# Looks Amazing! But Does It Sell More Burritos?

AI work looks impressive. Does it change the outcome? How to count the full cost, the maker's time and the reader's, against the results that matter.

By Steven Cappiello, Founder of JobKick · Published 2026-10-23
Canonical: https://jobkick.com/blog/looks-amazing-does-it-sell/
Topics: AI ROI, measuring AI productivity, cost of AI at work, AI at work, workslop, small business technology

*Part 2 of Looks Amazing. [Part 1](/blog/ai-recap-nobody-reads/) was about AI work nobody checked. This one asks whether it pays.*

I own a burrito shop. I put in all these ordering kiosks, and they're really amazing. They also cost me about $1,000 a month in hardware.

So here's the question I should have been asking from day one: am I selling that many more burritos because of them? I don't think so. But I don't actually know, because I haven't been measuring it. I should be.

![A heavy old mechanical cash register in green metal, with rows of number keys and a crank handle on the side, by a window.](/blog/looks-amazing-does-it-sell/old-cash-register.webp)

*The kiosk is the part everyone sees. The register is where you find out if it worked. Photo: rawpixel, CC0 public domain (via [Openverse](https://openverse.org/image/58203a04-1b1b-4ad7-916a-63826606768b)).*

That's the question I keep coming back to with AI at work. The output looks amazing. The deck is polished, the plan is long, the recap is on time. Did any of it change the outcome? Did it sell more burritos?

## The impressive part is easy to see. The result isn't.

New tools are great at looking like progress. You can see the kiosk. You can see the twelve-page plan. You can't see a sale that didn't happen, or the hour three people spent reading something that didn't need to exist.

And when people do check, the results so far are humbling:

- **25%** of AI initiatives delivered the ROI CEOs expected (source: IBM CEO study, 2025 · ~2,000 CEOs, https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles)
- **19%** slower with AI, while believing they were 20% faster (source: METR, 2025 · 16 developers, 246 tasks, https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)
- **≤2%** effect of AI chatbots on pay or hours: none found, and anything above 2% ruled out (source: Humlum & Vestergaard, NBER 2025 · 25,000 Danish workers, https://www.nber.org/papers/w33777)

IBM asked about 2,000 CEOs in 33 countries in early 2025. They said only 25% of their AI initiatives had delivered the return they expected, and only 16% had scaled across the company. A widely shared MIT NANDA report went further: for 95% of companies in its dataset, generative AI was "falling short." That one rests on 150 interviews, a survey of 350 employees and 300 public deployments, so treat it as a strong signal, not a census.

The second number is the one that gets me. In early 2025, the research group METR ran a controlled trial with 16 experienced software developers on 246 real tasks. With AI allowed, they took 19% longer. Beforehand they expected AI to make them 24% faster. Afterward they still believed it had made them 20% faster. They felt faster and were slower.

To be fair to the tools, METR has since said those results are out of date. Its February 2026 follow-up, with 57 developers, points to a speedup now. But METR also calls its own estimate "a bad proxy" for the real impact, because developers who didn't want to work without AI declined to take part. Measuring this is hard even for people whose whole job is measuring it. That's not a reason to skip it. It's a reason not to trust the feeling.

The third number comes from Denmark, where economists linked survey answers from 25,000 workers in 7,000 workplaces to official pay and hours records. Two years after ChatGPT launched, they found "precise null effects" on earnings and recorded hours, ruling out effects larger than 2%. Workers did report saving time, and 85% of users said they put it into other tasks. The time went somewhere. It just didn't show up in the numbers yet.

## The cost nobody counts: the reader

When we judge AI work, we usually count one cost: how long it took to make. With AI, that number is tiny, so it feels free.

But in [part 1](/blog/ai-recap-nobody-reads/) the workslop researchers put a price on the other side: about two hours to deal with each piece of AI work that missed the mark, and roughly $186 per employee per month. The effort didn't disappear. It moved from the person who made it to the people who have to read it.

The math is simple. A 2019 review of 190 reading studies put the average adult reading speed for non-fiction at 238 words per minute. Try your own numbers:

**Chart (calculator): The full cost of one AI document** — An estimate. Set the length, the readers and the time it took to make. The hourly cost is only an example; use your own.

```json
{"inputs":[{"id":"words","label":"Length of the document","min":200,"max":15000,"step":100,"value":3000,"unit":"words"},{"id":"readers","label":"People who receive it","min":1,"max":50,"step":1,"value":12,"unit":"people"},{"id":"make","label":"Time to make it (prompting and checking)","min":5,"max":240,"step":5,"value":30,"unit":"min"},{"id":"rate","label":"Hourly cost of a person (example)","min":10,"max":200,"step":5,"value":40,"unit":"$/hr"}],"constants":[{"id":"wpm","label":"average adult silent reading speed, non-fiction","value":238,"unit":"words/min","source":{"label":"Brysbaert 2019, 190 studies","url":"https://osf.io/preprints/psyarxiv/xynwg"}}],"outputs":[{"label":"Total cost to make it and read it","expr":"(readers * words / wpm + make) / 60 * rate","format":"currency"},{"id":"readHrs","label":"Hours spent reading it, all readers","expr":"readers * words / wpm / 60","unit":"hrs","format":"1dp"},{"id":"makeHrs","label":"Hours spent making it","expr":"make / 60","unit":"hrs","format":"1dp"},{"label":"Share of the cost paid by readers","expr":"readHrs / (readHrs + makeHrs) * 100","format":"pct"}],"chart":{"outputs":["Hours spent reading it, all readers","Hours spent making it"],"unit":"hrs"}}
```

Reading speed: Brysbaert, "How many words do we read per minute?" (2019) (https://osf.io/preprints/psyarxiv/xynwg), meta-analysis of 190 studies, 18,573 participants. Assumes everyone reads it once, all the way through.

With the defaults, a 3,000-word plan sent to 12 people takes half an hour to make and two and a half hours to read. The readers pay about 83% of the bill. And that's before anyone checks it, argues about it, or works from the wrong version.

The kiosk works the same way. The $1,000 a month is the easy part to see. The real question is what changed at the register.

## In fairness

Sometimes it clearly pays. The best-known study of AI at work followed 5,179 customer support agents. With an AI assistant, they resolved 14% more issues per hour on average, and newer, less experienced agents improved by 34%. That's a real outcome, measured where the work happens.

The lesson isn't "AI doesn't work." The ones that work are measured on the result, not the output. Nobody counted how many words the support agents wrote. They counted problems solved.

## What to do instead

1. **Name the number before you start.** Sales, tickets closed, hours saved, mistakes caught. If you can't name it, you're buying something that looks good.
2. **Count both sides of the cost.** Making it and reading it. A tool that saves the writer 30 minutes and costs twelve readers 30 minutes each loses money.
3. **Measure before and after.** Even roughly. Weeks with and without, the same days of the week, the same menu. A rough measure beats a confident feeling, and the developers in the METR study show how wrong the feeling can be.
4. **Try the simpler version first.** A one-paragraph update, a shorter menu board, a phone call. If the simple thing gets the same result, keep the simple thing.
5. **Set a date to decide.** Three months in, look at the number. Keep it, change it, or cancel it. Subscriptions are easy to start and easy to forget.

## What's the most impressive thing you've paid for that you never measured?

AI tool, kiosk, software, anything. Did it pay? How do you know? Tell me in the comments. I'm especially curious how small businesses measure this without a data team.

## Sources

- IBM Institute for Business Value, "[IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles](https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles)," May 2025 (about 2,000 CEOs, 33 countries, Feb–Apr 2025).
- Fortune, "[MIT report: 95% of generative AI pilots at companies are failing](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/)," Aug. 2025, on MIT NANDA's *The GenAI Divide: State of AI in Business 2025* (150 interviews, 350-employee survey, 300 public deployments).
- METR, "[Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)," July 2025 (16 developers, 246 tasks), and its [February 2026 update](https://metr.org/blog/2026-02-24-uplift-update/) (57 developers, 800+ tasks).
- Humlum & Vestergaard, "[Large Language Models, Small Labor Market Effects](https://www.nber.org/papers/w33777)," NBER Working Paper 33777, 2025 (25,000 workers, 7,000 workplaces, 11 occupations, Denmark).
- Niederhoffer et al., "[AI-Generated 'Workslop' Is Destroying Productivity](https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity)," *Harvard Business Review*, Sept. 2025 (1,150 US full-time employees).
- Brysbaert, "[How many words do we read per minute? A review and meta-analysis of reading rate](https://osf.io/preprints/psyarxiv/xynwg)," *Journal of Memory and Language*, 2019 (190 studies, 18,573 participants).
- Brynjolfsson, Li & Raymond, "[Generative AI at Work](https://www.nber.org/papers/w31161)," NBER Working Paper 31161, 2023 (5,179 customer support agents).