Skip to content
KiweelyKiweelyKiweely

Ask Claude, ChatGPT, or Gemini About a Failed Workday Run

An assistant can ask whether a run was sent, how many rows moved, and whether a change is waiting. It should see counts and outcomes, not the file.

  • By Kiweely
  • 9 min read
On this page
Asking Claude, ChatGPT, or Gemini about a failed Workday integration run

The question is worth connecting an assistant for: did the run go, what was the outcome, and is a change waiting. The question that is not worth connecting one for is: read the file and tell me who was underpaid. Kiweely Roots, coming next, draws that line in the product. Claude, ChatGPT, and Gemini reach Roots over MCP to check a run, start one, or ask for a change, as the person asking. The record those tools can be answered from is times, counts, and outcomes. The Workday rows are not kept after the run, so a later conversation cannot honestly be answered from a stored file. This article is that boundary, using the Roots page as read on 16 November 2026. The product is not live. The boundary is the requirement to take into a design, including a design you build before Roots exists.

What the page shows the assistant doing

The Roots page includes a pictured exchange. It is an illustration, not a measured result. A person asks whether last night’s payroll file reached the provider, and asks to add a new country entity. The assistant’s reply, as drawn, has three parts:

  • The run went at 02:01.
  • It included 27 workers, and none were rejected.
  • The country change has been tested on Sandbox and waits for approval, because the integration touches payroll.

That is the entire useful surface. A clock time, a count, a rejection count, and a statement of where a requested change sits. The assistant does not quote workers, rates, or account numbers. It does not claim the payroll provider’s internal posting if all it knows is that the file was sent and nothing was recorded as rejected. “None rejected” means the outcome record says none rejected. If your outcome record only knows that Workday delivered the file, the assistant must say delivered, not accepted. Those statuses are different, and the vendor rejection guide is why.

A second illustration on the same page is the benefits case the assistant is not required to narrate in prose: 14 rows rejected, a coverage tier mapped from elections, a Sandbox rerun, an approval waiting. The human card in Roots is the approval. The assistant’s job in the payroll picture is to say that the card is waiting, not to click it from chat.

Check, start, and ask are three different tools

MCP is the transport the page names. The tools on the far side are what you review in a security pass. A single “Roots” server that exposes the file as a tool has ignored the page, even if the transport is MCP.

Tool What the person said What the assistant may receive What it must not receive
ToolCheck a run What the person said“Did last night’s payroll file reach the provider?” What the assistant may receiveStart time, end time or send time, status, counts, rejection count, the integration’s name. What it must not receiveThe file, the worker list, pay, national identifiers, the request body.
ToolStart a run What the person said“Run the joiner feed now.” What the assistant may receiveWhether the tested code was started, and the run id or the time. What it must not receiveA chance to rewrite the map as part of starting it. Starting runs the code. It does not author the code.
ToolAsk for a change What the person said“Add the new country entity.” What the assistant may receiveThat a change was created, that it was tested on Sandbox, and whether it is waiting. What it must not receiveA production deploy from the chat. Payroll waits. The page says the Ireland-style change waits because it touches payroll.

“As the person asking” is the other half of the security sentence. The assistant acts as that user, not as a shared integration account that can see every feed. Workato’s Enterprise MCP page, read the same day, calls a similar idea Verified User Access: the action inherits the authenticated user. Roots’ page says “as the person asking” and does not publish Workato’s product name for it. The requirement to test in early access is the same shape: two employees, two sets of integrations, no cross-read. A directory owner who can ask about the joiner feed should not be able to ask about the payroll file unless they can already see that integration in Roots.

Workday’s own assistant story is a different surface. Developer Agent, on Workday’s page, builds apps and orchestrations from developer tools, through Extend Professional, from 2026R2+. That is a builder sitting with a developer. It is not documented on that page as “ask whether the 01:00 file was rejected.” Do not point employees at Developer Agent for operations questions about EIB and Studio runs. The Roots page is explicit that the assistant is asking about integrations. Workday’s page is explicit that Developer Agent is for building on Workday.

What a good answer sounds like

Write the tools so the model is forced into this shape. If the tool returns a paragraph of XML, the model will summarize the XML, and you have leaked the file into a chat log you do not control.

Good answers:

  • “Payroll to the provider ran at 02:01. Status: sent. Workers in the outcome record: 27. Rejections recorded: 0. That does not include a later vendor acknowledgment unless one was loaded.”
  • “The 01:00 benefits run recorded 14 row rejections. A fix was tested on Sandbox. Production is waiting for approval.”
  • “I started the joiner feed at 09:14. I don’t have a final count yet.”
  • “I can’t see that integration as you.”

Bad answers, which mean the tool returned too much or the prompt invited a guess:

  • A list of employee names “who might have been in the file.”
  • A reconstructed 834 segment.
  • “I fixed production so the file will pass.”
  • “The carrier accepted everyone,” when the tool result was only a Workday Completed status.

The last one is the common operational lie, and an assistant will repeat it if your run book uses one field called success. Split the fields. Delivered, vendor-accepted, and fix-waiting are different. The run book guide is the schema that makes the good answers possible. This article is the conversation on top of that schema.

Start and change are where chat gets dangerous

Checking a run is read-only and is the right first tool to expose. Starting a run is an action. On a directory feed, starting the tested code may be exactly what the owner wants. On a payroll feed, starting a run sends a file. The page says the assistant may start a run. It also says payroll changes wait. Read those together: starting the existing tested payroll integration is a run of code you already approved; starting it because the model invented a new map is not allowed, because there is no new map inside the run. If early access lets a conversation both rewrite and send, it has collapsed design time into the data path, which the FAQ on the same page forbids. Agents design and fix code between runs. The run is the code.

Asking for a change is design time, started from chat. The page’s own picture ends with Sandbox and a wait, not with a sent file. The approval is in Roots, on a card a person answers. That matches the wider rule that a chat transcript must not be able to approve a payroll change by wording the request cleverly. The person approves in the product, after the diff. The diff is the before and after on the failed rows. The assistant can say the diff exists and is waiting. The assistant is not the diff.

How this differs from pasting the error into a chatbot

Pasting an integration event, a stack trace, or a vendor rejection into Claude, ChatGPT, or Gemini works today and is the practice Roots is trying to make unnecessary. It is also how worker data enters a consumer chat product. Studio’s logging course already tells developers not to put the raw message on the integration event, because anyone who can open the event can read confidential data. Pasting that message into a personal chatbot is the same exposure with an extra company in the path.

A remote connector that returns counts does not require the employee to decide what to trim. There is nothing in the tool result to trim. That is the point of “nothing kept after a run” plus a run book of outcomes. The model can still mis-speak. It cannot quote a row it was never given, unless the user pastes one. Write the internal note in those terms: use the connector for “what happened,” and do not paste the file beside it “for context.” Context is how the boundary fails.

Other platforms will offer a similar connector with a different default. MuleSoft’s MCP Connector exposes the operations you choose to publish from a Mule app. If you publish “get file,” you published the file. Boomi Connect exposes connector actions as tools. Workato Enterprise MCP can turn a recipe into a skill, and the skill is only as narrow as the recipe. The Roots page specifies the narrow tool on the platform rather than leaving it as an exercise. If you build the exercise yourself while you wait, build the narrow tool. A gateway is not a policy.

What to test when early access opens

Sit two people down. One is allowed to see the payroll integration. One is allowed to see a directory feed. Use Claude, ChatGPT, and Gemini if all three are in the claim, not one of them.

Test Pass
TestPayroll owner asks if the file went PassA time, a count, a rejection count, and no names
TestDirectory owner asks the same payroll question PassRefusal
TestEither owner asks for the file or for a named worker’s pay PassRefusal, because the tool cannot return it
TestPayroll owner asks to add an entity PassSandbox result, waiting, no production send
TestSomeone tries to approve the waiting change inside the chat transcript PassThe change stays waiting. Approval is the card in Roots
TestA run is started from chat PassThe tested code runs. The map is unchanged
TestThe vendor later rejects rows PassA check after that acknowledgment reports the rejection count, and does not continue to say “none rejected”

If the third test fails, do not roll the connector out. The rest of the tests do not matter. The page’s retention sentence and the page’s assistant sentence are the same design. A tool that returns rows makes the retention sentence false in the only place employees look, which is the chat.

Frequently asked questions

Does the assistant call the model during the nightly run?

No. The nightly run is tested code. The assistant is a later, optional caller over MCP. Design time versus the run separates them. Connecting Claude does not put Claude on the 01:00 path unless you also built a runtime step that calls a model. The Roots page says that step is not there.

Can we use our existing Workday integration event as the tool result?

You can, and you should be careful. The messages tab and the output file are exactly the objects Studio warns you about. If the tool returns the output file or the raw message, you have rebuilt the paste problem. Return the status, the time, the processing duration, and counts you computed without attaching the document. If you cannot compute a rejection count from Workday alone, say so. Do not substitute the output file.

Which assistant does Roots support?

The page names Claude, ChatGPT, and Gemini, over MCP. It does not, on the page read today, name a proprietary chat as the only door. Early access should be checked on the three it names. A demo on one desktop app is not the claim.

Register interest in Kiweely Roots if the question your team actually asks in the morning is the payroll picture on that page. Bring the fields you are willing for an assistant to say out loud. If the list includes a worker’s name, it is the wrong list. Times, counts, and outcomes are the list the page already published.

Your Workday work, done by AI.

Kiweely Vine is the AI agent that builds, deploys and tests Extend apps as real people, makes changes in any Workday task screen, answers from your real tenant and builds reports.

Get the next one in your inbox.

One email a month for Workday teams: what’s changing in Workday and what to retest, one tip tried in a real tenant, and what’s new at Kiweely.

We’ll email you once to confirm. Unsubscribe in one click from any issue.