Researching a customer account.

Goal: Produce a one-page, sourced brief on one customer account that you would be comfortable reading aloud to that customer. Who it is for: Account managers, business development staff and owners who sell to other organisations. Time: 30 minutes, plus one real prospect from your pipeline and an AI assistant with web search turned on.

The test in that goal is deliberate. AI sales research rarely fails by being empty. It fails by being confident: the brief reads well, and one detail in it (a chief executive who left last year, a merger that was only rumoured, a product line that was dropped) ends up in your opening email. The customer notices before you do. This lesson is about producing less and being right about all of it.

What a useful account brief contains

A useful account brief answers four questions: what the organisation does, what has changed recently, who is likely to be involved in a decision like yours, and what you still do not know. Anything beyond that is decoration that someone has to check.

The unknowns row matters most. Ask for a complete picture and an assistant will fill the gaps with plausible guesses; ask what it could not confirm and you get your first call agenda.

Exercise: build a one-page brief on a real prospect

1. Pick one account and one reason. Choose a prospect you plan to contact in the next fortnight and write one line on why, such as “they are adding cool-store capacity and we service refrigeration”. The reason keeps the research narrow.
2. Start a fresh chat with web search on. Use ChatGPT, Claude, Copilot or Gemini with browsing enabled. A fresh chat stops earlier conversations leaking in. Do not paste CRM notes or contact details at this stage.
3. Run the research prompt. Adapt the prompt in the worked example below: company name, website, your company and your reason. Keep the rules section as written.
4. Open every link. Delete any claim whose source does not say what the brief says, is about a different company with a similar name, or is older than 18 months for a “recent change”.
5. Confirm the registry details yourself. Search the Companies Office register for the legal name and check that it is registered and matches the website. Name collisions are the most common error in AI company research.
6. Cut it to one page. Keep the four sections and the unknowns. Add the single question you most want answered in the first conversation.
7. Decide what goes into your CRM. Store organisation facts with their sources. Store only the people details your next step needs, with a note of where each came from.

Worked example: a regional packhouse

Illustrative example: a Christchurch refrigeration servicing firm preparing to approach a Hawke’s Bay fruit packhouse. Both businesses are fictional.

You are helping me prepare for a first sales conversation.

Company: Kōwhai Valley Packhouse Limited (website: kowhaivalley.example.nz)
My company: Southern Cold Services, Christchurch. We service and upgrade
cool-store refrigeration for food producers.
Reason for contact: they appear to be expanding cool-store capacity.

Research this company using public web sources only. Produce:
1. What they do: 3 bullets.
2. Recent changes in the last 18 months: up to 4 bullets, each with a date.
3. Roles likely involved in a refrigeration servicing decision.
   Roles only. No names, no personal social media.
4. Possible fit with our service: 2 bullets, each tied to a sourced fact.
5. Unknowns: what you could not confirm.

Rules: put a source URL after every bullet. If a fact has no source, leave
it out. If sources disagree, say so. Do not estimate revenue, staff numbers
or customers. Use plain New Zealand English.

Example output (illustrative, abridged):

2. Recent changes
- March 2026: resource consent application lodged for a new cool-store
  building at its Hastings site. [regional council notice URL]
- Sources disagree on the second packing line: the website says it opens
  "in 2026"; a June 2026 news item says commissioning is delayed.
  [two URLs]

3. Roles likely involved
- Site or operations manager; engineering or maintenance lead;
  finance, if the work is capital spend. No names collected.

5. Unknowns
- Current refrigeration contractor.
- Size of the planned cool-store and whether consent has been granted.

What makes this output good: dated changes, a disagreement surfaced rather than guessed away, roles instead of people, and unknowns you can turn into questions.

Why roles, not personal profiles

Researching an organisation is low risk. Researching its people is where New Zealand privacy law starts to apply. The Privacy Act 2020 covers information about identifiable people, and the Privacy Commissioner’s guidance on AI and the information privacy principles notes that this includes information that is inaccurate or made up, which covers an AI’s wrong guess about someone’s job history.

Since 1 May 2026, IPP3A also applies when you collect personal information from someone other than the person: a website, an enrichment tool or a colleague. Unless an exception applies, you must take reasonable steps to make the person aware of the collection, its purpose, who will receive it, who you are, and their rights to access and correct it (“for business purposes” is not specific enough). In practice: collect roles first, record the source of any named contact, and make sure your first message and privacy statement explain where the details came from. Our lesson on protecting customer information when using AI covers the wider rules.

Where AI sales research goes wrong

• Name collisions. Two companies with similar names in different regions get merged into one profile. Match the website domain and the Companies Office entry before trusting anything.
• Old news presented as current. A two-year-old expansion story becomes “they are expanding”. Require a date on every change and apply your cut-off.
• Invented specifics. Revenue, headcount and customer lists are the figures assistants most often estimate. Forbid estimates in the prompt and delete anything unsourced. The method in checking AI answers against sources applies line by line.
• Volume over accuracy. GTM-Bench, a 72-task benchmark for AI agents doing go-to-market research co-authored by Levy Street team members, scores each returned record plus one, zero or minus one so that weak records count against the agent. The published paper reports many generalist agents finishing with negative net scores. The lesson holds for people too: ten verified facts beat forty unverified ones.
• Leaking your own pipeline. Pasting a full CRM export into a personal AI account to “add context” exposes customer data. Keep public research and private context separate.

At dozens of accounts a week, purpose-built systems such as Bebop with Blackpearl Group (AI lead discovery, enrichment, research agents and sales playbooks) or a custom build from our AI agent development team become worth testing. Connecting research to HubSpot or another CRM comes later in this pathway.

Self-check

• The brief fits on one page and states why you are contacting this organisation.
• Every bullet has a source you opened, and the source says what the bullet says.
• Every “recent change” carries a date within your cut-off.
• The legal name matches the Companies Office register and the website.
• People appear as roles, or as named contacts with a recorded source and purpose.
• The unknowns section lists at least two things you will ask rather than assume.
• Nothing in the brief would embarrass you if the customer read it.

Next lesson: Creating a business proposal with AI.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.