The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Agents fail tax work by filling the gaps in the file.
Surge AI's new benchmark gives agents 50 real tax assignments, about 30 documents each, and the best one passes 10.8%. The failures in their write-up aren't misread PDFs. One agent decided a couple wasn't married from their Florida address, while the file showed a Colorado ceremony. Another assignment plants a ledger entry pointing at a receipt that isn't there.
That matches what I see on 200-page loan files. Reading the pages is the easy part. The hard part is a value with no page behind it, because the model fills it with something plausible and the output looks finished.
The Hacker News thread says better tooling around the model would fix most of this. I think it fixes part of it. The tooling has to make the agent cite a page for every answer and accept "not in the file" as a result.
If you think tooling closes most of that 90% gap, what would you add to it first?
The benchmark: https://surgehq.ai/blog/dayjob-tax and the thread: https://news.ycombinator.com/item?id=50275626
surgehq.ai
DAYJOB: Tax — Benchmarking AI Agents on Professional Tax Work | Surge AI
DAYJOB: Tax evaluates AI agents across 50 realistic tax environments, testing their ability to navigate messy records, exercise professional judgment, and deliver reliable tax work.
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Related posts
AI Automation Platform and Workflow Automation & AI-Powered Solutions
An AI-powered automation platform designed to simplify business processes, automate repetitive tasks, and improve overall productivity.
The platform combines a modern, user-friendly interface with intelligent automation to help businesses streamline workflows, reduce manual effort, and work more efficiently.
The motion, the UI, everything's amazing, Incredible work
Skycall 🎈: the dawn call, made automatically
An AI-powered operations brain for hot-air balloon operators, wrapped in a cinematic, living website. Tomorrow’s forecast comes in, the call gets made, and every passenger hears about it, with nobody working the phones.
Every balloon operator knows the feeling: it’s 6 pm and the forecast just changed. The pilot is reading wind at three heights while dozens of passengers set alarms for 4 am. Then the phone tree starts: voicemails, the wrong hotel, a family that flew halfway around the world, told at midnight. Tomorrow evening, the same again. Skycall turns that whole evening into one automated flow, so a small operator can run like a big one.
The problem: dawn flights live or die on the wind, and every “no” lands on a person with a phone. Rebooking is done by hand, empty seats go unfilled, and reviews blame the operator for the sky. The decision belongs to the pilot. The messaging shouldn’t eat their evening.
The solution:
The evening call. At 18:00 Skycall reads tomorrow’s forecast at the surface, 80 m and 180 m, checks it against the operator’s own limits, and calls it GO, HOLD or NO-GO, with the reasons in plain English.
Every passenger told automatically. Each one gets an email with the verdict, meeting point, what to wear and a calendar link.
The dawn recheck. Two hours before launch it looks again, and only writes if the sky changed.
One-tap rebooking. A no comes with two flyable dates, one tap each. Quiet after 24 h? A reminder. Still quiet? Pencilled onto the next flyable dawn.
Waitlist seat fill. On a GO with empty seats, the waitlist gets a 45-minute offer, and the first tap wins.
The crew manifest. The pilot’s phone shows passengers, basket load against the maximum, a hot-dawn warning and the hotel pickup order.
A Monday digest for the owner, plus a public ops log where every automated run is on the record.
A live demo console where anyone can run the call on any sky and watch it happen.
Built on Base44: 15 entities and 21 backend functions sharing one rules engine. The forecast comes live from Open-Meteo, and scheduling runs via an external cron hitting a Base44 backend function. This build runs a demo operator: the operator and bookings are demo data, and the passenger emails go to a demo inbox. The forecast, the decisions and the sends are real.
Skycall is a planning aid. Every verdict is advisory, and the pilot in command decides.
🔗 Skycall
Really interesting concept! The best part is that this goes beyond weather alerts and automates the entire operational workflow, from flight decisions and passenger notifications to rebooking and waitlists. Keeping the final decision with the pilot is an especially important touch. Great use of automation!
5 Common AI Prompting Mistakes That Lead to Generic Results
If AI keeps giving you generic answers, the problem might not be the model. It might be how you're prompting it.
AI tools like ChatGPT, Gemini, and Claude can help with writing, research, brainstorming, and business tasks. But the quality of their output depends partly on how clearly you communicate what you need.
Here are five common mistakes to avoid:
1. Being too vague.
“Write a good marketing email.”
Good for whom? About what? With what goal? The AI has to guess what you trully want, often leading to generic and/or false answers.
Clearly define the objective, audience, and desired outcome.
2. Leaving out context.
AI doesn't automatically know your business, brand voice, product, or situation. It can do a lot of thing, but it can't read minds... yet, maybe in the future. You should always provide the relevant background information it needs to produce a useful response.
3. Giving conflicting instructions.
Asking for a detailed explanation in 50 words, for example, can create competing demands. Prioritize your requirements and clarify what matters most when trade-offs are necessary.
Just imagine you're giving instructions to a toddler, you'd be surprised on how similar those two actually are. The core strategy is to say exactly what you want them to do using short, direct commands instead of asking questions or using negative phrases.
4. Forgetting to define the output format.
Need a table, email, checklist, or step-by-step guide? Say so. Specifying the format makes the result easier to use and reduces unnecessary restructuring.
We're trying to avoid you having to manually reformat a massive wall of text. You have a tool to do it for you, so use it.
5. Accepting the first answer without reviewing it.
Even a well-written prompt can produce inaccurate facts, weak reasoning, or unsuitable content. Review the output, verify important claims, and refine your instructions when necessary.
You'd never take something a toddler told you as law, and you should have the exact same behavior with AI.
The bigger lesson?
A prompt isn't about finding a magic phrase. It's about giving AI the information and direction it needs to do the job well, nothing more.
Better prompts won't guarantee perfect answers, but clear objectives, relevant context, and thoughtful iteration can make AI outputs more useful and consistent.
So, before switching models because you're unhappy with the results, try improving the instructions first.
Which of these prompting mistakes do you see most often?
