AI IN PRACTICE

What AI got wrong on real client work.

Two analyses I ran on live engagements, and the errors that looked like conclusions.

AI will give you a marketing plan in minutes. It will also be confidently wrong in ways that are invisible unless you know the category. These are two real examples, with what the output got wrong and what I did about it.

See how I work with clients

If a prompt could run your marketing, you would not need a CMO, fractional or otherwise.

Neil Henry, Fractional CMO

Start here

Can AI replace a marketing hire?

The tools are free. The judgement is not. Here's what marketing roles actually cost, and which ones AI comes closest to replacing.

Read the argument

Positioning Researcher

When to run it:: Onboarding a new client, or when the messaging stops working.
What to feed it:: The client's website, three to five competitor sites, any sales call notes, and a link to their review profiles.

What it got wrong

Corrections found after running this agent on live client work.

1
It assumed one buyer.
In most B2B markets there are two. Somebody specifies and somebody purchases, and they search completely differently. The agent now defines both separately and has to state which one every recommendation is for.
2
It assumed people search by problem.
In this market nobody searched for a supplier. They searched for a product code. The agent now checks whether a category is code-led before it does anything else, and treats each code as its own intent.
3
It only looked at direct competitors.
The most dangerous competitor turned out not to be a competitor at all. It was somebody selling the same thing on eBay with a basket instead of a quote form. The agent now has to include substitute purchase routes, not just the obvious rivals.
4
It let the client define their own strength.
That is how you end up defending a claim nobody believes. The agent now tests every candidate position against what customers actually wrote in reviews. If customers are not saying it, it does not count.

How I read the output

The first three things I check before trusting what comes back.

1
Check it has named more than one buyer.
If the output treats the person who signs and the person who uses the product as one audience, the positioning will fit neither of them.
2
Check the strength is evidenced by customers, not the client.
A claimed strength only counts if buyers mention it without being prompted. If the evidence comes from the business's own description of itself, treat it as a hypothesis.
3
Check the competitor set includes the alternatives.
The real competitor is often a spreadsheet, a cheaper supplier or doing nothing. If only direct rivals appear, the gaps it finds are probably not gaps.

Buying Committee Mapper

When to run it:: Before any campaign aimed at a purchase more than one person has to agree to.
What to feed it:: The client's website, two or three competitor sites, target customer, reviews and case studies; optionally, the Positioning Researcher output.

What it found

From a recent run on a specialist lender. Company details removed.

1
The people who paid never signed.
The loan was repaid from the beneficiaries' share of the estate, and they didn't have to be asked. One person signed, and a professional adviser chose the lender. The marketing spoke only to the person signing.
2
The biggest competitor was free.
A free government payment scheme is the standard route, and advisers look at it before any lender. The website never mentioned it.
3
The website was talking to people who couldn't buy.
The main product page addressed individuals, but the small print required a professional to be involved. The person who actually decided was barely addressed.

The recommendation was one change, and it would cut enquiries while improving them. Knowing whether that trade-off is right is the part a prompt cannot do.

What it got wrong

Corrections found after running this agent on live client work.

1
It mapped the seller instead of the buyer.
Every input described the business doing the selling, but the committee sits with the customer. It now needs a defined target customer, and treats the customer's own job adverts as the best evidence of who owns the budget.
2
It tried to map three products at once.
One company sold three loans, and each had a different committee. The result was a blur. The agent now maps one offer per run.
3
It implied it had checked things it couldn't reach.
Some sources were blocked, but the output read as if they had been reviewed. The agent now lists every source it could not access, and never implies a check it didn't make.
4
It ranked objections by feel.
Two objections could be equally fatal with no rule to separate them. The agent now ranks by how fatal each one is, and where two are equal, the one that arrives later ranks higher.

How I read the output

The first three things I check before trusting what comes back.

1
Check it looked for the payer and the chooser separately.
If the output says one person pays, chooses and signs, it has usually stopped looking.
2
Check the late objection isn't about price.
Deals that die at the end usually die over blame. If price tops the list, the map is too shallow.
3
Check the recommendation costs something.
One decision with a named risk is useful. A list of content ideas means it avoided making the call.