Skip to content
KiweelyKiweelyKiweely

Implementation and support need different rules for Workday AI

Deployment Agent inspects Implementation, Sandbox, Preview, and Customer Central. It is not a Production chat assistant. Support keeps events, integrations, and change control. A desktop agent still asks before it changes a tenant.

  • By Kiweely
  • 6 min read
On this page

Workday has not published an application-management policy, a support-desk standard operating procedure, or a rule that “AI used in implementation is allowed in Production support.” You have to write that policy. The product facts that constrain it are already public.

Deployment Agent is for implementers and administrators in Implementation, Preview, Sandbox, and Customer Central connected to non-Production. It is not available as a conversational assistant in Production. A support model that tells analysts to “ask Deployment Agent about the Production worker” is asking them to do something Workday’s setup topic says they cannot do.

Two modes, two rules

Implementation Support after go-live
Tenant the agent may chat in ImplementationImplementation, Sandbox, Preview, Customer Central Support after go-liveStill only those. Production chat is not a feature. A Sandbox that has just refreshed is a copy of Production, so it is the closest published place to inspect Production’s configuration.
What “done” means ImplementationA definition, policy, or package you can migrate, plus a test event Support after go-liveA Production event, integration event, or a change migrated on purpose
Change path ImplementationEdit in the build tenant, activate security, rehearse Object Transporter Support after go-liveThe same tools. A support fix that is configured only in Sandbox disappears on refresh.
Chat retention Implementation14 days, per tenant Support after go-live14 days. A support ticket that cites only a chat will rot.
Who signs ImplementationThe person who activates and migrates, then the client who runs the test Support after go-liveThe person allowed to change Production, under your change control

Adoption Agent is a different product. The catalog says you register it in Preview to use release-impact analysis. That is a support-relevant agent for “what did this feature release change,” not a substitute for Deployment Agent, and not a Production worker assistant. BP Optimize Agent analyzes business processes and recommends improvements. It is not a ticket bot. Do not write a support policy that treats every Workday-built agent as one queue.

A policy you can enforce

Build phase

  • Ask is allowed for named security groups on the Deployment Agent Users domain, in named non-Production tenants.
  • Prompts that inspect configuration must say “in my tenant” or name the connected tenant. Q&A must not be pasted into a design as if it had seen the tenant.
  • Anything the team will rely on is copied to the decision log the same day.
  • The team verifies Ask once per tenant with a question only the tenant can answer.
  • Setup is also done in Production when Sandbox should keep the agent across refreshes. The policy states that this does not authorize Production use. It states who may deactivate the Production registration so a refresh does not turn the agent back on.
  • Flex Credits: the enable topic says Ask consumes none in Production. The policy must not claim that for other agents.

Support phase

  • Production incidents are diagnosed from the business process event, the integration event, the domain policy, and, when the question is “does Production match what we deployed,” a Sandbox compare or Tenant Compare after a known refresh. They are not diagnosed by opening Deployment Agent in Production.
  • A proposed fix is built in an Implementation or Sandbox tenant, reviewed on an Object Transporter diff, and migrated. Extracts still cannot be loaded straight to Production. Security policy migration merges unless someone chooses overwrite. Group permissions on business processes are additive and will not remove a Production grant. The policy should say who is allowed to choose overwrite and who removes grants by hand.
  • No worker’s pay, medical detail, or national identifier goes into a prompt. The security example Workday published is about a field and a role, not a named person’s compensation amount. Follow that shape.
  • Customer Central access stays with the customer. Workday recommends at least two security administrators and two customer-central administrators when a partner was involved, so the customer can operate after the partner leaves. Support contracts that leave a single partner account as the only administrator fight that recommendation. The initial owner account deactivates after 60 days.
  • A desktop agent used by support, such as Kiweely Vine, uses a named person’s sign-in, asks before any change, and is scoped to one client per conversation. Proxy tests are watched and recorded as that security group. Production changes are still approved. Vine is not covered by the Deployment Agent setup, and a personal developer tenant being free does not make a client’s Production tenant free. A client tenant needs a Personal license at $99 a month or a firm seat at $149 a month.

What to measure

Measure records, not prompts.

Measure Implementation target Support target
MeasureDesigns that name a tenant object a second person opened Implementation targetAll of them Support targetAll fixes
MeasureMigrations with a saved post-migration summary Implementation targetAll packages that moved Support targetAll Production fixes that are configuration
MeasureIncidents closed with only a chat link Implementation targetZero Support targetZero, because the link dies
MeasureProduction changes made only in Sandbox Implementation targetZero, unless an extract exists Support targetZero
MeasureAgent users who are not on the domain policy Implementation targetZero Support targetZero

Workday will not produce this table for you. It will produce the migration summary, the event, and the agent’s 14-day history. Your policy chooses which of those counts.

Failure modes

  • One policy paragraph: “AI is encouraged.” Analysts then paste Production data into a non-Production agent and call it support.
  • Support “mirrors Production” by testing in a Sandbox that refreshed last night and no longer contains yesterday’s fix.
  • The implementation partner leaves, and the only people who could deactivate the agent or run Object Transporter are gone. The agent keeps returning on every Sandbox refresh because nobody can deactivate it in Production.
  • A support bot promise based on Payroll Agent or Financial Audit Agent. The catalog lines are payroll assistance and procure-to-pay audit samples. Write those down if you enable them. Do not widen them in the policy.
  • Treating Vine and Deployment Agent as interchangeable for a night-shift analyst. One inspects non-Production configuration inside Workday. The other signs in as the analyst and can change what that analyst can change, after approval. The approval is the control. The product name is not.

Frequently asked questions

Can support use Sandbox as “almost Production”?

Yes, for configuration that a refresh has copied from Production. No, for data or configuration that existed only in Sandbox, and no, as a way to chat with a Production assistant. Confirm when the refresh ran before you call Sandbox a copy.

Should AMS ban desktop agents?

Ban unapproved changes and shared sign-ins across clients. A desktop agent that asks before it changes, and that the analyst watches, is compatible with change control. A desktop agent that nobody approved is not. The policy should say which.

Does implementation’s Ask history transfer to the support team?

No. History is per tenant and deleted after 14 days. Transfer the decision log and the migration summaries. Do not transfer a login and hope the chats are still there.

Who updates the policy at the next feature release?

The same owner who reads What’s New in Workday in Sandbox Preview, where the HCM course says to run it when the preview window starts. Feature releases reach preview about five weeks before they reach every tenant. Agent behavior in the admin guide can change in those notes. The policy should name the person who rereads the Deployment Agent setup topic each release, not “the AI.”

Sources checked

Public Workday topics listed above, fetched 2026-11-16. Kiweely prices and approval behavior are from kiweely.com. Workday does not publish an AMS or support-policy document on the pages checked.

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.