Tarnfield Row

Tarnfield Row helps CIOs make sense of enterprise AI purchases: what each tool does, where products overlap and who will run the work.
Last updated:

Implementation plan

Who answers when the AI workflow breaks?

The assistant, search service and action tool may all have owners. The work still needs one person who can follow a request across them.

Published
Updated
Contents
  1. Name the person who owns the result
  2. Follow the accounts, not just the screens
  3. Separate build help from ongoing ownership
  4. Keep the request connected across tools

Name the person who owns the result

Imagine a new employee asking for access to a project. The assistant explains the process, search finds the policy, an agent prepares a request and another system grants access. If the person never receives it, “the chatbot worked” is no answer. Begin the delivery plan with the person responsible for the final business result. That person does not need to administer every product, but must be able to bring the right colleagues together.

For this example, the service owner might be the manager responsible for employee access requests. The identity team owns account rules, a project owner approves membership, and the support team handles unresolved cases. Use real roles from your organization. Do not assign responsibilities to a supplier without agreement, or assume the employee who built the first demonstration will remain available indefinitely.

Follow the accounts, not just the screens

Dust separates what an agent can access from who may use it, and syncs identity-provider groups through SCIM. Glean describes applying source permissions on every read and write. Google’s enterprise editions list identity management and source access controls.1,2,3,4

Make a row for each step in the access request: employee sign-in, policy lookup, approval, membership change and confirmation. Beside it, record the account doing the work and who maintains it. The employee’s account and a connection’s service account may differ. Ask the builder to explain which permissions actually govern the change, then have the identity team check that answer.

Rehearse a leaver and a transferred employee using approved test accounts. Remove access in the source system and check what the workflow can still read or change. Look for another group that might restore permission. The acceptance condition should describe the visible result: no restricted policy content disclosed, no unauthorized membership change, and an unresolved request routed to the right team.

Separate build help from ongoing ownership

Nexus describes an engineer building with the customer, while the team using an agent remains accountable for it. WRITER offers embedded transformation leads and deployment engineers. Microsoft documents maker tools and release approval gates.5,6,7,8,9

Put four responsibilities in the plan: build the workflow, approve its release, operate it and change it. Give each a named customer owner and, where agreed, a supplier contact. For the access-request example, the builder can prepare the connection, but the identity owner should decide its permissions. The process owner decides whether an unresolved case is acceptable. These decisions should survive the end of an implementation engagement.

Ask the supplier to hand over the task description, connected accounts, approval rules, known failures and recovery steps. Have a colleague who did not build the first version follow those instructions. They should be able to find a failed request, stop further changes and contact the right person. A presentation about the architecture is not a substitute for that exercise.

Keep the request connected across tools

Dust documents audit logs distinguishing human from agent actions, available on Enterprise. Microsoft documents maker audit logs in Purview. Google lists Cloud Audit Logs for enterprise editions.10,11,12,13

These records may answer different questions. A maker log can explain who changed the agent without showing whether a particular access request succeeded. Agree how your team links the original request, its approval, the attempted change and the result in the identity system. Ask each owner to show the same request from their side. Keep sensitive details with the people who need them; a shared incident note can point to restricted logs.

Try an awkward case before launch: the membership change succeeds, but the final confirmation fails. The support team should check the identity system before repeating the change. Record whether the task is complete, failed without a change, or still uncertain. Assign someone to resolve that uncertainty. Otherwise, each product team can report success while the employee remains stuck.

The finished handover can be one page with links: task, business owner, product owners, accounts, approval rules, evidence locations, stop procedure and support contacts. Add the change that requires another review, such as a new connection or wider group access. Keep the map with the operating instructions so it follows the service when people move roles.

How to read the evidence

Supplier material was checked November 8, 2026; this editorial guide was written November 9. Company confirmations are identified separately. Examples and trial checks are proposed exercises, not tests performed by Tarnfield Row.

  • This is our buying advice. The examples are proposed exercises, not results we measured.
  • This is a question to settle with the supplier before buying.
  • Based on supplier documents checked November 8, 2026. We have not tested the described features.

Sources

  1. Dust's permission model separates what agents can access from who can use them. dust.tt, checked November 8, 2026.
  2. Provisioned groups are synced from the identity provider through SCIM. docs.dust.tt/docs/user-documentation/admins/admin-governance/workspace-governance-roles-groups-and-permissions, checked November 8, 2026.
  3. Source-system permissions enforced on every read and write. glean.com/platform/security, checked November 8, 2026.
  4. Enterprise editions are administered with IAM, Workforce Identity Federation and source ACLs.. docs.cloud.google.com/gemini/enterprise/docs/editions, checked November 8, 2026.
  5. A Nexus engineer builds the first procedure with the team and shows it running in its own tools. agent.nexus, checked November 8, 2026.
  6. Layer 3: each agent is built, deployed and accounted for by the team that uses it. agent.nexus/telco, checked November 8, 2026.
  7. The vendor says it provides an embedded team of transformation leads and deployment engineers.. writer.com, checked November 8, 2026.
  8. Agents can be created by describing them in plain language, then tested before publishing.. learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio, checked November 8, 2026.
  9. Release pipelines and approval gates support Dev-Test-Prod promotion.. microsoft.com/en-us/copilot/products/copilot-studio, checked November 8, 2026.
  10. Audit logs distinguish between actions by human users and actions by AI agents. docs.dust.tt/docs/user-documentation/admins/audit-logs/audit-logs, checked November 8, 2026.
  11. Audit logs are only available to workspaces on the Enterprise plan. docs.dust.tt/docs/user-documentation/admins/audit-logs/audit-logs, checked November 8, 2026.
  12. Admins have full visibility into maker audit logs in Microsoft Purview.. learn.microsoft.com/en-us/microsoft-copilot-studio/security-and-governance, checked November 8, 2026.
  13. Enterprise editions support Access Transparency and Cloud Audit Logs.. docs.cloud.google.com/gemini/enterprise/docs/editions, checked November 8, 2026.