Staff Your AI Initiative: Three Roles That Ship

In May 2025 the founder of a freight brokerage in Gdansk showed me three model invoices. His dispatcher bot had run on three flagship models in four months. Each switch cost his contractor two weeks of re-tuning. Each switch barely moved answer quality. His dispatchers still copied every load number by hand, because the bot died on real paperwork. I asked him who owned the bot between switches. Nobody. One freelancer tuned prompts in spare hours. The models changed. The team stayed one person wide. Nothing shipped.

I told him to stop shopping for models and hire one AI engineer for six weeks. She kept the cheapest model, connected it to his document store, and seated a dispatcher at the review screen. Six weeks later the bot cleared seventy percent of routine loads alone. The model mattered least. The three seats around it mattered most. Here are those three seats.

Models change weekly, crews carry the year

I made the culture case in an earlier post: crews that log failures and review them together outlast bigger tool budgets. I will not repeat it here. This post names the crew. One model swap changes answer style for a week. One hire changes what the team ships for a year.

Seat 1: the AI engineer

The AI engineer owns production. She connects the model to company data through retrieval (RAG) over the document store, and she signs every release that touches customers.

Her week runs on three dials: cost, accuracy, response time. On the Gdansk bot she routed load-number reads to a cheap model at two cents per hundred calls, and saved the large model for damage-claim letters where one wrong clause costs real money. She set a four-second budget per answer and trimmed the document set when answers ran slow.

She also owns the dull parts that decide whether a demo survives contact with customers: an eval set of two hundred fixed questions, rerun on every release; a fallback model when the first one times out; logs that tie each answer to its source pages. When dispatchers flagged a wrong delivery window, she traced it to one stale PDF in the store and replaced it the same day.

Hire this seat first when nothing runs in production yet. No prompt survives a missing retrieval layer.

Seat 2: the prompt engineer

The prompt engineer owns instruction reliability. He mixes three crafts: enough code to read logs and version prompts, enough analysis to count failure modes, enough language sense to hear why one sentence misleads a tired driver at midnight.

His material is the conversation itself. He writes the system instructions, few-shot examples, and output schemas the code parses. He keeps a regression file of past failures, rerun before every release. On the Gdansk bot, drivers ignored the bot confirmation and phoned dispatch anyway. He read forty transcripts; one line held the cause: the bot confirmed the pickup town and skipped the street. He rewrote the confirmation to name street, dock number, and time window, and added one rule: the bot asks a follow-up when any of the three is missing. Callback calls fell by half in two weeks.

He also shapes the human side of the exchange: when the bot states facts, when it asks, when it hands the chat to a person. Drivers trust a bot that admits its limit and says 'I cannot read this document, Anna from dispatch takes it from here' over one that guesses. His work ends where guessing starts.

Seat 3: the responsible-AI lead

The responsible-AI lead owns the guardrails. She decides what the bot may do alone and what waits for a person.

Her tools are dull and she likes them dull: an approval gate on every outbound message to customers, a redaction pass that strips phone numbers and ID numbers before anything reaches logs, a thirty-day retention stamp with nightly deletion, a monthly bias watch that compares refusal and error rates across customer groups. She maps every stored field to its GDPR legal basis, so the team answers regulator letters from one page. Twice a year she hands an outside auditor the log tables before he asks.

On the Gdansk build she stopped one release the day before launch. The bot quoted tariffs from a driver contract that had expired in March. Her pre-launch review caught the stale file, because her review always opens the source list behind quoted numbers. The delay lasted one day. A wrong quote to forty carriers would have fueled disputes for months.

Bring this seat before the bot touches personal data or regulated money. Past that point her absence costs more than her salary.

Three seats, one crew

Three separate hires form a queue. Four habits turn them into a full-stack team.

First, Friday demos. Each of the three shows one weekly failure and one fix in fifteen minutes, transcript on screen. The Gdansk crew ran eleven of these in a row. The prompt engineer learned which retrieval misses to expect. The AI engineer learned which instructions confused drivers. Knowledge crosses seat borders every week.

Second, a shared failure log. Every wrong answer lands in one table with transcript, source pages, and an owner name. No private notebooks. Anyone reads it. Anyone adds to it.

Third, shared release rights. Any of the three stops a release, and the stop needs no defense beyond one failing example. The responsible-AI lead used this right once. The team thanked her, because the rule had stood since day one. That is shared ownership in practice: the release belongs to all three.

Fourth, a human in the loop. Edge cases route to a dispatcher, and each dispatcher decision flows back into the eval set. The crew studies new model versions together each month and reruns the regression file before adopting anything. Learning follows the calendar. Knowledge lives in files all three touch, so no departure takes it away.

Where a specialist plugs in

Hire order follows pain. When no bot runs in production, hire the AI engineer first. She builds the retrieval layer and the eval set that later hires stand on. When a bot runs and misunderstands people daily, hire the prompt engineer first. He repairs the conversations the team already has. When the bot handles personal data, money, or hiring decisions, bring the responsible-AI lead before the next release. Her review costs one week. A regulator letter costs a year.

Outsource the rest. An outside reviewer sets up the first eval set in days. An outside red team probes the bot once a quarter. An outside lawyer maps the data fields once. Keep three things inside: the prompt files, the release decision, and the failure log. Those three carry the product memory, and product memory stays home.

Start with one seat this month. Add the second when customers talk to the bot daily. Add the third before the data turns sensitive. Three seats, one crew, and the model becomes the cheapest decision you make.