Start Small: A Deliberate First Fortnight with AI

The usual way an organisation meets AI is all at once. Accounts multiply across the office in a single enthusiastic week, everyone experiments in private, and a month later nobody can say what actually worked, what it cost, or whether anything sensitive was pasted somewhere it shouldn't have been. The antidote isn't caution for its own sake. It's deliberateness: one workflow, one fortnight, one honest review.

A fortnight is long enough for the novelty to wear off, which is precisely the point. What survives the third repetition of a task is what the tool is actually worth.

Choose a workflow with visible edges

The right first candidate is predictable, text-rich, and has inputs and outputs you can point to. The follow-up email after a branch meeting is a good example: the input is a page of notes, the output is a short message to a known audience, and the same shape recurs every month. So is summarising a council agenda before your delegation walks in, translating a flyer for a suburb where English is most residents' second language, or digesting a two-hundred-page inquiry report into the six points that matter for your campaign. Each of these has a clear before and after, which means you can judge the tool's work instead of guessing at it.

Resist the glamorous candidates. Strategy, messaging that will carry your name in public, anything where the cost of a subtle error is high — these make terrible first workflows precisely because you can't yet tell a plausible output from a good one.

Day one is setup, not magic

Before the first real task, spend an hour on the unglamorous parts, whatever assistant you've chosen. Create the account under an organisational email rather than a personal one. Open the privacy settings and find out what happens to your conversations — whether they're used to train the model, how long they're retained, and what you can switch off. Most assistants now offer a workspace or project feature for recurring work: make one for your pilot, and give it a short reference document explaining who your organisation is, who it speaks to, and the tone it uses, so you aren't re-explaining yourself every session.

Finally, look at what any connected tools can reach. If the assistant can see a calendar, an inbox or a shared drive, know that before the fortnight starts, and narrow the access to what the pilot needs. Nothing else on day one matters as much as knowing where your words go.

Learn to ask properly

The gap between a mediocre result and a useful one is usually the request, not the model. A good ask has four parts: clear instructions, the relevant context, the format you want back, and, where possible, an example of what good looks like. Ask for 'a 150-word supporter email for members who attended their first branch meeting, warm but not gushing, ending with one specific ask' — not 'write an email to our members'. The specific request can fail informatively; the vague one can only succeed by accident.

Keep a running note of the phrasings that worked. By the end of the fortnight that note is the beginning of a shared practice, and it's worth more than the outputs themselves.

Where the tool belongs, and where caution rules

AI earns its keep on work that is repetitive, text-rich, and reviewed by a person before the result matters. That last clause is the load-bearing one. The technology is at its best where a human reads the output anyway and the cost of a bad draft is a minute of editing.

Caution rules wherever being wrong carries real consequences, and wherever the work runs on human trust — the conversation with a longstanding member, the judgment call about how hard to push an MP's office, the doorknocking script that has to survive contact with an actual front porch. Political judgment and lived experience don't summarise well, and a fortnight's pilot is not the place to pretend otherwise.

The decision at the end

When the fortnight closes, sit down with everyone who touched the workflow and choose among three options: adopt it, adjust it and run it again, or abandon it without embarrassment. All three are wins — the abandoned pilot cost you a fortnight and taught you where the edge is. If you adopt, expand the same way you started: one workflow at a time, each with edges you can see, each with a person who owns the review. Slow is what deliberate looks like from the inside.