interview ends.
None of these sources publishes an error rate. Treat this as a supervised hire.
This page covers entering data into systems. For pulling data out of complex documents, see document processing; for invoices, accounts payable; for keeping CRM fields current, CRM data hygiene.
What the agent does and what stays with you
Human
- Decide which fields to fill, the allowed values and the mapping from source to system
Agent with review
- Read emails, forms and documents and pull out the fields to enter
- Check each record against business rules before it is entered
Agent
- Type the record into the target system through its screens
Agent with review
- Correct values that do not match the system, such as item or order numbers
- Ask a person when a value is missing or ambiguous
Human
- Compare entered records with the source to find errors
Agent
- Keep a log of completed work
8 tasks: 2 agent, 4 agent with review, 2 human
| Task | Who does it | Why |
|---|---|---|
| Decide which fields to fill, the allowed values and the mapping from source to system | Human | a16z says runbooks, internal terminology, preferred formats and who to escalate to are specific to a company and often to one team. You set the field rules; no source read for this page shows an agent setting them. |
| Read emails, forms and documents and pull out the fields to enter | Agent with review | Microsoft says Graebel's agent extracts key data from unstructured emails into structured form with confidence scoring and sends low-confidence cases to people. |
| Check each record against business rules before it is entered | Agent with review | Graebel's agent validates each request against business rules, service logic and compliance requirements before any action, and exceptions go to a person. |
| Type the record into the target system through its screens | Agent | Microsoft's documentation lists automated data entry as a computer use task, and Graebel's agent enters data and completes transactions in a system without an API; a person spot-checks. |
| Correct values that do not match the system, such as item or order numbers | Agent with review | Asurion's agent trims or adjusts item and purchase-order numbers by defined instructions before reprocessing; we keep a reviewer on corrections because a changed value can look plausible. |
| Ask a person when a value is missing or ambiguous | Agent with review | Copilot Studio's human supervision feature pauses the run and sends a clarification request, but Microsoft says it may not trigger every time a person would want a pause. |
| Compare entered records with the source to find errors | Human | O*NET lists this as a core data entry task; no source read for this page shows an agent checking its own entries, and a16z describes wrong values that pass visual checks. |
| Keep a log of completed work | Agent | Copilot Studio keeps a run history showing what the agent saw and clicked, with logs sent to Purview and Dataverse according to Microsoft. |
The split follows checkability. Typing a record that passed your rules is cheap to sample and fix, so the agent does it and a person spot-checks. Corrections and ambiguous records go to a person, because a changed value can look right. The a16z authors found agents strongest "on standardized, repeatable tasks with a clear, well-defined path" and weakest where nobody can tell whether a task succeeded.
The O*NET profile for data entry keyers lists the work a person did: read source documents and enter data, compile and verify data before entry, compare data with the source to find errors, and correct errors or report them to supervisors. On the evidence above, Graebel's agent extracts the fields and checks them before entry. Comparing entries with the source stays with you; that comparison is how you know the agent is working.
Speed is not the reason to hire one. The a16z article says that in agentic mode a task a person finishes in two to three minutes can take an agent eight to ten. The gains come from running around the clock and from volume.
A worked example: one change-of-address email
Assume, for this example, that customers email address changes to a shared mailbox, and a clerk enters each one into an order system that has no API.
An email arrives: a customer gives a new street address and asks that open orders ship there. The agent reads it, finds the customer number in the signature, and checks it against the system: the number exists and the name matches. It opens the customer screen, enters the new street, city and postal code, and logs the old and new values.
The email also asks to change a delivery already in transit. Your rules say in-transit orders go to a person, so the agent stops there and puts the item in the exceptions queue with that reason. The named reviewer calls the carrier. On Friday, the reviewer's sample of 50 records includes this one; the address matches the email, and the record passes.
What you need to provide
The agent touches whatever your records contain: usually personal data (names, addresses, phone numbers) and sometimes customer or financial data. It needs:
- A field list and mapping, with 10–20 real examples and their correct entries. Microsoft's documentation says adding detail to the instructions improves accuracy.
- A dedicated account in the target system with only the access the job needs. Microsoft warns that with maker-provided credentials, anyone using a shared agent "can act with the original author's access on the configured machine".
- A machine for the agent. Copilot Studio's computer use runs on a Windows computer you set up; Microsoft lists Citrix and other virtualized environments among app types that might not be supported for password sign-in, so check your target system first.
- Business rules and a named person for exceptions, with a response time.
Keep the agent away from payments and approvals. Microsoft's FAQ says its computer use tool is not intended for making financial transactions or for recommending or scoring in sensitive domains such as hiring, healthcare or finance.
Job brief template
Give this brief to a vendor, an operator or your own team. Replace the bracketed values with your own.
Copy-ready brief
- Goal
- Every [request type] received in [mailbox or folder] is entered into [system] within one business day, or sent to a named person with the reason it could not be entered.
- Deliverable
- Records created or updated in the target system, an exceptions queue with the reason for each item, and a daily log listing each source item, the record it produced and any value the agent changed.
- Inputs
- A field list for the target system, with allowed values, formats and which fields are required
- A mapping from the source (email, form, document type) to each field, with 10–20 real examples and their correct entries
- A dedicated, least-privilege account in the target system and access to the source mailbox or folder
- Business rules: what makes a record valid, what counts as a duplicate, and what to do with each common exception
- The name of the person who answers the agent's questions and their expected response time
- Acceptance criteria
- Every source item received in the period is either entered or in the exceptions queue; the daily counts reconcile.
- On a weekly sample of 50 entered records, at least 49 match the source on every required field.
- No record is written outside the agreed system, record types and fields.
- Every value the agent changed from the source (for example a corrected item number) is in the log with the original value and the rule it applied.
- Exceptions reach the named person within 2 business hours of the run that found them.
- Systems
- source mailbox, shared folder or web form
- target system of record (for example an ERP, CRM, order system or an internal web app without an API)
- agent platform with screen automation (for example Microsoft Copilot Studio computer use)
- Constraints
- Records may contain personal data (names, addresses, contact details); the agent reads only the agreed sources and writes only the agreed fields.
- No payments, approvals or deletions; Microsoft says its computer use tool is not intended for making financial transactions.
- The agent runs on an isolated machine with an allow list of sites and applications.
- Use a dedicated account, not a person's credentials.
- Volume
- Example: 50–1,000 records a week
- Cadence
- Example: hourly or daily runs; weekly sample review
- Human checkpoints
- Before launch: the process owner signs off the field list, mapping, business rules and example set.
- Every exception: the named person completes or rejects the record and records why.
- Weekly: a reviewer checks 50 entered records field by field against the source.
- After any change to the target system's screens: rerun the example set before the agent resumes.
The reconciliation criterion matters most. If every source item is either in the system or in the exceptions queue, nothing disappears silently.
What it costs
Human benchmark
Data entry keyers
$41,340
per year (median annual wage)
BLS OEWS 2025 national annual median for SOC 43-9021 (series OEUN000000000000043902113); the hourly median, $19.88, is a separate BLS series (OEUN000000000000043902108); wage only, no benefits or overhead
Source: OEWS national annual median wage, Data Entry Keyers (series OEUN000000000000043902113)
Agent price range
$0.04 to $0.12
per computer use step (one record takes several steps)
Copilot Studio capacity pack, $200 per 25,000 Copilot Credits per month ($0.008 per credit), × 5 credits per step on a standard model or 15 on a premium model; the pay-as-you-go meter needs no up-front license commitment and an Azure subscription is required; its per-credit rate was not on the page; machine costs excluded
Sources: Microsoft Copilot Studio Plans and Pricing, Automate web and desktop apps with computer use
The two figures use different units. Compare them at your own volume.
The human figure is the BLS 2025 national annual median wage for data entry keyers, $41,340 a year, before benefits and overhead; BLS's separate hourly median series gives $19.88 an hour. The agent figure uses one vendor's published price: Copilot Studio sells capacity packs of 25,000 Copilot Credits at $200 per pack per month, which is $0.008 per credit, and Microsoft's documentation bills 5 credits per computer use step on a standard model and 15 on a premium one.
Assume, for this example, that one record takes 12 steps on a standard model: 12 × 5 = 60 credits, and 60 × $0.008 = $0.48 per record. Your step count depends on the form; measure it in the trial. Microsoft's pricing page also lists a pay-as-you-go meter and a pre-purchase plan. A16z puts agent inference at roughly $6–8 per hour, citing a founder's estimate; treat that as an order of magnitude.
Risks and human checkpoints
A value is read wrong but looks plausible, such as payment terms of net 60 entered as net 30.
How you notice it
Sample records field by field against the source each week; a16z gives this example of an error that passes every visual check until an invoice goes out wrong.
The system's screens change, or a dropdown, date picker or dialog trips the agent, and it loops or stops.
How you notice it
Alert on failed or long runs and reconcile source counts with created records daily; Microsoft lists sensitivity to layout changes, difficult controls and getting stuck as known limitations.
Text inside an email or document instructs the agent to do something else (prompt injection).
How you notice it
Run on an isolated machine with an allow list and a least-privilege account, and review the run history for actions outside the allowed sites; Microsoft warns that hidden commands in screenshots, pages or inputs can influence the agent.
The agent proceeds on an ambiguous record instead of asking.
How you notice it
Track the share of records sent to exceptions; a sudden drop is a warning sign. Microsoft says clarification requests might not trigger in every situation where a person would want a pause.
The entry shows as submitted, but the record stalls downstream and nobody follows up.
How you notice it
Check next-day status of submitted records in the target system; a16z describes an insurance portal submission that showed as received and then stalled on a phone call the agent never saw.
Everyone who uses a shared agent acts with the builder's access.
How you notice it
Use end-user or dedicated service credentials and review who triggered each run; Microsoft warns that with maker-provided credentials anyone using a shared agent acts with the original author's access.
Human checkpoints
- Before launch: the process owner signs off the field list, mapping, business rules and example set.
- Every exception: the named person completes or rejects the record and records why.
- Weekly: a reviewer checks 50 entered records field by field against the source.
- After any change to the target system's screens: rerun the example set before the agent resumes.
The quiet failure is a wrong value that looks right. The loud ones, a stuck run or a changed screen, show up in run history and in the daily count. Microsoft lists sensitivity to layout changes, difficulty with dropdowns and date pickers, and loops among the tool's known limitations, so the count is your first alarm.
How to run a trial
Run the agent for four weeks on one record type before it touches others.
- Pick one source and one record type with clear rules, such as address changes or new supplier contacts.
- Run the agent on last month's items in a test environment and compare every record with what your team entered.
- Go live on new items, with a person sampling 50 records a week field by field.
- Track the exceptions rate, failed runs, steps per record and reviewer minutes per week.
- Change one screen setting or form field on purpose in the test environment and check that the agent stops or escalates rather than guessing.
Pick one source and one record type
With clear rules, such as address changes or new supplier contacts
Run on last month's items in a test environment
Compare every record with what your team entered
Go live on new items
A person samples 50 records a week field by field
Track exceptions, failed runs, steps and reviewer minutes
Exceptions rate, failed runs, steps per record, reviewer minutes per week
Change a screen setting or form field on purpose
The agent should stop or escalate rather than guess
Our recommendation: count a record the agent sent to exceptions separately from a record it got wrong. An exception costs a few minutes of someone's time; a wrong record can sit in your system until a customer or an invoice exposes it.
If two weekly samples in a row meet the acceptance criteria, add the next record type. For structuring the test, see how to run a paid trial with an AI agent; for writing the brief, how to write a job description for an AI agent. Related: the administrative assistant role and the operations overview.
Sources
- Data Entry Keyers (43-9021.00)
O*NET OnLine (U.S. Department of Labor)PrimaryAccessed Nov 14, 2026
- OEWS national annual median wage, Data Entry Keyers (series OEUN000000000000043902113)
U.S. Bureau of Labor StatisticsPrimaryAccessed Nov 14, 2026
- OEWS national hourly median wage, Data Entry Keyers (series OEUN000000000000043902108)
U.S. Bureau of Labor StatisticsPrimaryAccessed Nov 14, 2026
- Automate web and desktop apps with computer use
Microsoft LearnPrimaryAccessed Nov 14, 2026
- FAQ for the computer use tool
Microsoft LearnPrimaryAccessed Nov 14, 2026
- Microsoft Copilot Studio Plans and Pricing
MicrosoftVendorAccessed Nov 14, 2026
- Computer-using agents in Microsoft Copilot Studio are now generally available
Microsoft Tech CommunityVendorAccessed Nov 14, 2026
- Asurion gets invoice automation back on track with Microsoft Copilot Studio
Microsoft Customer StoriesVendorAccessed Nov 14, 2026
- Can Agents Use a Computer Yet? We've Got the Data
Andreessen HorowitzVendorAccessed Nov 14, 2026
What we could not verify
- Named production evidence comes from customer stories published by Microsoft (Graebel, Asurion); neither publishes an error rate or record volume for the agent. The other deployments cited come from anonymized interviews in an Andreessen Horowitz article.
- No independent study of field-level accuracy for AI agents doing data entry was found.
- Copilot Studio's pay-as-you-go price per Copilot Credit was not readable on the pages fetched; the agent cost uses the published capacity pack price ($200 per 25,000 credits per month). The cost of the Windows machine the agent runs on was not checked.
- How many computer use steps one record takes depends on the form and was not measured; the per-record figures on this page are labeled examples.
- Data-protection duties for the personal data in the records (GDPR, US state laws) were not researched for this page.
Related
- AI agents for business operations: what to hand over
- AI agent for document processing and extraction
- Can AI replace administrative assistants? Task by task
- AI agent for accounts payable: what it does and costs
- AI agent for CRM data hygiene: clean and dedupe records
- How to test an AI agent before buying: run a paid trial
- AI agent job description: how to write one, with template