Can an AI executive assistant connect ERP, sales, and collections?

In 60 seconds: Yes, but begin with a read-only briefing assistant. Give it governed views of CRM opportunities, ERP orders, invoices, and collections activity. Make every answer show its source and data timestamp. Let it draft a follow-up or create an approval request, but do not let it change credit terms, post payments, or promise delivery by itself. The first release should answer a short list of recurring management questions and fail clearly when data is missing.

The useful question is not whether a model can chat with an ERP. It is whether a manager can trust the answer enough to act. That depends more on definitions, permissions, and data freshness than on eloquent prose.

Decision framework

Choose the first use case by checking four things:

• The question recurs every week, such as “Which invoices need attention today?”
• The answer exists in named systems and has an agreed definition.
• A wrong answer can be detected before money or a customer is affected.
• One person owns the follow-up.

A daily briefing is a better first release than an autonomous collections agent. It creates value while exposing disagreements about pipeline stages, overdue balances, disputed invoices, and promised payment dates.

A concrete workflow

1. At a fixed time, the service reads approved CRM and ERP views using a service identity with minimum permissions.
2. It normalizes customer IDs, currency, owner, due date, dispute state, and latest activity.
3. Rules select exceptions: overdue invoices without a dispute, opportunities without a next step, or orders blocked by missing information.
4. The model summarizes those records and links each statement to its source record.
5. The manager asks a follow-up in Teams or another controlled interface.
6. The assistant can draft an email, create a CRM task, or open an approval. A person confirms any external communication or financial change.
7. The system records the question, source set, response, chosen action, and approver.

Microsoft describes Copilot in Dynamics 365 Sales for opportunity summaries and sales preparation. Its Sales agent introduction explains Microsoft 365 and CRM integration. Product availability and licensing depend on the tenant and plan, so confirm current terms directly rather than designing from a marketing screenshot.

If the ERP is not Dynamics, inspect its API, authorization model, and change history. A custom connector may be appropriate, but a connector is only transport. It does not resolve mismatched customer codes or competing definitions of revenue.

Controls that make the answer useful

Every response should include “as of” time, source links, filters used, and omitted records. Separate observed facts from generated suggestions. For example: “Invoice 184 is 12 days overdue” can come from ERP data. “Call this customer first” is a recommendation and should be labeled as such.

Use row and field permissions. A regional manager should not gain access to every subsidiary because the assistant has a broad service account. Microsoft documents governance capabilities in the Copilot Control System, but tenant administrators still have to configure access and review usage.

Pilot checklist

• Write ten management questions in plain language.
• Define each metric, currency rule, and cutoff time.
• Map source fields and owners.
• Start read-only and deny unknown actions.
• Test missing, stale, duplicate, and disputed records.
• Require links to underlying entries.
• Log access and approvals.
• Measure correction rate and action completion, not answer length.

For prioritization, connect this work to what the business should automate first. The assistant should surface a bottleneck, not hide a broken process behind a chat box.

Limits

Natural-language access does not make data complete. Model output can be wrong, prompt injection can arrive through documents, and a summary can omit a critical exception. Do not ask the model for legal conclusions about collections or let it generate customer commitments from inference. Keep sensitive fields out unless they are necessary and authorized.

Latency also matters. A briefing built from last night’s snapshot must say so. If a manager needs a real-time available-to-promise answer, that requires a different integration and stronger validation.

A Kiia-sized first release

Kiia would begin with one role, two data sources, and a read-only morning briefing. We would add a draft action only after managers can trace each answer. That creates a foundation for WhatsApp and Teams follow-up while keeping the ERP and CRM, not the model, in charge of business truth.