By Greg Nowak. Last updated 2026-07-06.
Agentic AI changes the practical risk of automation. A chatbot can give a poor answer and stop there. An agent connected to a CRM, helpdesk, mailbox, file store, or website can give a poor answer and then act on it.
That is why the unglamorous work needs to happen early. Before an OpenAI-powered agent gets access to sales or support systems, the business needs a clear map: what it can read, what it can write, whose authority it is using, where a person must approve the next step, and how the team can reconstruct the activity later.
OpenAI's ChatGPT agent points in this direction. It can browse, use connectors, interact with websites, work with files, and operate inside a virtual computer. The OpenAI Agents SDK shows the same pattern from the engineering side, with tools, handoffs, tracing, and structured workflows. The shift is from “help me draft this” toward “complete this task across several systems.”
That can be commercially useful. It can also become difficult to govern if the permissions discussion starts only after a pilot has already touched production customer data.
Start with systems of record, not prompts
The first planning question is not “Which model should we use?” It is “Which system owns the truth?”
In a sales workflow, the CRM may own accounts, contacts, opportunities, notes, tasks, and pipeline stage changes. In a support workflow, the helpdesk may own ticket status, customer messages, SLA state, internal comments, escalation flags, and linked knowledge-base articles. File storage may hold contracts, policies, call notes, exports, and sensitive spreadsheets. Email and calendar systems may expose customer conversations and meeting context.
The more an agent can see, the more capable it may look. It also has a larger blast radius when something goes wrong. OWASP's 2025 Top 10 for LLM Applications gives this a useful frame through risks such as prompt injection, sensitive information disclosure, excessive agency, and insufficient logging or monitoring. The problem is not only whether the model can produce a harmful sentence. It is whether the surrounding system has given the agent too much authority, too little supervision, or no useful audit trail.
This is where CRM and helpdesk integrations often get messy. A team starts with a productivity goal, adds connectors, and then discovers that the agent has inherited permissions from a personal login or an application registration with far broader access than the workflow needs. The demo works. Governance is already behind.
| Boundary | Decision to make | Control to put in place |
|---|---|---|
| System of record | Which platform owns the customer, ticket, deal, or case data? | Document the source of truth and prevent duplicate writes elsewhere. |
| Read scope | Which exact records, files, fields, or messages does the agent need? | Use least-privilege service accounts and narrow API scopes. |
| Write scope | Which fields, statuses, comments, or tasks may change? | Separate draft, recommend, and execute permissions. |
| Approval gates | Which actions need a person before they happen? | Require review for customer-visible messages, deletions, escalations, refunds, and stage changes. |
| Tool handoffs | When does work move between agent, API, CRM, helpdesk, and human? | Define handoff rules, fallback paths, and failure states before rollout. |
| Observability | Can the team see what the agent saw, decided, and did? | Keep tracing, logs, and reviewable records for important actions. |
Permissions are part of the product design
Enterprise permission models are not small. Microsoft Graph's permissions reference shows how broad and granular application permissions can become across mail, calendars, files, users, groups, sites, chats, security data, and more. That is not a flaw in Microsoft Graph. It is what an enterprise API is built to support. The integration designer still has to decide what the agent genuinely needs.
The same principle applies inside CRM and service platforms. Salesforce's Einstein Bots setup documentation is a useful reminder that bot and service experiences already sit inside permissioned business systems. External AI integrations should work with those boundaries, not bypass them with a generic user account, administrator token, or shared personal login.
A controlled rollout should separate access into three practical levels.
First, read-only context. The agent can retrieve relevant records, policy text, product data, prior case history, or CRM metadata, but it cannot change business state. This is usually the right starting point for discovery, summarisation, routing recommendations, and internal drafting.
Second, draft-and-queue authority. The agent can prepare a support reply, propose a CRM note, suggest a deal-stage update, or assemble a task list, but a person approves the final action. This mode is often best for customer-visible communication and sales operations where judgement, tone, and commercial context matter.
Third, bounded execution. The agent can perform a small set of low-risk actions without approval: adding an internal tag, creating a follow-up task, enriching a record from an approved source, or routing a ticket according to a documented rule. Even then, the allowed actions should be explicit, logged, and reversible where possible.
The danger is the vague middle: “the agent can update CRM records” or “the agent can work tickets.” Those phrases hide too much. Fixing a typo in a contact title is not the same as changing an opportunity stage, deleting a note, sending a customer reply, merging accounts, or escalating a complaint.
Prompt injection belongs in the workflow test
Prompt injection is often treated as a prompt-writing issue. In connected business systems, it is a workflow issue.
If an agent can read a customer email, browse a website, open a document, call a tool, and update a CRM field, then untrusted content can enter the agent's working context. OWASP highlights prompt injection as a major LLM application risk. Recent research on LLM-enabled open-source vulnerabilities also connects LLM-integrated systems with architectural risk patterns, including prompt injection, supply-chain exposure, and excessive agency.
For CRM and helpdesk use cases, testing should place hostile or misleading content in the places the agent may actually read: email bodies, ticket descriptions, uploaded files, website pages, knowledge-base articles, CRM notes, and third-party data fields. The test is not just whether the agent repeats bad instructions. The real question is whether those instructions can make it ignore policy, reveal data, call the wrong tool, update the wrong record, or skip an approval step.
The controls need to work together. Tool instructions should be narrow. System prompts should define the agent's role and firm constraints. API permissions should block unauthorised actions even if the model attempts them. Approval gates should catch risky transitions. Logs should support later review. None of those layers is enough on its own.
Logs need to explain the business action
Technical logs matter, but agentic workflows also need business-readable traceability. If a support ticket was escalated, the team should be able to see why. If a CRM task was created, the source record and rule should be clear. If a customer reply was drafted, the reviewer should be able to inspect the material the agent used. If an action failed, the fallback path should be visible.
The OpenAI Agents SDK documentation is relevant here because tracing and orchestration are core parts of agent systems. When agents use tools and handoffs, the implementation should preserve enough context to understand the path from input to output. For regulated, contractual, or commercially sensitive workflows, that record may matter as much as the automation itself.
A simple readiness test helps: if the business could not comfortably explain an agent action to a customer, manager, auditor, or account owner, the workflow is not ready for unsupervised execution.
A sensible rollout pattern
The best commercial path is usually narrow at first. Choose a workflow with clear inputs, low ambiguity, and a known owner. Good early candidates include internal ticket summaries, CRM research briefs, duplicate-check suggestions, renewal-preparation notes, and support triage recommendations.
Map the boundary before building the integration: system of record, data sources, API scopes, write permissions, approval points, failure states, logging, and rollback options. Only then should the team decide which agent tools, connectors, and service accounts are needed.
For many businesses, the right first milestone is not “the agent updates the CRM automatically.” It is “the agent prepares a useful recommendation, grounded in approved sources, with a visible audit trail, and a person can approve or reject it quickly.” That is still valuable automation. It also shows the organisation where more autonomy may be acceptable.
As confidence grows, selected actions can move from recommendation to controlled execution. Each new permission should be earned. The agent should not receive broader access simply because the previous demo worked.
Where an external project manager helps
This work sits between business process, data governance, integration architecture, and daily operations. Sales teams know where CRM friction hurts. Support teams know which ticket decisions are sensitive. Developers know what the APIs allow. Security and operations teams know what must be restricted, logged, and reviewed. Someone has to turn those views into a rollout plan that can actually be built.
That is where Greg can be useful as a freelance project operator: scoping the workflow, documenting systems of record, translating business actions into permission boundaries, setting approval gates, planning prompt-injection tests, and making sure OpenAI/API workflows use least-privilege service accounts rather than personal logins.
The boundary map is what makes the exciting part usable. Before agents touch the CRM, define the rules well enough that a mistake stays small, a person can intervene, and the business can explain what happened.
Related on GrN.dk
- ChatGPT apps need a permissions map before they touch company data
- Agentic AI: what it is, how it works, and why it matters now
- OpenAI Computer-Use Rollouts Turn Browser Tasks Into a Scoped-Credentials Project
Need help with this kind of work?
Plan a controlled AI agent rollout Get in touch with Greg.