Picking which accounts get a seat: account-based events with AI

Field marketers already own account-based marketing. Forrester found that 79% of them are responsible for ABM strategy and 73% are measured on pipeline or revenue influenced (Forrester, 2025). Yet the guest list for most hosted events is still built the old way. A sales leader forwards some names, someone pulls last year's registrants, and the room fills with whoever replied first.

That was tolerable when rooms were big. They are getting smaller. In Forrester's Q1 2026 survey, 63% of B2B organizations plan more hosted networking events under 20 attendees, against 18% planning more large hosted events over 200 (Forrester, 2026). And 97% of B2B events leaders say securing the right attendees is a priority (Forrester via Marketing Week, 2025).

When a room holds fewer than 20 people, every seat is an ABM decision. This workflow treats it as one: a ranking problem over accounts, with a layer of human judgment on top. AI is good at the first part and bad at the second.

Step 1: Build one row per account from four inputs

Start at the account, not the contact. Forrester's 2026 buying research puts a typical purchase at 13 internal stakeholders and 9 external influencers (Forrester, 2026). You choose the companies first, then decide which of their people belong at the table.

Pull four sources into a single sheet:

• Target account list. The ABM tier agreed with sales, plus the account owner.
• Intent signal. Whatever your team uses: a third-party intent score, web visits, content engagement. Record the source in its own column, because these are modeled signals, not observed behavior.
• Open pipeline. Opportunity stage, amount, expected close date and the contacts attached to each opportunity, exported from HubSpot or Salesforce.
• Past attendance. Every event the account's people registered for and actually attended over the last 18 months.

The join is where the hours go. Company names will not match across files. Some guests registered with personal email addresses. Some contacts have changed jobs since they last came. Match on email domain where you can and fix the remainder by hand. Validity found that 76% of organizations say less than half of their CRM data is accurate and complete (Validity, 2025), so expect this step to take longer than every AI step combined.

The output is one row per account with columns for tier, intent, opportunity stage and amount, number of contacts on the opportunity, events attended, date last attended and owner.

Step 2: Write the rules before the model sees anything

Decide what the event is for. A dinner meant to move late-stage deals needs a different list from one meant to open doors at tier-one accounts with no pipeline. Then write two or three hard rules, for example:

• Cap any single account at two seats.
• Reserve at least a third of seats for accounts with no attendance history, so the list does not simply recycle familiar faces.
• Exclude customers in an active renewal dispute.

These rules are yours. Do not ask the model to infer them from the data.

Step 3: Have AI rank and explain

Upload the table with a prompt along these lines:

You are helping choose accounts for a hosted dinner for [N] guests
in [city] on [date]. The goal of this event is [move open
opportunities in stages 3 and 4 / start conversations with tier 1
accounts that have no open pipeline].

The attached table has one row per account. Columns: [list them].

Hard rules:
1. [rule]
2. [rule]
3. [rule]

Rank the top [2x N] accounts. For each, give:
- Rank and account name
- A two-sentence rationale that cites only columns in the table
- Any rule it came close to breaking
- Confidence (high, medium, low) and the reason

Use no information outside the table. If a field is blank, say it
is blank. Do not guess job titles, deal status or company news.

Ask for roughly twice as many accounts as you have seats. Sales will veto some and guests will decline others. Even executive dinners, the best-attended format in one vendor dataset, converted 70% of registrants into attendees (Clutch Events, 2026). Plan the long list with that in mind.

The rationale column matters more than the rank. A ranked list without reasons is a black box your account executives will ignore. A ranked list with reasons is a draft they can argue with.

Step 4: Make the model argue with itself

Run a second prompt on the same conversation:

Review your ranking and report:
1. Accounts in the top 15 that rank there mainly because of one
   strong signal. Name the signal.
2. Accounts outside your list with open opportunities above
   [amount]. Explain why each fell out.
3. Any pattern suggesting one column is over-weighted, such as most
   accounts sharing a region, an owner or an intent source.

This catches a common failure: a single column dominating. An account whose people downloaded three reports last week can outrank an account with a late-stage deal if the model reads recent activity as importance. You want to see that happen on screen, not discover it at the table.

Step 5: Choose the people, then hand it to sales

Now go down to contacts. For each shortlisted account, list who is attached to the opportunity and who has attended before. In a Forrester client story, opportunities with multiple contacts attached were eight times more likely to advance than single-contact opportunities (Forrester, Palo Alto Networks client story, 2024 to 2025). Inside each account, the useful question is which member of the buying group is missing from the relationship. The account owner answers that, not the model.

Send the ranked list, with rationales, to each owner. Ask one question per account: invite, skip, or invite someone else. Give them two working days.

This is where knowledge no dataset holds comes in. The owner knows the champion just resigned. They know the economic buyer and the champion do not get along, that two guests from rival firms will not speak candidly in front of each other, and that an account on the edge of churning would read an invitation as the wrong gesture. None of that is in the model. Most of it is not in the CRM either.

The review is also an alignment exercise. In Demand Gen Report's 2025 ABM benchmark, 43% of practitioners reported sales and marketing alignment challenges (Demand Gen Report, 2025). A list that sales has marked up before invitations go out is a list sales will follow up on afterward.

What to check

• Every rationale against its source row. Models sometimes cite a figure that is not in the table or blend two similar accounts. Check at least ten rows by hand, and all of the top five.
• Titles. Never let the model state or infer a current job title. Confirm with the owner or a public profile.
• Intent-only picks. If an account made the top ten on intent alone, ask the owner whether the signal matches anything they are seeing.
• The shape of the room. Scan the final list for balance by stage, industry and seniority. A table of prospects with no customers in the room has nobody to vouch for you.

Where it goes wrong

The ranking inherits the bias of its inputs. Past attendance rewards accounts you already know, which is why the reserved seats in step 2 exist. Stale CRM data produces confident rationales about deals that closed or died months ago. And no model can tell you who will make the evening work.

Keep the AI's ranked list and the final guest list side by side. After the event, compare both against who attended and which opportunities moved in the following 30 days. The gap between what the model suggested and what your owners chose is the most useful record you will have when you build the next list.