AI Writing ROI in 2026: Beyond Apollo's Zero-Payoff Claim
What one software renewal can tell you that an economic headline cannot.
October 25, 2026
AI writing ROI depends on what a tool helps you finish, what that result is worth, and its full cost. Apollo's latest report questions whether AI spending is improving broad corporate results. It doesn't establish whether a particular writing tool pays for itself in your business. Keep that distinction when deciding which subscriptions to renew.[1]
Apollo's report led Techmeme when I checked on October 25. The claim is blunt. Torsten Slok argues that revenue per worker and profit margins are flat for the S&P 493, while those measures are rising for the Magnificent Seven. His note is dated October 24. Today's attention is fresh; the research wasn't published this morning.[1][6]
I publish this on DictaFlow's site, so I have a commercial interest in voice software. That makes the buying question worth asking about our product too. A busy microphone or a long AI-generated document isn't a business result.
Apollo's claim has a different denominator
Slok uses broad company and economic measures to question the payoff from AI spending. That is a useful challenge to claims that adoption alone creates value. But a flat aggregate cannot tell you whether one employee's correspondence got easier. Nor does a faster reply prove that company-wide margins improved.[1]
Larry Dignan's October 25 analysis at Constellation Research offers a useful counterweight. He cites Slok's concerns while arguing that useful enterprise AI projects can coexist with shaky infrastructure economics. His assessment is independent commentary, not a controlled measurement of a writing product's returns.[5]
For an owner buying software, both pieces leave work to do. You need evidence from the task you are paying to change. If a vendor advertises better productivity, ask which cost disappears, which capacity becomes usable, or which result improves. Those are different claims and need different records.
Start with one recurring document. A proposal revision is a better candidate than a vague goal to use more AI. You can inspect whether the revised proposal answers the request and whether the customer comes back confused. You also know how often you face that task.
A cheaper draft can still be an expensive reply
Consider a fictional repair business sending customers updates about equipment awaiting parts. The owner knows the status. They need to explain it clearly, without promising a delivery date that nobody has confirmed.
The original note says a supplier expects an update on Thursday. An assistant rewrites that as a part arriving on Thursday. The sentence is fluent and the email took less effort to draft. It also created a promise the owner cannot support. Any later correction belongs in the cost of producing that reply.
Compare the actual alternatives. Typing from scratch might work well. A saved template might solve the repeated wording. Dictation could help someone who knows what to say but struggles to get it onto the page. A generative assistant could help organize scattered notes. You don't need to buy all four approaches to test that question.
Give each method the same source facts and judge the saved message. Keep the supplier's update date separate from the delivery date. Check that the customer can understand what happens next. If one version needs a follow-up to repair a misunderstanding, attach that work to the original task rather than treating it as an unrelated email.
I would reject a faster method that repeatedly creates false commitments. A dollar estimate of someone's time won't rescue a message the business shouldn't send. Once the versions meet the same quality standard, comparing their costs becomes useful.
Put the renewal decision on one page
For a small trial, record the tasks attempted, the usable results, the time spent correcting them, and the subscription or usage charge. Include the setup work. If another person has to review every output, count their time too. These are suggested trial records, not findings from Apollo's report.
Here is a purely illustrative calculation. Suppose a tool costs $20 for the month. You complete 80 suitable tasks and save three minutes on each, after checking the result. That is four hours of released capacity. At an assumed $30 an hour, the capacity has an imputed value of $120. Subtract the $20 fee and the arithmetic leaves $100 before setup or other costs. Those are invented inputs, not a forecast or a product benchmark.
That $100 is not automatically cash in the bank. If you still pay the same wages and produce the same work, the immediate benefit may be breathing room. It becomes a different business case if the capacity lets you avoid overtime or complete additional paid work. Record which outcome actually occurred instead of reporting an assumed hourly rate as money earned.
Usage changes the decision too. If the same monthly tool helps with only eight tasks, three minutes each releases 24 minutes. At that same assumed rate, the capacity value is $12, below the $20 fee. A tool can be pleasant to use and still fail that narrow renewal test.
Accessibility deserves its own line. A person who cannot comfortably type may need a tool to complete the work at all. Comparing their voice workflow with an imaginary painless typing session would give the wrong answer. Record whether the task became possible and what help it still required. Don't reduce that decision to a typing-speed contest.
Test the input problem before buying more generation
If your recurring problem is getting your own explanation into a text field, try an input tool before paying for more generated prose. DictaFlow's documented desktop workflow is hold, speak, release, with text inserted at the cursor. That makes it a candidate for the repair-business example, where the owner already has the facts and wording.[4]
The current DictaFlow product facts distinguish ordinary paste from Typing Mode for fields that resist paste. That is a practical thing to test in the actual work app, not a reason to assume every field will work perfectly. Count corrections and failed insertions in the trial. A free built-in tool or an existing template may already meet your needs.[4]
Try the next batch of routine updates without changing the business's approval process. Keep the original facts beside each saved reply. When the subscription comes up for renewal, inspect those messages and the work attached to them. You should be able to say which task improved, how often it occurred, and whether the benefit justified the cost. Leave the broader claim about corporate AI returns to evidence broad enough to answer it.
Sources
[1] Apollo: No Signs of AI Payoff Outside the Tech Sector, October 24
[4] DictaFlow: current consumer product facts