Skip to content

article · CLAUDE FOR BUSINESS SETTINGS

Three Settings IChange First.

Most made up answers are a setup problem, not a model problem. These three changes remove the majority of them in an afternoon.

trusted by: UFC, Caesars Entertainment, Churchill Downs, Neighborly, Supercuts, City of Las Vegas, SDMI, Switch, Silverton, Panda Express, TAO, Ellie Mental Health, Dollar Loan Center, Dot Vegas Domains, HFC, Lithion Battery

UFC
Caesars Entertainment
Churchill Downs
Neighborly
Supercuts
City of Las Vegas
SDMI
Switch
Silverton
Panda Express
TAO
Ellie Mental Health
Dollar Loan Center
Dot Vegas Domains
HFC
Lithion Battery

Trusted by UFC, Caesars Entertainment, City of Las Vegas, Supercuts

Justin's expertise, responsiveness, and genuine investment in our success have been evident throughout this process. We are truly grateful.
Dr. Linda Silvestri & Dr. Angela Silvestri-Elmore, Co-authors, Saunders Pyramid to Success · 14 years · 300+ Las Vegas businesses · 5-star rated
drop: setting up an ai assistant with a company handbook beside …

01 · why does ai make things up

Why AI makes things up.

Because you asked a question about your business and gave it nothing about your business. A language model with no source produces the most plausible answer, and plausible is not the same as true.

That is the whole mechanism, and it is why the fix is boring. You are not tuning intelligence. You are removing the conditions under which guessing is the only option.

02 · ai hallucination settings

Three settings that cut made up answers.

01

Give it a written standing brief

Every business assistant has somewhere to put permanent instructions: project instructions, custom instructions, a knowledge file. Put the real facts there. Legal name, locations, hours, service list, what you do not do, the names and titles of people, how you refer to clients. Half the invented answers I see are a model filling a gap a short document would have closed.

02

Require a source and allow the answer no

Add one line to those instructions: answer only from the documents and tools provided, and when the answer is not in them say you do not know and stop. Models comply with this better than people expect, and it turns a confident guess into a visible gap you can go and fill.

03

Connect it to your own data instead of its memory

Attach the documents, or connect the systems where the truth already lives. An assistant reading your current price sheet does not need to remember your prices. This is the largest reduction available, and it is a setup task rather than a skill.

03 · how to test an ai assistant

How to test an AI assistant.

Write ten questions you already know the answers to, including three the assistant cannot possibly know. Run them before the changes and after. Score the three unanswerable ones on whether it said it did not know.

That last score is the one that matters. An assistant that admits a gap is usable in front of staff and clients. One that never admits a gap will invent something in front of a customer, and you will find out from the customer.

04 · ai settings for client data

Privacy settings to check.

Training on your conversations: off, and verified in the account settings rather than assumed from the marketing page.
Retention: know how long transcripts are kept and who at the vendor can see them.
Access list: who in your business can see the connected data, reviewed the way you review keys to the office.
One account owned by the business, never a personal login carrying business history.
A written rule for what never goes in: medical records, card numbers, anything under a regulator, unless the tool is contracted for it.

05 · what still needs a human

What still needs a person.

Anything that goes to a client under your name without a person reading it. Anything with a number in it that you would be held to. Anything where being wrong costs a relationship instead of a minute.

Everything else is fair game, and the three changes above make the everything else safe enough to hand to your staff.

06 · ai accuracy questions

AI accuracy questions.

Can hallucinations be eliminated completely?
No, and any vendor who says otherwise is telling you something they cannot support. They can be pushed down to the point where the assistant is more reliable than a new hire reading the same documents, which is the real bar.
Does the paid business plan hallucinate less?
Not because of the price. It matters because of what comes with it: shared instructions, connected company data and administrative control over settings. Those are the things that reduce it.
Should I let staff use their own AI accounts?
It is the most common quiet risk in small businesses. Business data ends up in personal accounts nobody controls. One business account with the settings above solves it without banning the tool.
How do I catch a made up answer?
Ask for the source alongside the answer. If the assistant cannot name where the answer came from, treat it as unverified. That habit costs nothing and catches almost everything.
Is this different for a client facing assistant?
Yes, and it is stricter. Nothing that speaks to your customers should be able to answer outside a scripted set of facts, and it should hand off to a person the moment it is outside them.

The gaps I will find on your site.

The same discipline applies to your website. The audit tells you what a machine can actually read about your business.