The 9 questions (with what “good” looks like)

Why connectors matter (more than the model)

AI agents become valuable when they can pull context from where work actually lives and then take controlled actions across systems (email, docs, ticketing, ERP/finance apps, data warehouses, etc.). That’s not an opinion; it’s how agentic platforms describe real-world automation: the “agent” is only effective when it’s integrated with third-party services/tools. (Beam AI)

But the moment an agent can act across systems, governance becomes existential. Identity and authorisation bodies are already warning that traditional human-centric IAM/authorisation patterns struggle when agents create sessions, run at machine speed, and act across multiple services on behalf of users. (ISACA)

So connectors aren’t plumbing. They’re risk, performance, and adoption.

The 9 questions (with what “good” looks like)

1) Are you indexing data, or running a federated fetch at query time?

This is the foundational architecture choice. Many enterprise AI vendors argue indexing is critical for speed/consistency, while federated approaches can be latency-prone and return partial results due to mismatched data models and ranking. (Glean)

What good looks like

• Clear explanation of which sources are indexed vs fetched on demand (and why).
• Measurable retrieval latency targets and how they’re achieved.

2) How do you enforce source permissions end-to-end?

Finance cannot tolerate “it usually respects access.” The connector must enforce the same permission model as the underlying system—continuously.

Leading enterprise search/AI providers explicitly emphasise “permissions-aware” retrieval and the complexity of mapping changing permissions into an identity model. (Glean)

What good looks like

• Permissions captured with content (not an afterthought).
• Explanation of how permission changes propagate (and how quickly).

3) How are you protecting sensitive data (financial, payroll, customer, contracts)?

A finance user asking a question in chat can accidentally trigger exposure if the system doesn’t implement strong controls for sensitive content and role-based access.

What good looks like

• Clear policies for redaction, classification, and restricted fields.
• Configurable controls at source/field/document level (not just “admin settings”).

(Ask for evidence: architecture notes, security docs, or third-party assurance where available.)

4) How fresh is the data: minutes, hours, or days?

Finance outcomes (cash, disputes, credit holds) are time-sensitive. If the “AI layer” is stale, it will look smart but make the wrong call.

Enterprise connector guidance commonly highlights the need for frequent updates (e.g., webhooks/incremental updates) to keep context current. (Glean)

What good looks like

• Stated sync method by connector: webhook, incremental crawl, batch.
• A documented “freshness SLA” by system (and what happens when it fails).

5) Who owns data modelling, and what happens when your business changes?

Your org structure, entities, products, approval matrices, and naming conventions shift constantly. If the vendor can’t explain how they model your data and keep it aligned, accuracy degrades quietly.

Glean’s connector discussion explicitly calls out data modelling responsibility as a differentiator question—because it’s where reality breaks. (Glean)

What good looks like

• A clear answer: vendor-led, customer-led, or hybrid.
• A change-management plan: what happens when you add a new entity, COA segment, or approval flow?

6) What exactly are you indexing: metadata, content, attachments, comments, workflow states?

“Connected to X” is meaningless unless you know what is connected. Finance use cases often depend on edge objects: approval comments, exception reasons, dispute codes, credit memo status, audit evidence attachments.

What good looks like

• A per-source inventory: objects indexed + fields included + exclusions.
• Ability to scope indexing (and prove scoping is enforced).

7) Are you organising information into a knowledge graph (or equivalent relationships)?

Finance decisions depend on relationships: vendor ↔ contract ↔ PO ↔ invoice ↔ payment ↔ entity ↔ approver. Without relationships, you get “answers,” but not reliable traceability.

Enterprise AI vendors increasingly describe structured relationship layers (knowledge graphs / identity schema) to make retrieval more accurate and permissions-aware. (Glean)

What good looks like

• Ability to represent and query relationships (not just keyword retrieval).
• Traceability back to source objects and relationships.

8) Where is the index stored, and what are the data residency implications?

CFOs and Controllers care because this influences regulatory posture, vendor risk, and cross-border data constraints.

This is explicitly one of the “9 connector questions” raised in enterprise AI connector guidance. (Glean)

What good looks like

• Clear storage location(s), tenancy model, encryption posture.
• Region controls and deletion/retention policy (including backups).

9) How does your connector infrastructure scale (and what breaks first)?

Scaling isn’t just volume—it’s permissions complexity, API rate limits, changes in schemas, and thousands of users querying simultaneously.

Connector guidance frames scaling as a core differentiator for “reliable, scalable access” across dozens of applications. (LinkedIn)

What good looks like

• Rate-limit handling, retry logic, backoff policies.
• Observability: connector health dashboards, ingestion lag monitoring, alerting.
• A clear “failure mode” story (degraded service vs silent inaccuracies).

The governance add-on CFOs should insist on (especially for agents)

If the product is moving from “answering” to acting, require a governance story that survives audits:

• Authenticate/authorise on every external call when acting on a user’s behalf (to prevent privilege misuse across contexts). (ISACA)
• Revocation and session control that works across systems (agents create new failure modes for traditional IAM). (ISACA)
• Action logs: who asked, what the agent did, what systems it touched, and evidence links.

If they can’t explain these simply, you’re buying operational risk.

A practical way to use this in vendor evaluation

Run a 30-minute “connector deep dive” with your IT/Sec + Finance Ops leads and ask these 9 questions verbatim.

Then score the vendor on three outcomes:

1. Trust (permissions + auditability)
2. Freshness (real-time relevance)
3. Completeness (can it see what finance actually needs, not just top-level docs)

This is how you avoid the common trap: a great chat demo that can’t survive real controls.

At Farben, our view is that the AI layer in finance must be context + controlled execution: deep connectors, permissions-aware retrieval, and human approval gates so automation produces cash and control outcomes, not just “nice answers.”

If you like our articles, consider subscribing!