Casebook · Venue & Park Leasing · China

Conversational CRM for venue sponsorship sales: one sentence in, a preview out, saved only after confirmation

Pilot · DemoRecords as of 9 Oct 2026BEE Sigma Delivery Team

The client is a Chinese company that sells park and venue leasing and sponsorships. We connected AI to Lark (Feishu) groups and Lark Base: a salesperson describes a visit in one sentence, the agent parses it, checks for duplicates and previews seven CRM fields, and the record is saved only after they reply “confirm and write”. MVP v3 was built and demoed on 4 Sep 2026; as of the 10 Sep 2026 records, the client has not formally accepted it.

IndustryPark and venue leasing, sponsorship sales
MarketMainland China
Delivery statusPilot · Demo (MVP v3 demoed, not yet accepted by the client)
TimingStarted Jul 2026; MVP v3 demoed 4 Sep 2026; latest test round 10 Sep 2026
Systems connectedLark groups, Lark agents, Lark Base (client / project / follow-up tables); only 2 bots remain on DingTalk
Decisions kept by peopleEvery CRM write, contact record and decision-chain diagram runs only after the requester confirms it
01

How it worked before

Follow-ups lived in chats and in salespeople's heads. CRM fields sat empty, and history was spread across each salesperson's own Excel file, with different columns and headers. The old records sat in a DingTalk CRM plug-in that was slow to fill in, so front-line sales rarely completed it.

Before and after
Before and after
02

What AI does now, and what people do

The main platform is Lark (confirmed 20 Aug 2026; only 2 bots stay on DingTalk). AI turns a sentence into a structured preview; people decide whether it belongs in the system.

  • AI does: parses the visit description, works out which salesperson owns it, checks for duplicates and previews seven fields; tags contacts with five roles and relationship tiers; draws a decision-chain diagram on request.
  • People do: check the preview and reply exactly “confirm and write” before anything is saved to the linked client, project and follow-up tables; contact records and decision-chain diagrams need the same confirmation.
Process: what the AI does and what people do (gold = people)
Process: what the AI does and what people do (gold = people)
03

Timeline

Dates below come from the project records.

  • 11 Jul 2026Strategy kickoff.
  • 15 Jul 2026First demo of conversational capture, writing to the three tables (demo environment).
  • 24 Jul 2026Main flow confirmed working; test-data isolation rule added the same day.
  • 14 Aug 2026Decision-chain module designed (contact roles and relationship tiers); manual entry then took about 3–4 minutes per client, logged as a pain point.
  • 20 Aug 2026Lark confirmed as the main platform; 2 bots kept on DingTalk.
  • 1 Sep 2026New CRM MVP demoed with real leads as test input.
  • 4 Sep 2026MVP v3 built and demoed; not formally delivered.
  • 10 Sep 2026One-click decision-chain diagram passed testing: 8 contacts and 5 direct reporting lines in one run, about 4 minutes.
Timeline: every point comes from project records
Timeline: every point comes from project records
04

Current status and limits

This is a pilot at demo stage. MVP v3 has been demoed, but as of the 10 Sep 2026 records the client has not accepted it and it is not rolled out to the whole sales team. The records contain no conversion, deal or time-saved figures, so we don't claim any. Project-stage review is still being planned.

05

What we learned

Conversational capture can mistake a test for a real visit: an invented sample once produced a real-looking preview, so from 24 Jul 2026 test samples are isolated by default. A “saved” message must match what is actually in the table, or salespeople stop trusting it. And model choice should rest on data: we ran a cost-and-quality comparison on 26 Aug 2026.

Scope note: an anonymised case is a summary of project records with client names, people and amounts removed. This page draws on those records; client names, people and amounts are withheld. Dates and figures come from project records and some have not been confirmed in writing by the client. The delivery status is as labelled at the top; results vary with each business’s systems, data and involvement and are not a promise to other businesses.

FAQ · Quick answers

Is this conversational CRM live?

Not yet. MVP v3 was built and demoed on 4 Sep 2026 and tested again on 10 Sep 2026, but as of those records the client has not formally accepted it, so we label it Pilot · Demo.

Why must sales reply “confirm and write” before anything is saved?

When AI parses one sentence it can get the client, project or contact wrong. A seven-field preview plus an exact confirmation keeps mistakes out of the CRM, and every saved record has a person who approved it.

Why Lark rather than DingTalk?

The old records were in DingTalk, but this round's group-chat capture, agents and Base tables were built on Lark. Lark was confirmed as the main platform on 20 Aug 2026; only 2 bots remain on DingTalk.
More anonymised cases
← All anonymised cases

Hand your first workflow to the agents

One free scan shows how visible you are in the AI era; the AIM assessment finds your best angle of adoption.