How to research a commercial prospect before the first meeting (with AI)

Why the first meeting is won before you walk in

Most commercial clients still buy through a broker. In Australia, brokers placed 46% of all general insurance written in FY25. About four in five small and medium businesses use one (Insurance Business, June 2026). In the UK, brokers write the large majority of commercial premium. That's the good news. The less good news is that the business owner sitting across from you has heard the pitch before.

What wins the room is simple: showing you already understand the business. Not a generic "tell me about what you do", but "I saw you've started exporting to the US, so let's talk about products liability". Buyers are tired of sellers who haven't done the homework. In Gartner's 2025 survey of B2B buyers, 73% said they actively avoid suppliers who send irrelevant outreach (Gartner, June 2025).

There's a second reason. Anyone who's worked in underwriting will feel it. Every proposal form asks the same twenty or so things: what the business does, how long it's traded, who runs it, where it operates, who it works for. Half of those answers are public. Finding them before the meeting means you spend the meeting on the half that isn't.

Step 1: pin down who you're actually meeting

Start with the boring bit, because everything else hangs off it. The name on the shopfront is often not the name on the policy. "Kestrel Engineering" might be a trading name of a family trust, a company with a slightly different name or one of three related companies. Insure the wrong entity and you've got a problem the day there's a claim.

Both countries give you the basics free, with no login. Australia has 3.8 million registered companies (ASIC, October 2026) and 2.7 million actively trading businesses (ABS, June 2025). The UK register holds 5.5 million companies (Companies House, June 2026). Plenty of them have similar names.

Red flags worth a question

• Status isn't active. Deregistered, in external administration (AU) or with a proposal to strike off (UK).
• A recent name change. Ask about claims under the old name.
• Overdue accounts or confirmation statement (UK). Often admin, sometimes cash flow.
• Charges registered against the company. Lenders may need noting on the policy. It also tells you about debt.
• Directors with a trail of dissolved companies. Not a verdict, but worth understanding.
• The trading name doesn't match any registered entity. Find out exactly who you'll be insuring.

Step 2: read the signals

Once you know who they are, look at what they're doing. Businesses leave a surprising trail: a job ad for night shift machinists tells you about wages and hours, a planning application tells you about a new building. A work safety prosecution tells you something no proposal form will volunteer. Here are the sources we'd check, roughly in order, with the cover each one touches.

A few signals tend to mean the cover needs a fresh look. Cyber is the obvious one: in the UK, 47% of businesses have some cyber insurance but only 10% hold a standalone policy (UK Cyber Security Breaches Survey 2025/26). In Australia, take-up is lower still: industry data cited during a parliamentary inquiry in November 2026 puts it at fewer than 4% of small businesses (Insurance Business, November 2026). A website that takes online orders or stores customer records is a conversation waiting to happen.

Step 3: let AI do the legwork, safely

Everything above is a solid hour by hand. An AI tool with web search can get you most of the way in a few minutes. ChatGPT with search or deep research, Claude, Perplexity and Microsoft Copilot can all read websites and registers and summarise what they find. The catch is that they're fluent whether or not they're right, so the prompt has to do the heavy lifting.

Six rules that make AI research trustworthy

1. Identity first. Give it the legal name and the ABN or company number. Make it confirm the entity before anything else. If more than one business could match, it should list them and stop.
2. A source and a date for every fact. No link, no fact.
3. "Not found" beats a guess. Tell it, in so many words, never to estimate an identifier, a figure or a date.
4. Keep facts and guesses apart. Ask for three columns: verified (with source), inferred (with reasoning) and ask the client.
5. Ask for contradictions. Website says 30 staff, LinkedIn says 80? You want to know.
6. Make it check its own work. Finish with a second pass that removes anything its sources don't support.

That's a lot to type each time, so we built it for you. Fill in what you know and copy the prompt into whichever AI tool your brokerage allows.

You're helping a commercial insurance broker in Australia prepare for a first meeting with a prospective client. Research this business using web search.

Business: [business name]
ABN or ACN: not given
Website: not given
Location: NSW, Australia
What they do (as I understand it): not given
Covers I want to discuss: Liability, Property and BI

Rules:
1. First confirm the exact legal entity on ABN Lookup and ASIC. If more than one entity could match, list them with their ABN or ACN and stop. Never guess or reconstruct an ABN or ACN.
2. Give a source link and the date of the source for every fact. If you can't find something, write "Not found". Don't estimate figures, dates or headcounts.
3. Sort what you find into three groups: Verified (with source), Inferred (with your reasoning) and Ask the client.
4. List anything that doesn't match between sources, such as the website against the register or LinkedIn.
5. Treat anything older than two years as possibly out of date and say so.

Check these sources where they exist:
- ABN Lookup (abr.business.gov.au): entity name and type, ABN status and date, GST, business names, location
- ASIC company search (connectonline.asic.gov.au): status, registration date, former names
- SafeWork NSW prosecutions and enforceable undertakings
- the NSW environment regulator's public register
- AusTender contract notices (tenders.gov.au)
- AustLII (austlii.edu.au) for court and tribunal decisions
- job ads on Seek and LinkedIn, Google reviews and recent news
- the Wayback Machine (web.archive.org) for older versions of their website

Then give me:
- A short summary of what the business does, where, for whom and how big it looks.
- A table with columns: line of cover | what you found | source | confidence | question to ask in the meeting. Cover: Liability, Property and BI.
- The ten things you could not find that an underwriter will ask about.

Finally, re-check every fact against its source and remove anything the source doesn't support.

Check what the AI tells you

AI tools get things wrong in very consistent ways. The evidence is piling up. When the EBU and BBC tested AI assistants on news questions in 2025, 45% of answers had at least one significant issue, often with sourcing (EBU and BBC, October 2025). In Australia, Deloitte agreed to refund part of a $440,000 government contract after its report was found to contain made-up references and a fabricated court quote (The Register, October 2025).

The worrying part is us, not the machines. In the University of Melbourne and KPMG global study of trust in AI, 66% of people who use AI at work rely on its output without checking it (University of Melbourne and KPMG, 2025). In broking, an unchecked fact can end up in a submission.

The mistakes to look for

• Invented identifiers. An ABN or company number that looks right but isn't. Copy identifiers from the register yourself, never from the AI.
• Namesakes. A similar business in another state or a related company with a different number. Match on the number, not the name.
• Stale pages. A director who left years ago, a site that closed, a product they stopped making. Check the date on every source.
• Confident filler. Turnover "estimated" from nowhere. If it isn't sourced, it's a question for the client.

Want to test your eye? Each of these AI research notes about a made-up business has one mistake in it. Tap the line you wouldn't trust.

Keep client data out of free AI tools

Researching a prospect from public sources is one thing. Pasting last year's schedule, a claims list or a director's details into a free chatbot is another. The regulators in both countries have been clear about where the line sits.

• Public information about a business is fine to research with AI: its website, registers, news and tenders.
• Client documents and personal information belong only in tools your brokerage has approved, with a business agreement that covers how data is stored and whether it's used for training.
• Write it down. A one page AI policy that says which tools staff can use and for what protects the brokerage as much as the client.

Turn your research into the first meeting

Good research doesn't replace the meeting. It changes it. Instead of an hour of "so what do you do?", you confirm what you found in five minutes and spend the rest on the things only the client knows. Split your notes in two before you go.

Then open with what you found and ask about what changed. "I saw the new site in Beresfield on your website. Is that running yet?" lands better than any question on a form. If you found something sensitive, like a prosecution, raise it plainly and early: "I saw the SafeWork matter in 2024. What changed afterwards?" Underwriters will ask, so it's better coming from you.

For the questions insurers will ask in a particular industry, pick it below. These come from our free fact finders, which ask every question on the insurers' proposal forms in plain English.

A worked example: thirty minutes on a machining business

Let's put it together on a made-up prospect: Kestrel Engineering, a precision machining business in Newcastle, New South Wales. You've got a meeting on Thursday.

Step 1: the entity (5 minutes)

ABN Lookup shows Kestrel Engineering Pty Ltd, a company with an active ABN since 2009, registered for GST, main location NSW 2304. One registered business name: Kestrel Precision. That's worth a note, because their newer website uses it.

Step 2: the signals (15 minutes)

• Website: components for mining and defence customers, ISO 9001 certification plus a line saying they "supply OEMs in Australia and North America". That's products liability with a US and Canada exposure question.
• Job ads: two CNC machinists and a second shift leader. Wages are growing and they're running longer hours.
• Street View: a metal clad factory with a sawtooth roof. There's a timber pallet yard against the back fence.
• AusTender: a Defence contract worth about $480,000 last year. Contracts like that come with insurance requirements.
• SafeWork NSW, NSW EPA and AustLII: nothing found. That's a good sign and it goes in the notes.

Step 3: the AI pass (10 minutes)

The deep research prompt confirms most of the above, adds a LinkedIn estimate of 35 to 45 staff and finds a trade magazine story about a new five axis machine. It also gets two things wrong. It lists a director who resigned in 2021, taken from an old About page. And it calls the Defence contract "ongoing" when the notice shows it ended in June. Both caught by checking the sources.

What you walk in with

• What share of sales goes to North America? Is it direct or through an Australian customer?
• What does the Defence contract (or the next one) require on liability and products limits?
• What's the new machine worth? Has the sum insured caught up?
• Is the second shift running? How many people are on the floor?
• How far is the pallet yard from the building? Is anything stored against the walls?

Five questions, none of them generic. Every one is something the underwriter will ask. That's the meeting you want to be in.

Or let Coverbound do the research

Questions brokers ask

How long should prospect research take?

About thirty minutes for a first meeting with a small to medium business: five to confirm the entity, fifteen on the public signals and ten on an AI pass and checking it. Spend longer on bigger or more complex risks.

Is it OK to use ChatGPT to research a prospect?

Researching public information about a business is fine, as long as you check every fact against its source. Don't paste client documents or personal information into a public AI tool. Use only the tools your brokerage has approved for that.

Can AI make up an ABN or company number?

Yes. AI tools can produce identifiers that look valid but belong to no one or to a different business. Always copy the ABN, ACN or company number from ABN Lookup, ASIC or Companies House yourself.

What's free on ABN Lookup and Companies House?

ABN Lookup is free and shows the entity, ABN status, GST registration, business names and location. Companies House is free and shows officers, people with significant control, charges, filing history and accounts. ASIC's basic company search is free, but extracts cost a small fee.

Should I tell the prospect I researched them?

Yes. It shows you've done the work. Open with what you found and ask them to confirm or correct it. If you found something sensitive, such as a prosecution, raise it plainly and early.

Does researching a prospect replace the proposal form?

No. Research tells you what's public. Turnover, wages, claims history and contract terms still have to come from the client. The client is responsible for their disclosure.