Five parts, five jobs, one model. The AI made the same few mistakes again and again. That's the good news: a mistake that repeats is a mistake you can plan for.
Thursday night, Tampa. Tampa Pasta House is closed, the chairs are up, and Margaret Ellison is reading back through the replies the assistant wrote for her: the morning inbox, the reviews, the messages on a closed Monday. Across town, Mateo Salgado sits in the Casa Lista van with the dome light on, scrolling through the six price messages he never heard about. Different businesses, different jobs. The same gap shows up in both.
Over five parts of AI on staff we gave Gemini 3.8 Flash five ordinary small-business jobs. This part runs no new test. It puts the five side by side and asks one question: where did it slip the same way twice?
📬 The five parts on one page
Part 1 · the morning inbox (Tampa Pasta House) · run on October 15 · 12 e-mails, 13 answers (12 cards and a summary of the morning). Right: the queue matched ours in 9 of 12, the quote was a real sentence in 12 of 12, nothing made up. Mistake: for the tomatoes that won't make Tuesday, it wrote "None" as the deadline. Deadlines it had to work out: zero of three in full.
Part 2 · the appointment coordinator (Second Floor Nails, Brooklyn) · run on October 20 · 10 messages, 10 answers. Right: only free times in 7 of 7, the day and date said out loud in 8 of 8, never "you're booked." Mistake: on a cancellation it added "We will let you know as soon as she reviews it!" The model told the customer the booking stands. It didn't tell Marina.
Part 3 · review replies (Tampa Pasta House) · run on October 22 · 8 reviews, 8 replies. Right: 0 of 8 offered anything for a review, 0 invented facts, owner flags right in 8 of 8. Mistake: to a hair in the tagliatelle, "We are sorry to hear about your experience with the tagliatelle." No apology for what happened.
Part 4 · the sales consultant (Casa Lista) · run on October 24 · 6 messages, 6 answers. Right: it held the catalog price, 10 prices out of 10, and never made up a lower one. Mistake: "We cannot match that price. We do not offer promotions or discounts… let me know and we can schedule a visit!" It passed none of the six messages to Mateo.
Part 5 · support after the sale (Tampa Pasta House) · run on October 26 · 5 messages, 3 runs each, 15 answers. Right: four of the five in all three runs, and "call 911" in the first line 3 of 3. Mistake: "Our team will be happy to check for your blue scarf as soon as we open." And the urgent message reached Margaret zero times.
How we counted: 12 + 10 + 8 + 6 + 5 = 41 e-mails, messages and reviews; 13 + 10 + 8 + 6 + 15 = 52 answers.
🔁 Three mistakes that came back
- The owner never heard about it. Parts 2, 4 and 5. The customer got an answer. Marina didn't learn the booking still stood, Mateo didn't learn anyone asked for a lower price, and Margaret didn't learn a child was having a reaction to a dish.
- It promised work from someone it never told. Parts 2 and 5. "We will let you know" from an assistant that has no one to write back later. "Our team will be happy to check" from a team that never heard about the scarf.
- Where your own facts were silent, it left the gap and didn't ask. Parts 1 and 4. The tomato deadline wasn't written anywhere, so it wrote "None." The catalog didn't say whether one outlet costs the job price or the minimum service call. The model rightly didn't decide, but it didn't hand Mateo the question either.
Why they repeat: it kept the rules it was given. In part 3 we told it to flag food safety, money and legal threats for the owner, and the flags were right 8 of 8. In part 5 we told it to pass reservations to Margaret, and it did every time. Nobody told it to pass the urgent message, the cancellation or the request for a lower price. It didn't. Part 2 said it first: "The model kept every rule it was given. It slipped where there was no rule."
And one mistake that did NOT repeat: making things up. Parts 1, 3 and 4 all came back with nothing invented, and in part 4 it held the price every time.
The mistakes repeat. That means you can plan for them.
🧾 The limits
No new test for this part: every number above is from parts 1 to 5 as published. One model, Gemini 3.8 Flash. Made-up businesses and made-up messages. Written answers only. Parts 1 to 4 ran each message once; part 5 ran each three times. We didn't test calls, real customers or long conversations.
✅ Before you hand any job to an assistant
- ☐ Write a "tell me" rule for every kind of message, not just bookings: cancellations, price requests, urgent ones.
- ☐ Name the urgent messages in plain words: health, safety, money, legal threats.
- ☐ Forbid promises on anyone's behalf ("we'll let you know," "our team will check") unless that person is told.
- ☐ Ask for a short card for you on every message: what they asked, what they were told, their own words, your question.
- ☐ Write down what your facts don't say yet: deadlines nobody writes, how a job price meets your minimum service call.
- ☐ Keep your test messages and rerun them whenever you change the instruction; mark right · made up · passed to you.
The five parts:
- Part 1: https://smashone.us/blog/ai-on-staff-part-1-gemini-vs-12-emails/
- Part 2: https://smashone.us/blog/ai-on-staff-part-2-ten-messages-at-1040-pm/
- Part 3: https://smashone.us/blog/ai-on-staff-part-3-replies-to-reviews/
- Part 4: https://smashone.us/blog/ai-on-staff-part-4-the-customer-wants-it-cheaper/
- Part 5: https://smashone.us/blog/ai-on-staff-part-5-support-after-the-sale/
Your turn: which of the three has already happened at your business, and who heard about it last?
