The forecast was beautiful. Revenue to one decimal place. Six weeks later, the biggest customer walked — and the warning signs were everywhere except the ledger.

Episode 2 of Beyond Automation takes last week's fork — automation replaces the task, augmentation multiplies the person — into the one function where mistakes come with signatures: finance. Alex traces the three-way split the best finance leaders are already making, stress-tests it against three objections (including the strongest case for the machine), and lands the controls mindset that makes confident wrongness manageable: treat AI output like junior-staff work — reviewed, sampled, signed off.

In this episode

• Why AI adoption in finance has plateaued at about six in ten functions — and why the plateau is the interesting part
• Bucket one: the close and reconciliations — automate aggressively, sample like an auditor
• Bucket two: forecasting — the machine's breadth plus the human's context (the cold open is what happens without it)
• Bucket three: capital allocation, disclosure, going concern — the estimates are "opinions wearing decimals", and accountability can't be delegated
• The question that decides sign-off: show me why the model said that
• The override log, finance edition — settle gut-versus-model with a ledger, not anecdotes
• Two beliefs, clearly labelled as beliefs — including why Alex thinks the manual close is dead inside ten years, and won't miss it

Chapters

00:00 — Edge Igniter intro 00:09 — Cold open: the forecast that was precisely wrong 01:51 — The season's three beliefs 02:58 — The three-way split: automate, augment, ring-fence 07:06 — The stress-test: three objections, steel-manned 10:46 — The playbook: the three-bucket sort 13:31 — One action this week (analyst + CFO) 14:10 — Transparency disclosure 15:10 — Sign-off and next episode

One action this week

Analysts: pick one recurring report and automate the assembly — keep the commentary. Your name stays on the thinking.

CFOs and finance leaders: sort this quarter's finance calendar into the three buckets, then automate one, augment one, ring-fence one — and write down who signs each. If you can't name the signer, it isn't ring-fenced; it's abandoned.

Companion

The New Finance Tech Stack — the three buckets mapped to actual tool categories, vendor-neutral, with the judgement layer drawn on top. Free in the Edge Brief: https://eitheedgebrief.beehiiv.com/

Transparency

AI helped research, verify and draft this episode; a human wrote the final script, made the calls, and owns every claim. The cold open is a composite drawn from patterns finance people will recognise, not one real company. A note on the adoption figure: several conflicting numbers circulate; we quoted the one from a named survey with a stated sample and dates — and what it shows is a plateau, which we think is the more useful truth.

Important: this episode is general information — not financial, accounting or audit advice. Rules differ by jurisdiction and change; before altering controls, reports or processes that someone signs, consult your advisers.

Two workers. One company. One AI rollout. Two completely different futures.

Season two of Edge Igniter opens where season one ended: the machines learned fast — now what? Alex traces the fork between automation and augmentation through three independent datasets that have nothing to do with each other, and keep agreeing anyway: payroll records covering millions of workers, millions of anonymised AI conversations, and a workplace experiment with more than five thousand customer support agents.

The short version: automation replaces the task; augmentation multiplies the person — and which one your organisation gets is a design choice someone is making right now, possibly without knowing it.

In this episode

• Why the entry-level door is quietly narrowing in AI-exposed occupations — and why experienced workers in the same jobs are thriving
• The usage data showing both futures happening at once, at scale
• What happened when one company deliberately chose augmentation: newer staff 30% more productive, two months of experience performing like six
• The stress-test: three objections to the "design choice" claim, taken seriously
• The playbook: three questions that tell you which side of the line any task sits on
• Where Alex stands — clearly labelled as belief, not finding

Chapters

00:00 — Edge Igniter intro 00:09 — Cold open: two workers, one rollout 01:53 — The three beliefs this season runs on 03:35 — The evidence: three datasets that agree 08:38 — The stress-test: three objections, steel-manned 11:56 — The playbook: three questions and a redesign map 15:24 — One action this week (individual + leader) 16:01 — Transparency disclosure — including a correction to a season-one statistic 17:13 — Sign-off and next episode

One action this week

Individual contributors: write your role as ten tasks and mark each one — automate, augment, or human. That's your redesign map. Take it to your manager before someone else takes theirs.

Leaders: run the same exercise for one team before the next tool purchase. If you can't say which tasks a tool automates and which it augments, you're not buying a strategy — you're buying a subscription.

Every day you face the same decision dozens of times: the machine hands you an answer — do you trust it?

The research says we get this wrong in both directions. Decades of automation studies show that the more reliable a system becomes, the worse we supervise it — operators of highly reliable systems were about 50% less likely to catch failures. Meanwhile, Wharton's "algorithm aversion" experiments show we abandon algorithms after a single visible mistake, and in Stanford's diagnostic trials, doctors overruled correct AI answers so often that the AI alone beat doctors using the same AI.

Alex Rivers closes the opening Edge Igniter series with the judgement skill in daily practice: a three-question framework — What does it cost to be wrong? Is this situation normal? Can I check it? — that sorts every AI answer into accept, adjust, or reject. Plus the aviation lesson of "children of the magenta line", why your GPS sounds equally confident driving you into a bay, and three habits that calibrate your trust over time, starting with an override log almost nobody keeps.

Chapters (approximate — verify against final audio)

00:00 — Intro 00:10 — The GPS that drove into the bay 01:30 — Failure one: automation bias (why reliable machines put us to sleep) 03:00 — Failure two: algorithm aversion (one strike and we're out) 04:15 — Why machines fail differently: confidence isn't competence 05:10 — The framework: three questions, ten seconds 07:00 — Accept, adjust, or reject: five worked examples 09:10 — Calibration habits: the override log, weird zones, flying manual 10:30 — Closing the series: the machine gets a vote, never the final vote 10:55 — This week's action 11:10 — How this episode was made (AI disclosure) and sign-off

Key takeaways

The most dangerous AI in your workplace is the excellent one nobody checks anymore — reliability breeds complacency, and training alone doesn't fix it. The skill isn't trusting AI more or less; it's calibration: matching scrutiny to stakes, normality, and verifiability, case by case. High stakes or abnormal situations shift the vote to the human; cheap-to-verify answers get verified before trusted. In safety-critical work, the machine can talk you into stopping, but should never talk you out of your own alarm. And calibration is a learnable skill: keep an override log, probe your tools' weak zones, and regularly work without the machine so the backup system — you — actually functions.