Back
Blog / 
Implementation

Is your customer data ready for an AI agent?

written by:
David Eberle
A cable plugged into a blue record frame with an empty central space.

Published · Updated

Moving beyond a chatbot to AI that completes customer-service requests means connecting it to the systems where the work happens. The agent needs access to the information and actions your employees use, from finding the right order to recording a return. A useful answer is only part of completing that work.

Uncertainty about data readiness can then hold up the whole AI project before a pilot starts. Missing fields, inconsistent records and information spread across systems make that concern understandable. But treating readiness as a requirement to perfect all enterprise data leaves the team with a much larger task than assessing its first AI workflow.

Assess readiness against the service you want the agent to deliver. Establish which records and actions that scope requires, whether the integration can use them reliably, who will build each connection and what should happen when something is missing. Watching employees carry out the work gives you a concrete basis for that assessment, and for deciding which gaps block the proposed workflow, which can be handled by a person and which are outside its scope.

Trace the records and actions employees use

In a 2026 discussion with Typewise, a furniture retailer evaluating AI described a gap in its existing service data. Tracking numbers were available in its ERP, but not all carrier events were mapped into it. Staff still logged into carrier portals to get information that the connected system did not expose.

The retailer already used rule-based automation. This was not a report of a Typewise AI deployment or its results.

Sit with an employee handling a real request and trace the work across the ERP, carrier portal and any other system involved. Record the objects they open, such as an order, delivery note or return record, and how they identify the right one. Separate the fields they read from the fields they write or change, then record the actions they take and the confirmation each action returns.

Use that trace to define the integration requirements. The AI may need an authorized lookup of order items and delivery status, a way to create or update a return record, and a carrier action that generates a label. Check which of those operations your actual APIs support, what inputs and permissions they require, and how the workflow will recognize a successful result. A field appearing in an ERP screen does not establish that the integration can read or change it.

Test delayed and partially shipped orders as well as the easy cases. A schema containing a delivery-date field does not mean every relevant order has a date.

Set freshness by the request

How old can a record be before the answer becomes misleading? An invoice copy and a promise about stock availability may need different answers.

During a Typewise proof-of-concept kickoff, a wholesaler's IT contact judged a daily export sufficient for invoice and shipment questions about events already days or weeks old. The proposed workflow would hand over when it could not find a record. This was an interim pilot choice, not a validated production service level.

For your workflow, specify the maximum acceptable age, who refreshes the data and what happens when a refresh fails. Measure handovers caused by missing or stale records. If those grow, revisit the cadence rather than treating the pilot compromise as permanent.

Confirm what a value means

"In stock" does not necessarily mean "available to this customer tomorrow."

The same wholesaler, while evaluating Typewise, explained that large-customer delivery commitments required coordination with its internal teams. A shipment-tracking question was different: staff could look up the delivery note and the tracking number behind it. The first needed judgment and coordination; the second was a direct lookup.

Ask a service expert to explain what each important field allows them to conclude, and what it does not. Record the additional conditions needed before making a promise. Test those conditions with an engineer, using the values the system actually returns.

Find the authoritative source

For each fact, name the source your team relies on when records disagree. Customer-specific terms or product-code mappings may sit outside the main business system.

If a local file is authoritative, decide how the workflow will read it, who maintains it and how changes are checked. Moving data into a new tool is not enough if nobody owns its accuracy. Our guide to mapping customer-support sources covers the wider source hierarchy.

Enforce access before serving the data

Verify identity and authorization in the system that returns the records. A logged-in customer should see only the records they are entitled to see. An instruction asking the AI to respect that boundary is not a substitute for enforcing it.

Test anonymous requests, the correct account and a different customer's account. A failed access check must block the data, including drafts or recommendations that would reveal it. Keep the read and write permissions separate; the permissions guide provides a practical starting point.

Generate and confirm the return label

In the wholesaler's pilot discussion, its logistics contact questioned the value of drafting an email when he would still have to generate the return-label PDF himself. Generating the label and writing a short reply was already quick for him. The team parked that case while the carrier connection was missing. This was his assessment of the workflow, not a timing experiment.

For each candidate, list the final actions, whether they are technically possible, whether they are allowed and what human work remains. Include a fallback for a missing record, a failed lookup and an unconfirmed action.

Name who builds each connection

Readiness depends on people as well as data. The lookups and actions in your trace may be built or approved by someone other than the AI provider: your own IT team, a CRM partner or the company that runs a system for you. Their capacity, data-protection reviews and approval process can set the pilot's timeline.

For each connection, record who builds it, who approves it and when it will be available. Where an outside provider runs the system, or the data still needs a privacy review, start that conversation when you choose the workflow rather than after the pilot is designed.

Some information staff rely on may sit where no interface can reach it, such as notes in a free-text history or a log rather than a field. Treat that as a gap in the trace. Decide whether the workflow can do without it or whether a person keeps that step.

Then check each connection end to end. Send a test request and confirm what actually arrives: the record, the case history and the fields the AI needs. A connection marked as done in a project plan has not yet shown that the expected context reaches the AI.

Agree with the service owner how much case history each request type needs. Fetching everything for every request can slow the workflow, and the right depth is a service decision as well as a technical one.

Check the return path too. When the AI closes, ignores or reroutes a case, that status has to reach the system where your team tracks open work. Otherwise someone still has to close those cases by hand.

Bring one service expert and one integration engineer together around a real request. If they cannot trace it through these checks, keep the gap explicit in the pilot scope rather than marking the workflow ready.