How to build a better invite list with AI, starting from last year's export

The small room is where B2B event budgets are heading. In Forrester's Q1 2026 survey, 63% of B2B organizations said they plan more hosted intimate networking events under 20 attendees, the highest share of any format, while only 18% plan more large hosted events over 200 (Forrester, 2026). A year earlier, 97% of B2B events leaders said they prioritize securing the right attendees (Forrester, 2025).

At a 300-person conference, a few wrong invitations disappear into the crowd. At a 16-seat dinner, each one is a chair a target account did not get. The invite list stops being admin and becomes the strategy.

Most teams still build it the same way: a target account list from marketing, a round of "send me your top five" to sales, and whatever the field marketer remembers from last time. The richest data the team owns, the record of who actually attended past events, usually sits unused in old exports.

This workflow starts there.

Step 1: Pull attendance, not registration

Export every event from the last 12 to 24 months into one sheet. One row per person per event, with these columns: name, work email, company, event name, event date, event format, and status (attended, no-show, or cancelled in advance).

Status is the column that matters. A registration is an intention; an attendance is a behavior. Vendor data from Clutch Events, covering 191 events and 13,155 registrations, found attendance rates ranging from 70% at executive dinners to 19% at vendor-owned conferences (Clutch Events, 2026). A list built on registrations blends people who came with people who never intended to.

If your check-in data lives on paper or in someone's memory, this is where you find out. Fix capture for the next event before you worry about anything downstream.

Step 2: Join it to the CRM

Match each row to a CRM contact and account. Bring back the account name, account owner, target account flag, lifecycle stage, and any open opportunity with its stage.

Two rules apply here. Keep unmatched rows on their own tab, because a guest who attended twice and is not in the CRM is a finding, not an error. And match with deterministic rules first (exact email, then email domain plus name), reviewing the fuzzy cases by hand. Do not hand the matching itself to a chatbot.

Expect the join to be messy. 76% of organizations say less than half of their CRM data is accurate and complete (Validity, 2025), and 48% of CRM admins saw customer data decay faster over the previous year (Validity, 2024). Both figures come from a data quality vendor, and both will feel conservative once you start matching.

Step 3: Score it with rules you can explain

Before any AI touches the list, build a simple score in the spreadsheet. For example:

+3  open opportunity at stage 2 or later
+2  target account
+2  attended 2 or more past events
+1  attended 1 past event
+1  director level or above (CRM title, as recorded)
-2  no-show 2 or more times without cancelling

The weights are yours to argue about, and that is the point. A score your sales leader can read in ten seconds survives the review meeting. A black-box score does not.

The seniority line has some support in vendor data. Clutch found that directors and above attend at a 62% rate, and that technical practitioners show three times the no-show rate of senior leaders (Clutch Events, 2026). Whether that should move your list depends on who you actually need in the room.

Step 4: Have AI draft tiers and rationales

Now AI is useful. Its job is to sort a scored table into tiers, write a one-line reason for each contact, and catch coverage gaps a tired human would miss. Researching people or "improving" the data is outside its job.

Strip personal details first. Pass a contact ID, company, CRM title, target flag, opportunity stage, attendance counts and score. Leave emails and phone numbers out.

You are helping plan the guest list for a 16-seat executive dinner.
Below is a table with these columns: contact_id, company, crm_title,
target_account (Y/N), opp_stage, events_attended, no_shows, score.

Sort every contact into one of three tiers: Invite first, Waitlist,
Not this time. Put [N] contacts in Invite first.

For each contact, write one sentence explaining the tier, using only
the columns provided.

Rules:
- Use crm_title exactly as given. Do not add, correct or infer titles.
- If a field is blank, write "unknown". Do not guess.
- No more than two Invite first contacts from the same company.
- Do not reference any information that is not in the table.

Return a table: contact_id, tier, rationale.

Set N with your own acceptance and show rates in mind. If Clutch's executive dinner figure holds for your audience, roughly three in ten registrants will not make it on the night.

Then run a second pass focused on accounts rather than people:

Using the tiered table above, list every company with an opp_stage
of 2 or higher. For each, show the number of Invite first contacts.
Flag any company with zero. Then list any company with three or
more contacts in Waitlist and none in Invite first.
Do not add information that is not in the table.

This pass matters because a typical buying decision now involves 13 internal stakeholders and 9 external influencers (Forrester, 2026). The unit that matters is the account, and a contact-level score can quietly leave a live opportunity with nobody at the table.

Step 5: Review with sales in the room

Take the draft to the account owners with the tiers and rationales attached. Ask them whether anyone listed has left the company, whether there is a relationship reason to add or remove someone, and whether the right person from each open opportunity is included.

This is the step AI cannot do. It does not know that a champion just resigned, that two guests from rival firms should not share a table, or that an executive had a bad support experience last month. The account owner usually does.

What to check before invites go out

• Titles. Language models will quietly update or embellish job titles. Compare every title in the AI output against the CRM field. Any mismatch is the model's invention.
• Stale data. A title copied faithfully from the CRM can still be two years out of date. For everyone in Invite first, confirm current role on a public profile or with the account owner.
• Invented rationale. Read the one-line reasons. If one mentions something that is not in your columns, such as a funding round or a product interest, the model made it up. Delete it and rerun.
• Coverage math. Count Invite first contacts per open opportunity yourself. Models miscount.
• The unmatched tab. Review people who attended but are not in the CRM before you finalize. Some of them belong on this list, and all of them belong in the CRM.

Doing it again next quarter

The first three steps are the part most teams skip, because rebuilding the join by hand every quarter is tedious. Socially covers that data work: it syncs event attendance against HubSpot or Salesforce, flags attendees who are not in the CRM yet, tiers contacts by real attendance behavior across events, and surfaces who is worth inviting next based on past attendance and CRM data. The AI drafting and the human review in steps 4 and 5 stay with your team.

The list you end up with will not be perfect. It will be explainable, which matters more when a sales leader asks why their account did not get a seat. And next year's export will be better than this year's, because this time you recorded who came.