Mistakes to avoid

We have seen the same mistakes in firms of 8 and of 80. They are not “technology maturity” failures. They are scope, vanity, and not protecting two hours. This list is more useful than another adoption framework.

Read it before you sign. If your proposal commits three, send it back. Including ours.

The ten

• Start with the website chatbot. The visitor is the hardest judge. Start inside.
• No named owner. “The department” does not operate a Friday.
• No pilot sentence. If it does not fit in a sentence, it will not fit in a month.
• No baseline week. Then every number is theatre.
• Seats for everyone “to experiment”. Noise, leaks, zero habit.
• Pasting the customer spreadsheet into a free account. A delayed incident.
• Wiring the whole ERP in month one. A maze. File or queue.
• Training the whole company on a Friday and not reviewing at day 14. An event, not training.
• Measuring tokens or smiles. Measure hours, errors, accept rate.
• Painting day 30 because “it is already paid”. Cut or stop. Do not bury it.

Why they happen (and how to stop them)

They happen because the salesperson shows the pretty case, because leadership wants a photo, because the voucher is tight, because nobody wants to be the person who says “not yet”. The antidote is a visible list (the start checklist) and someone allowed to say no.

The cheapest no is in week zero. The dearest is in month six, with the ERP half-wired and the team burned.

Vendor mistakes (they exist too)

Promising adhesion. Promising an accuracy percentage without your sample. Hiding the integration price. Vanishing until justification. Pushing their only partner. Calling a webinar “training”.

Ask for a week-2 deliverable, a sample on your documents, and who picks up the phone on Friday. If the answer is fuzzy, they are not your vendor.

If you have already made them

Stop the scope, not necessarily the relationship. Write the sentence. Name an owner. Take a baseline week now (better late). Close the stray accounts. Open one queue. Review prompts. If after 15 days of this cut there is no use, switch off and take the old Monday back. That is fine. Pretending is not.

A well-closed dead pilot is experience. A zombie pilot is a fixed cost and a toxic story (“AI does not work here”).

Risks and the programme

Each mistake on this list is one of the five risks (hallucination, leak, lock-in, bias, empty file) landing by design. The programme does not ask you to make them in order to “go fast”. It asks for coherence. Coherence is slower in week one and cheaper the rest of the year.

How to send a proposal back (sample paragraph)

“Thanks. Before we go on we need: one-flow sentence, a named owner, a baseline week, a written data perimeter, a week-2 deliverable, and the price split into design / implementation / integration / training / operations. Without that we cannot compare. A public chatbot, if it enters, enters after a measured internal month.”

Copy it. Paste it. It is not rude. It is the filter that saves you month six. If the vendor takes offence, you have avoided one of the ten.

Apply the same paragraph to our quote. If it does not meet it, send ours back too.

The silent eleventh mistake: writing nothing down

No sentence, no poster, no sheet, no vendor card. Everything “was discussed”. Two months later nobody remembers the perimeter and everyone pastes what they like. Writing one side of a page is not bureaucracy. It is the only thing that survives the holiday.

If you can only do one thing from this list, write the pilot sentence and the yes/no/ask poster. The rest can be cut. Those two cannot.

If you are already in the middle of three of these mistakes, you do not need a crisis. Fifteen days of cutting is enough: one sentence, one owner, one queue, personal accounts closed. The grand announcement can stay. The scope cannot. That cut is the difference between a zombie and experience.

Print the ten and let someone outside (us or another) tick them against your proposal. Self-check forgives. An outside check does not. Fifteen minutes of that reading prevent the month when you no longer want to hear the word pilot.

Frequently asked questions

Is one mistake worse than the others?

Too much data in an account with no contract, and no kill switch. The rest can be cut. Those two follow you.

Is a chatbot always a mistake?

A chatbot on day one, yes. An assistant that cites a FAQ after an internal month, no.

What if leadership already announced “AI everywhere”?

Translate the announcement into one flow and one number. The announcement stays; the scope is cut. Someone has to do it.

The positive list is the start checklist and where to start. We will review a proposal (including someone else’s) against these ten.