Are AI agents worth it for business?

What the evidence says

Most "ROI" numbers you'll see come from vendors. Treat them as marketing. Intercom, for example, quotes customers reporting a 50% resolution rate and a 96% cut in resolution time4; those are real testimonials, but they are chosen by the seller.

The strongest independent evidence is the study "Generative AI at Work" by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in the Quarterly Journal of Economics in 20253. Looking at 5,179 customer support agents, it found that an AI assistant:

• raised issues resolved per hour by 14% on average3;
• helped novice and lower-skilled workers most, at 34%, with minimal impact on experienced, highly skilled workers3;
• improved customer sentiment and employee retention3.

Two lessons follow. AI that helps people has the best track record. And the gains are biggest where people are still learning the job. Good, independent evidence on fully autonomous agents running business processes end to end: unknown. We haven't found a study of the same quality.

What agents actually cost

Hidden costs to budget for: setup time, someone to own it, cleaning up the help articles or data it relies on, and usage beyond the included limits. ChatGPT Business and Claude Team both sell extra credits12; Lindy pauses when credits run out until you top up at $10 per 1,0009; unused Agentforce Flex Credits don't roll over10.

A break-even you can run yourself

This is an illustration with made-up inputs; plug in your own. Say a seat costs $25 a month and an hour of the person's time costs you $40. The seat pays for itself if it saves 37.5 minutes a month (that's $25 ÷ $40). If it saves an hour a week, about 4.3 hours a month, you get back roughly $172 for a $25 spend.

For per-outcome pricing, compare cost per resolution. If a ticket handled by a person costs you $6 in time and the AI resolves it for $0.994, each resolution saves about $5, provided the customer really was helped. Check customer ratings, not just resolution counts.

The question isn't really "is AI worth $25?" It's "will people actually use it every week?" A seat nobody opens is worth nothing. Start monthly, measure use, then switch to annual (about 20% cheaper on ChatGPT and Claude seats12).

Where agents are worth it

• High-volume, repetitive questions with good written answers, such as support FAQs. See the customer support guide.
• First drafts: emails, proposals, reports, slides. A person still reviews.
• Research and summaries of many sources. See the research guide.
• Newer staff, where the study found the biggest gains3.
• Admin glue: inbox triage, scheduling, CRM updates. See email and calendar and sales and CRM.

Where they're not (yet)

• Anything irreversible without review: payments, contracts, deleting data. OpenAI warns agents can be manipulated by hidden instructions on web pages and in email (prompt injection), and that safeguards don't remove all risk13.
• Work you can't measure. If you can't tell whether the output is right, you can't tell whether it's worth it.
• Self-hosting without a technical owner. OpenClaw, for example, has published hundreds of security advisories this year that someone has to keep patched14. See Is OpenClaw safe?
• Experts who are already fast. The study found minimal gains for the most experienced, highly skilled workers3.

A 30-day test

1. Pick one workflow and one measure (tickets per hour, hours on reports, reply rate).
2. Record the current number for two weeks.
3. Buy monthly seats for the people who do that work, minimum two on most business plans12.
4. Train them for an hour: what to delegate, what to check.
5. After 30 days, compare. Keep it if the gain beats the cost and quality held; otherwise cancel.

Choosing a tool? Start with the buyer's guide, compare ChatGPT vs Claude or Grok Bot vs ChatGPT dots, and see every price on /pricing.