2nth · Plain-language briefing for finance leadership
No jargon. One idea, four numbers, one spending rule — and the question to ask before signing anything.
AI assistants answer from your company's filing. Tools like Microsoft Copilot read what your business already has — documents, spreadsheets, email — and answer questions from it. That is the whole product.
Which means: if your filing is five versions of the same report, a tracker spreadsheet someone keeps "their copy" of, and folders shared to everyone since 2023, you are about to pay a monthly fee per person for an assistant that confidently summarises the wrong version. The assistant isn't broken. The filing is.
Nobody sells you this sentence, because the fix isn't a licence — it's work.
The pattern repeats everywhere: seats bought for everyone, no preparation, and within 90 days the organisation's verdict is "we tried AI and it was mediocre." That verdict is the real cost — not the wasted subscription, but the two years of internal cynicism that follow it, while competitors' teams quietly get faster.
The counterintuitive truth: the IT tidy-up you've deferred for years and the AI programme are the same project. Sorted filing, sensible permissions, one authoritative version of each number — that isn't preparation for AI. On current technology, it is most of what makes AI good.
In organisations under ~200 people that have never audited licences, 15–30% of the spend is typically recoverable: seats for people who left, duplicate subscriptions from old entities, add-ons nobody opens. The audit usually part-funds the AI programme. Ask for this number before approving anything new.
An AI seat costs roughly a coffee-shop lunch per person per week. It earns that only if the person uses it on real work — a training and habit question, not a licensing one. Demand a usage report at 90 days, not a feelings survey. Unused seats get cancelled.
Assembling packs by copy-paste is where document-heavy businesses bleed time — and the first place a well-grounded assistant pays for itself. Baseline it now so the after can be measured against a before.
These assistants respect your existing permissions — exactly as they are. If a confidential folder was over-shared years ago, the assistant will cheerfully surface it to whoever asks. It doesn't create the exposure; it reveals it. A permissions review before rollout is incident insurance, and it's cheap.
AI programmes fail as big-bang purchases and succeed as gated investments. Each stage produces evidence before the next rand is approved:
| Gate | You're buying | Evidence before the next gate |
|---|---|---|
| 1 · Audit | 2–4 weeks of discovery | The waste number, the risk list, a costed plan |
| 2 · Foundations | Tidy filing, fixed permissions, staff trained | Permissions test passed; literacy sessions done |
| 3 · Pilot seats | A small licensed group, defined use cases | Usage data + hours saved, in writing |
| 4 · Rollout | Seats where the pilot proved value | Adoption holding; dormant seats culled |
| 5 · Automation | AI doing multi-step work, human sign-off | Every flow with an owner and audit trail |
The board never approves "an AI strategy." It approves the next gate, every time, on evidence. This is also your protection against vendor enthusiasm — including ours.
Literacy for everyone — a few hours, not a course: what these tools are genuinely good at (drafts, summaries, finding things), where they fail (they state errors with total confidence), and the one rule that prevents most incidents: if you didn't sign in with your work account, don't put work information into it. Your staff are already using free AI tools with company data — the choice is between governed use and invisible use, not between use and no use.
Accountability stays human — the person who sends it, owns it. AI drafts; people decide. No AI-made decisions about people or money. That sentence is most of your AI policy, and your auditors will ask for it in writing.
Measured, not vibed: licence waste recovered; reporting-cycle hours down against the baseline; every seat active or cancelled; zero data incidents; one or two carefully-governed automations doing real work with a named human approving the output. Not "transformation" — a tidier, faster, better-protected business where the AI spend is justified line by line, and a foundation on which the next wave of AI (coming regardless of what you approve this quarter) lands well instead of badly.