Inspiration
Firms are paying for AI licences, but most staff do not know what to hand to AI. Managers cannot see where hours leak. And since February 2025, Article 4 of the EU AI Act says every company using AI must take measures to make its staff AI literate. Most firms have no record that the training happened.
We wanted something a busy person could act on in one minute, built from data the firm already has.
What it does
Handoff has three views, all built around one real deadline: the ROS pay-and-file date on 18 November.
- Employee: Aoife, a junior tax assistant, asks in plain words: "What should I hand to AI before 18 November?" Handoff checks the request is clear, asks one question if needed, then gives her one skill, not a menu. She sees about 15.7 hours and €1,337 back, an 11-day countdown, and a check step, because she verifies only 18% of AI output today.
- Partner: engagements ranked by money lost. Walsh Holdings is €8,677 over budget, and its biggest block of work has a 0.95 AI usage score in the Economic Index.
- AI Lead: adoption by team, verify rate, CSO and Eurostat benchmarks, a short course on Anthropic's 4D AI Fluency framework, and a one-click export of the Article 4 literacy log.
When Aoife clicks "I'll try it", that click lands in the AI Lead's log. The training record comes from doing the work.
How we built it
The engine joins four data sources:
- Anthropic Economic Index: how much people already use AI for each task.
- O*NET: the official task list for accountants.
- Revenue deadlines and Irish public holidays: real filing dates.
- Company data: timesheets, AI usage and engagements.
Each timesheet line maps to an O*NET task. Each person gets one recommendation, ranked by:
$$\text{score} = \text{hours} \times \text{AI usage share} \times \text{deadline weight} \times \text{verify bonus}$$
Every real number shows its source on hover. We had no access to FRAME's real data, so the company data is synthetic and every row says so.
Claude writes the plain-English answers, but only over the engine's numbers, and it names each source. Every Claude call has a fallback, so the demo works with no API key.
Two of us built it in one day. We split the work into tasks on Relay, our own coordination board. Each task's AI agent opened a pull request, and the other person reviewed it.
Challenges we ran into
- A demo click that went nowhere. In our final review we found "I'll try it" saved only in the browser and never reached the AI Lead's log. We moved it to a server route and added a test.
- A score of zero. Bank reconciliation itself scores 0.0 in the Economic Index, yet it was Aoife's top skill. The hours came from the record-keeping work around it, which scores 0.97. We made sure the app could explain that.
- Getting the law right. Article 4 was softened in 2026 to "take measures to promote" literacy, and national regulators enforce it. We checked our wording against the source rather than quote a fine that does not apply to it.
What we learned
- One clear action beats a dashboard. Every view ends in a next step.
- Real public data makes synthetic company data believable, as long as you label which is which.
- Two people can coordinate several AI agents in a day if each task owns its own files and has a test written before the build.
What's next
A pilot with real timesheets to measure the hours saved, rather than estimate them. The engine takes a CSV (a plain spreadsheet file) in the same shape as real timesheets, so any firm can plug in its own.
Built With
- anthropic-economic-index
- anthropic-sdk
- claude-api
- next.js
- node.js
- o*net
- papaparse
- react
- relay
- typescript
- vitest
- zod
