Airclerk is new, and the most useful thing we've learned so far came from the deals that didn't close.
We were not wrong about what frontier AI can do for regulated financial services firms. We were wrong about the size of the first step. We pitched the destination. Most firms needed a doorway.
We believed in the destination strongly enough that we built it into our own product: an end-to-end agentic workflow for regulated knowledge work. The conviction was right. The sequencing was wrong. People aren't ready to adopt what's possible, not because they don't believe it works, but because the organisational distance between where they sit today and where the demo lives is too far to cross in one step.
So we changed what we sell. This post is the longer version of why, and why we think the gap is going to keep getting wider, not narrower.
The gap is real, and it's widening
On one side: what's actually possible with Claude in May 2026. Workflows that read policy documents, compare them against rules or internal guidelines, draft customer-facing responses, file the supporting evidence, and flag edge cases for a human. Not as a demo. In production, connected to real systems, with logs and audit trails. The same shape applies to an accountant reconciling a year-end pack, a wealth adviser preparing a statement of advice, or (well outside our wheelhouse) a legal team triaging a discovery dump.
On the other side: where most regulated firms in NZ and Australia actually sit. A handful of people experimenting with the ChatGPT app on personal accounts. A nervous IT-led "AI policy" that mostly says don't paste customer data into anything. Maybe one Copilot pilot nobody's quite measured. A general sense that "we should be doing more on this" coexisting with no clear sense of what or who owns it.
The gap between those two states isn't a step, it's a chasm. And the capability side is moving faster than the readiness side, so the chasm is getting wider, not narrower. Every six months the frontier moves another half-step out. Adoption-readiness inside a regulated firm moves on a much slower clock: governance, risk, vendor onboarding, change capacity, internal politics. That clock doesn't speed up because the model got better.
That gap is the reason our own end-to-end product, despite the conviction we built it with, sits on the wrong side of the chasm for most of the firms we talk to. It assumes a level of readiness most aren't yet at. The work that's actually needed right now isn't selling the destination. It's helping people take a much smaller first step, small enough that conviction gets earned rather than assumed.
Worth saying out loud: we're putting most of our attention on Claude because it's where we see the strongest fit for regulated knowledge work today, not out of vendor loyalty. The argument that follows would still hold if the leading model were called something else.
Why the obvious response is the wrong one
The obvious commercial response to a gap that wide is to bridge it in one move: build the case, sell the vision, sign the multi-year transformation. There's a whole tier of consulting firm that lives on exactly this play. It fails for three structural reasons.
1. The buy-in is based on the happy path. Aaron Levie at Box recently made the point that CEOs are particularly exposed to this: they sit far enough from the last-mile work that when they play with AI themselves, they see the demo result and underweight the next ten or twenty things that have to happen to make it sustainable. "Look, I generated a contract" misses verifying the terms, wiring in the back-catalogue of past contracts, and building a review process for the bits the model gets wrong. The transformation pitch gets a sympathetic hearing at the top, on the strength of that incomplete picture. Then the engagement meets the real systems, the follow-on work becomes visible, and the conviction evaporates somewhere between vendor risk and the change advisory board.
2. The org can't absorb the change even if you ship it. The mental model most people walk in with is that AI is going to replace a person: a whole role, a whole headcount line. That framing is wrong, and it's actively damaging to adoption. What AI actually replaces, in practice, is roughly one fiftieth of what one person does, on day one. One task. Then another. The compounding is in the count of small workflows absorbed over time, and the time it frees up for that person to do the other things they should already have been doing. A pitch that asks an organisation to swallow the whole bird in one go triggers the immune system every time. A pitch that asks them to swallow one bite doesn't.
3. The goalposts keep moving. What we believe works in May 2026 might not be the right approach by November. Locking a firm into an 18-month roadmap is locking them into a snapshot of what was possible the week the contract was drafted. The honest response to a frontier moving this fast isn't a bigger plan, it's a shorter one. Commit to small, ongoing deployments. Ship something real every fortnight. Reassess every quarter.
The smaller pitch, and why it compounds
What works is deliberately proposing something smaller than the client thinks the moment calls for. One workflow. One team. One use case where the ROI is plainly visible: claims triage, broker submission summarisation, statement-of-advice drafting, the thing the operations lead has been complaining about for three years. Six to eight weeks. A budget that doesn't need board approval. A scope that can be killed without anyone losing face.
The pitch sounds almost embarrassingly modest. "We're going to make this one thing work, in production, with the people who do it today, and we'll show you it working in eight weeks. If it doesn't, you've spent less than a quarter of a senior hire."
What that pitch actually buys is the only thing that matters: the first instance of "we tried Claude on a real piece of our work and it did the thing." Once that exists, the conversation in the building shifts from "should we be using AI?" to "where else?". Different cast of people, different objections, different cadence. It's the conversation you want to be in.
The risk function gets a concrete artefact to assess instead of a hypothetical. Risk committees are extraordinarily good at saying no to abstractions, and surprisingly reasonable when looking at a real thing already running with logs they can read.
And the people inside the firm get to be the heroes of their own AI story. They commissioned the small thing. It worked. Their internal credibility goes up. Their appetite for the next small thing goes up. That is how change actually moves inside a regulated firm.
What this means commercially
We're aware this isn't a flattering pitch from a "scale the consultancy" standpoint. The temptation, and most of our peers give in to it, is to bundle the small thing inside a bigger thing, sign the bigger contract, and deliver the small thing first as "phase one." We've decided not to do that, for a reason that has more to do with compound interest than virtue. Each small thing we ship cleanly is the strongest possible sales artefact for the next one. A botched transformation is the strongest possible sales artefact for our competitors.
This is why the Readiness Sprint is structured the way it is: short, fixed scope, ends in a governed, working example in the firm's own environment, not another strategy deck.
The gap between what's possible and what your firm is ready to do is genuinely enormous, and it's getting bigger by the quarter. You don't close a gap that big in one jump. You close it by taking the first step often enough that, eventually, you look up and you're somewhere new.