AI: What's Actually Possible

Nobody is an AI expert yet

AI isn't a search tool any more. It's the newest member of your team. That was the line I opened with, and the rest of the session was about what that new team member can actually do today, and how you onboard it.

Mainstream generative AI is about four years old. Most people in that room had 10 to 30 years in analysis and change. The skills that matter most with AI are the ones BAs already have: eliciting what people actually need, reading a room of stakeholders, knowing which requirement is the real one. And most BAs already see the upside: 69% say AI is positive for their career and only 5% see it as negative (IIBA Global State of Business Analysis 2026, 2,120 professionals in 122 countries). 84% of NZ knowledge workers already use generative AI (Accenture and Microsoft, 2024).

Meanwhile the capability is moving fast. The length of task a model can finish on its own has gone from about four minutes of human work (GPT-4, 2023) to 17.4 hours (Claude Mythos Preview), per METR, which flags readings above 16 hours as less reliable.

What it actually is

It predicts the next word. It read the internet, but it has never read your Confluence pages, process maps, backlog, stakeholders' emails, or what the last workshop agreed. Which is why it can be confidently wrong. My example: asked whether NZ requires reporting a privacy breach, it searched, found a 2019 page, and confidently said notification was voluntary under the Privacy Act 1993. The Privacy Act 2020 made it mandatory from 1 December 2020.

So the whole skill is the brief. "Write up the workshop" gets generic output. "Requirements in MoSCoW, from this transcript. Flag conflicts. Don't invent." narrows the beam. Anyone plus an AI tool gets generic output; your years of experience, ways of working and domain knowledge, coaching the same tool, get great, specific output. That's where a BA's IP sits.

Before any real customer data goes near it: on a personal or free account the individual sets the rules, training is a per-person setting and there's no audit trail. On a plan your organisation runs, the organisation sets the rules, it isn't used to train the model, and there's one record with admin controls. Customer and citizen data only goes into a tool your organisation controls. The Privacy Act 2020 applies to AI tools, and public sector teams also have the GCDO Public Service AI Framework (2025).

A Skill is your team's playbook

A Skill is a written-down way of working the AI follows every time. Written once, used by everyone: every BA, the testers, next year's grad. I run 17 Skills of my own, each written once, including an email drafter that writes in my voice.

The example on screen (the photo at the top of this page) was an "Our Story Standard" Skill:

• Our story format and definition of ready.
• Acceptance criteria in Given / When / Then.
• Every story gets a negative path.
• Split anything that fails the INVEST check.
• Never invent a requirement nobody stated.
• A BA reviews it before it hits the backlog.

Connect it, and a chatbot becomes an agent

Connect it to what you already run: email, files, meeting transcripts and calendar, the wiki, the backlog, and anything else through the browser. A chatbot answers and you go do the work. An agent finds what it needs, does the work, and you check and approve. It reads 40-page specs and transcripts, sees whiteboard photos and scanned forms, listens to interviews and voice notes, and clicks through legacy systems. The management model I gave the room: AI is the intern, you are the senior. It never gets bored of the twelfth user story, but it has no idea how your organisation does things and can be confidently wrong.

Four demos, one fictional insurer

All four demos were recorded and talked over live, and all ran on the same scenario: Northgate Mutual, a fictional NZ insurer modernising claims intake.

• Every stakeholder's view at once. Five AI agents each took a persona (the head of claims as sponsor, the contact-centre lead, the IT integration lead, privacy and compliance, and a policyholder lodging a claim on their phone at 9pm) and reviewed the same online claim form design. In about two and a half minutes: every view, ranked, conflicts mapped, and the questions to be ready for. Preparation, not a replacement for talking to people.
• Speed to prototype. Signed-off Given/When/Then acceptance criteria went in. Claude Code wrote the tests first, and all eight failed because nothing was built. Then it built until all eight passed and deployed a live link the room could open. About eight and a half minutes end to end.
• Buried in the threads. Ten email threads, three meeting transcripts and 21 backlog stories. Claude read all of it through the Drive and Notion connectors and wrote a brief: a decision agreed in a corridor that a story still contradicts, a story nobody asked for, and a privacy lead never invited while the form collects medical details. On a schedule, that brief is waiting at 6am.
• The system nobody documented. "ClaimsMaster 4", a fictional claims system built in 2009 whose rules left with the people who knew them. Claude in Chrome clicked through every screen, tried lapsed policies, late claims and big estimates, and wrote the spec nobody wrote: a field dictionary, a process map and the hidden rules, like claims over $25,000 routing to a senior assessor and HEIC phone photos (most iPhones) being rejected.

The line that does not move

AI prepares: specs, briefs, tests, a first build. You approve: read it, change it, own it. Then it goes out. Priorities, trade-offs, sign-off, and anything a stakeholder or customer sees all stop at the approval step.

What shifts is the work around it. Less box-ticking documentation, reverse-engineering old systems, chasing what was decided where, and waiting for a build to test an idea. More facilitating the room, judgement on trade-offs, time with stakeholders, and testing ideas with real users.

Start with the most painful thing

Like any new team member, it needs a clear brief, your know-how written into a Skill, and your sign-off every time. Where I left the room:

• Name it: what was the most painful task last week?
• Get it out: write it down or say it out loud.
• Hand it over: tell an AI tool and ask how it could build a Skill for it.
• Test it on real work, not a toy example.
• Rinse and repeat next week.