How to Automate B2B Order Desk Calls: Orders, Truck Status, Claims and Reorders

Automating a B2B order desk by phone means a voice agent answers the calls your order entry team takes every day (new orders, "where's my truck" status checks, damaged-goods claims and scheduled reorders), writes the result into your ERP or order system through your API, and hands anything that needs judgment to a person with the details already captured.

For manufacturers and distributors, the order desk is often a large team keying phone orders by hand, with the heaviest volume in the first hours of the morning. The calls repeat, the data is structured, and a wrong unit or a missed add-on turns into a wrong shipment and a credit. That makes the desk a good fit for a voice agent, as long as each call type is designed on its own.

Which order desk calls can a voice agent take?

Four call types cover most of a B2B order desk, and each one touches a different system and has a different point where a person takes over.

Start with the call type that has the most volume and the cleanest data behind it. For many desks that is status calls, because they only read from your system.

How does the agent take an order without keying errors?

It identifies the account first, confirms every line with the unit said out loud, and reads the whole order back before writing anything. In Retell, an inbound webhook fires before the call connects and includes the caller's number, so your system can find the account and pass its details into the call as dynamic variables. During the call, custom functions call your API to check products, prices and stock, and the agent can save fields from each response for the rest of the conversation.

Units cause many order desk errors. "Ten of the fifties" can mean ten bags, ten pallets or ten cubic meters. Write the agent's instructions so it:

• asks for the unit whenever the caller leaves it out, using the units your catalog sells in
• repeats each line as quantity, unit and product
• asks whether anything else goes on the order before moving to delivery, because add-on items are easy to lose
• offers only the delivery dates and windows (morning or afternoon) your system returns
• reads the full order back and waits for a clear yes before it writes the order

Turn on Talk While Waiting for lookups so the caller hears a short line such as "let me check that" instead of silence. Product, plant and job site names can be added as boosted keywords so transcription picks them up.

Retell does not ship a prebuilt ERP connector. Orders reach your ERP through a custom function, an MCP server you already run, or a webhook after the call. Keep live stock and pricing out of the knowledge base: auto-refresh runs every 24 hours and spreadsheet files are capped at 1,000 rows, so anything that changes during the day should come from an API lookup.

How should "where's my truck" calls work?

The agent looks up the order in your system and reads back only what the system returns. Most status calls come from known accounts, so it can find open orders by the caller's number and ask which one the caller means when there are several. A custom function returns the fields you choose to share, such as order status, ship date and delivery window.

Tell the agent never to estimate an arrival time on its own. If the system shows no status, or the delivery is past its window with no explanation, the agent should open a ticket or transfer the call. The reasons people still phone about orders they can already track are covered in WISMO: why "where is my order" is still a phone problem.

How do you handle damaged-goods claims and credit requests?

Let the agent take the claim, not decide it. A complete claim record includes:

• order or PO number, and the delivery date
• each item and quantity affected
• what happened, and whether it was refused at delivery or found later
• whether the damage was noted on the delivery paperwork
• the best contact for follow-up

The agent then creates a ticket through a custom function, or through the Zendesk integration tool if your claims already live there. With the caller's consent, it can text a link to your photo upload form during the call, once your number is registered for texting. The credit decision stays with your team. Claims above a value you set, or any dispute, go to a person by warm transfer, and the agent can brief that person privately before the caller is connected. See how call transfers work between AI voice agents and human teams.

How do you absorb the morning call spike?

Plan capacity for the peak hour, not the daily average. Pay-as-you-go workspaces include 20 concurrent calls, and Enterprise plans set concurrency to your volume. With concurrency burst turned on, calls above your limit still connect, up to the lower of three times your limit or your limit plus 300, at an extra $0.10 per minute for those calls. If every slot is busy, an inbound call waits about 40 seconds for one to open, then goes to the fallback number you set.

If you also run outbound reorder calls, reserve inbound concurrency so a batch can never crowd out the morning order line, and schedule batches outside the morning peak. Test the full call path at peak volume before launch, as described in voice AI load testing on the real call path.

What should change after hours?

After hours, the agent should stop trying to transfer and take a ticket instead. A transfer to an empty queue only adds hold time. Because the inbound webhook can override which agent answers each call, your system can send after-hours calls to a version with no transfer step that creates a ticket and promises a callback during business hours. You can also give the agent the current time with the {{current_time}} variable and write the rule into its prompt.

Can the agent check inventory or book a dock slot during the call?

Yes, when those systems have an API. A custom function can check stock at a plant or warehouse, and a second one can list open dock doors and book a slot for a pickup. If the API is not ready, start manual: the agent collects the request, a webhook sends it to your team's inbox or sheet after the call, and staff confirm it. Swap in the live lookup later without changing what callers hear.

Does it work in Spanish and English?

Yes. Spanish is supported for speech recognition and voice in both Spain (es-ES) and Latin American (es-419) variants, and a multilingual agent can follow a caller who switches between English and Spanish mid-call. Every order is written to your system in the same format whatever language the call used. Test regional product terms in your simulated calls before going live. More on Spanish voice agents.

How do scheduled reorder calls to store locations work?

The agent calls each location on a schedule, takes counts, compares them to the last order and writes the confirmed reorder. Batch calls take a CSV with one row per location, and any extra column, such as store name or last order quantities, becomes a dynamic variable in the call. A good reorder call:

1. Confirms who is speaking and whether they can give counts. If your retail accounts require a passcode for vendor calls, build it into the opening.
2. Asks for counts one item at a time.
3. Waits when the person says "hold on, let me check." Prompt the agent to stay silent in that moment, and raise the silence settings: by default the agent reminds a quiet caller after 10 seconds and ends the call after 10 minutes of silence.
4. Compares each count to the last order and asks for a recheck when a number looks far off.
5. Writes the order through a custom function and reads back the delivery date.

Store phone menus can be handled with the Press Digit tool. Confirm with your counsel which calling rules apply to your outbound program, even for business lines. The fuel and cylinder version of this call is covered on propane delivery ordering.

How do you roll it out and measure it?

Start with one call type on one line, prove it on test calls, then expand. A practical sequence:

• Write test cases from real calls, including wrong units, unknown products and callers with two accounts. Simulation tests can mock your custom functions, so nothing is written to the ERP during testing.
• Run the full set as a batch test after every prompt change and compare pass rates.
• Define post-call extraction fields: call type, order written (yes or no), line count, claim opened and transfer reason.
• Set alerts by email or webhook for when call success rate drops.
• Add the next call type only when the first one holds.

Retell's AI phone ordering page shows how teams connect catalogs and order APIs and start on a single order line. Pay-as-you-go pricing runs $0.07 to $0.31 per minute with $10 in free credits, so you can test your own order desk calls before talking to sales.