AI agents for business: what the evidence says about ROI in 2026

The adoption gap in numbers

The headline figures on AI in business point in two directions at once. In McKinsey's November 2025 State of AI survey of 1,993 participants, 88% of respondents said their organisation regularly uses AI in at least one business function.^[1] B In the same survey, 23% were scaling an agentic AI system somewhere in the enterprise and a further 39% were experimenting with agents.^[1] B And yet only 39% attributed any EBIT impact to AI at all, and only about 6% reported an EBIT effect of 5% or more, McKinsey's bar for a "high performer".^[2] B We read these through secondary coverage because McKinsey's own pages did not load for us; the figures are consistent across two outlets.

The pattern repeats elsewhere. In PwC's 2025 Global CEO Survey, 56% of chief executives reported efficiency gains in employees' time from generative AI over the previous year, but only 32% saw revenue increases and 34% saw profitability gains, against 46% who had expected them.^[3] A Deloitte's survey of 2,773 senior leaders found 26% exploring autonomous agent development to a large extent and 42% to some extent, with regulatory compliance the top barrier for 38%.^[4] A Gartner, an analyst firm, predicts that over 40% of agentic-AI projects are cancelled by the end of 2027 because of costs, unclear value or weak risk controls; that is a press-release headline we could not read in full and we grade it C.^[5] C

For an owner of a small or mid-size business the gap between 88% and 39% is the whole story. Almost everyone is using AI; a minority can show it on the profit line. The rest of this page is about ending up in the minority, which means choosing use cases with evidence behind them and pricing the review work honestly.

Demand for the question is new. US searches for "ai agents for business" average about 3,600 a month, up 5,556% year on year, with an advertiser cost per click of $87.72; "ai customer service agents" averages 1,600 a month at $142.69 per click (DataForSEO, retrieved 2026-11-09). Those click prices tell you how much vendors expect to earn from your decision. Read the evidence before you read the ads.

Use cases ranked by evidence of payoff

We ranked the common business use cases by the quality of independent evidence that they improve a measured outcome, not by how often they appear in vendor decks. "Agent" here covers the whole range from an assistant that drafts for a human to a system that acts on its own; we note which level the evidence covers, because most of it is about the assisted end.

Rank	Use case	What independent studies found	Level of autonomy studied	Grade
1	Customer support (chat and ticket replies)	5,172 agents at one firm resolved about 14–15% more issues per hour with AI-suggested replies; gains of 30% or more for the least experienced, near zero for the most experienced; customer sentiment improved	Assistant drafts, human sends	A [6]
2	Routine business writing (emails, summaries, descriptions)	453 professionals finished mid-level writing tasks about 40% faster with about 18% higher rated quality; weaker writers gained most	Assistant	A [7]
3	Analysis and consulting-style tasks	758 consultants were 12.2% more productive and 25.1% faster inside the model's competence, but 19 percentage points less likely to be correct on a task just outside it	Assistant	B [8]
4	Software development	One task experiment: 55.8% faster with a coding assistant (95 developers); a 2025 randomised study on real tasks: experienced developers 19% slower with agentic tools allowed	Assistant to agent	B/C [9] [10]
5	Internal operations and admin (data entry, routing, scheduling)	No controlled study we could find; vendor case studies only	Workflow to agent	C
6	Sales outreach and marketing campaigns	No controlled study we could find; vendor case studies only	Agent	C
7	Decisions (pricing, credit, hiring screens)	Across 106 experiments, human plus AI performed worse than the better of the two alone on decision tasks; gains were concentrated in content creation	Assistant	A (negative) [11]

Two honest caveats. The strongest evidence is for assistants that draft while a person sends, not for autonomous agents. And the two use cases with the loudest marketing, sales and marketing agents, have no independent outcome data we could find; that is not proof they fail, only that you would be the experiment. Our page on why AI agent projects fail collects the documented failure patterns.

Customer service agents: what the studies show

The customer-support study is the best evidence a business owner has, so it is worth reading closely. The AI did not answer customers; it suggested replies and the human agent chose. Productivity rose about 14–15% on average, measured as issues resolved per hour, and the gain was almost entirely among newer and lower-skilled staff, with the most experienced agents gaining close to nothing. Customer sentiment improved and staff turnover fell.^[6] A The study was quasi-experimental, at one firm, on one task type.

What that means for a small business: an assisted support desk pays off fastest where you have turnover, new hires or a long ramp-up. It pays off least where one experienced person already knows every answer. A fully autonomous support agent, the product category with the highest click prices in our data, is a different and less-studied thing: the benchmarks for tool-using agents in retail and airline customer service show even state-of-the-art agents succeeding on fewer than half of tasks and inconsistent across repeated runs.^[12] B Start assisted, measure resolution time and reopens, and only then remove the human from the send step for the narrowest ticket types.

Sales and marketing agents

Content production is where the writing evidence applies: first drafts of product descriptions, campaign copy and follow-up emails are exactly the mid-level writing tasks that got about 40% faster in the randomised experiment.^[7] A A person still edits and sends.

Autonomous prospecting agents that find leads, write personalised outreach and send it without review sit at the other end. We found no controlled evidence of their effect on revenue, and two documented risks apply: inconsistency across runs, and reputational cost when an unsupervised message is wrong about the recipient. Searches for "ai sales agents" fell 45% year on year and "ai marketing agents" fell 34% (DataForSEO, retrieved 2026-11-09), which is at least consistent with early buyers cooling. Treat a sales agent as a drafting tool with a human send gate until you have your own reply-rate data.

Internal operations and admin agents

This is where most small businesses actually start, and where the evidence is thinnest: invoice matching, inbox triage, meeting notes, moving data between systems, scheduling. There are no randomised trials. What exists is engineering experience, and it points toward workflows rather than agents: a fixed sequence with a model filling one or two steps is testable, while an open-ended agent in your accounting system is not.

The automation platforms are the practical layer. Zapier offers a free tier of 100 tasks a month and a Professional plan from $19.99 a month billed annually or $29.99 monthly for 750 tasks (checked 2026-11-09);^[16] Make and n8n are the alternatives, with n8n also self-hostable, and both publish prices on their sites (see pricing). All three support approval steps, run logs and execution caps, which are the three controls that make a level-3 agent safe to run. For meeting notes specifically, Fathom has a free plan and Team seats at $15 per user per month billed annually or $19 monthly (checked 2026-11-09).^[17] Our comparison of no-code AI agent builders covers which platform fits which job.

Cost model and break-even

No vendor will give you a break-even, so here is the arithmetic, with the inputs you have to supply.

Monthly benefit = hours the task takes now × realistic gain × loaded hourly cost of the person doing it.

Monthly cost = licences and platform fees + model or task usage + setup time amortised + review time (hours spent checking and correcting × loaded hourly cost).

Three rules make the estimate honest:

1. Take the gain from the table above for the nearest studied use case, not from the vendor. For assisted support, 14–15% on average; for routine writing, up to 40% on the drafting step only; for anything in rows 5 to 7, start at zero and measure.
2. Budget review time explicitly. With agent completion rates below 50% on open tasks in the benchmarks, every unreviewed run is a liability, and every reviewed run costs minutes.^[12] If review takes longer than doing, stop.
3. Count the expert's time at the expert's rate. The gains concentrate among less-experienced staff; an agent handed to your most experienced person may return nothing.^[6]

An illustration with made-up inputs, so you can see the shape: a task taking 40 hours a month at a loaded cost of $40 an hour is $1,600 of time. A 15% gain is worth $240. If the tool and platform cost $60 and review takes four hours ($160), the net is $20 a month, before setup. Change the gain to 40% on a drafting step and the net turns clearly positive; change the review to eight hours and it turns negative. The result swings on the review line, which is the one vendors omit. The Tool Finder runs this estimate with your numbers and the dated prices from our catalog.

Governance: what the EU AI Act asks of a business that deploys agents

If you operate in the EU, two sets of duties apply to you as a deployer, not just to the vendor.

The AI Act's AI-literacy duty (Article 4) has applied since 2 February 2025: organisations that deploy AI systems must ensure the staff who operate them have sufficient understanding of how they work. The high-risk obligations under Annex III, which include AI used for recruitment, task allocation and monitoring or evaluating worker performance, now apply from 2 December 2027 after the 2026 Omnibus amendment, and deployers of those systems take on duties such as human oversight and logging.^{[13] [14]} A Inferring employees' emotions at work is a prohibited practice.^[14] B Separately, GDPR Article 22 gives people a right not to be subject to solely automated decisions with legal or similarly significant effects, which is directly relevant to any agent that approves, rejects or ranks people.^[15] B

In practice: an agent that drafts support replies or reconciles invoices is low-risk and mostly a literacy and data-protection matter; an agent that screens applicants or scores staff is high-risk by definition and needs a compliance plan before the pilot. A simple run log and a written brief per agent are the lightest documentation that moves you toward both. This is not legal advice; check your own counsel and your vendor's terms.

Checklist before you buy

• The use case is in rows 1 to 4 of the table, or you have a plan to measure it yourself.
• You know which autonomy level you are buying and where the human gates are (send, pay, delete, publish).
• You have a baseline: current hours, error rate and outcome measure for the task, recorded before the pilot.
• Review time is in the budget, at the rate of the person who will do it.
• The first month runs with draft-only or sandbox permissions.
• You know which data leaves your environment, to which model provider, under what training and retention terms.
• The vendor can show completion rates on a sample of your tasks across repeated runs, not a single demo.
• The pilot has a stop rule: a date and a number below which you switch it off.

Our comparison of the best AI agents for work applies this checklist to named products.

Where demand goes from here

In our Future Demand Radar, "AI agents and delegated work" scores 59.7 out of 100 in the baseline scenario (band 55.2–64.1) and is classed "build now"; "team productivity and AI collaboration" scores 51.3 (band 48.6–54.1), classed "validate in 12 months". Our baseline scenario expects search demand for agent-related terms to reach about 1.7 times today's level by 2028, with a trough risk in 2027–28 if the project cancellations Gartner predicts materialise. Under the conservative scenario the category drops to 56.0 and stays "build now". Forecasts are model outputs with stated uncertainty, not facts.

What to do next

Pick one task from rows 1 to 4, record its baseline this week, and run the Tool Finder: it asks for the hours, the hourly cost, your platforms and your tolerance for review, and returns the break-even and a shortlist from our catalog with prices dated 2026-11-09. Then run the pilot with a human send gate and a stop rule.

Frequently asked questions

Do AI agents actually deliver ROI for small businesses?

Sometimes, and mostly at the assisted end. The best evidence is a study of 5,172 support agents whose productivity rose about 14–15% with AI-suggested replies, concentrated among newer staff.^[6] A Across organisations, 88% use AI but only 39% report any EBIT impact.^{[1] [2]} B The difference is use-case choice and whether review time was budgeted.

Which business use case should I automate with an AI agent first?

Customer-support drafting or routine writing, because they have grade-A evidence of benefit and a cheap check: a person reads the draft before it goes out.^{[6] [7]} Decisions about people or money come last; a meta-analysis of 106 experiments found human-plus-AI combinations under-performing the better single agent on decision tasks.^[11] A

What does an AI agent cost a business per month?

Licence or platform fees plus usage plus review time. Zapier Professional starts at $19.99 a month billed annually for 750 tasks and Fathom Team seats at $15 per user per month billed annually (checked 2026-11-09); the larger cost is usually the hours a person spends checking output, which scales with the error rate. Use the break-even formula on this page or the Tool Finder.

Does the EU AI Act apply to a small business that only uses AI agents?

Yes, as a deployer. The AI-literacy duty (Article 4) has applied since 2 February 2025; high-risk obligations for Annex III uses such as recruitment, task allocation and performance monitoring apply from 2 December 2027 after the 2026 Omnibus amendment.^{[13] [14]} A Drafting replies or reconciling invoices is low-risk; screening applicants is high-risk. This is not legal advice.

Are AI sales and marketing agents worth it?

For drafting campaign copy and follow-ups, the writing evidence applies: about 40% faster in a randomised experiment, with a person still sending.^[7] A For autonomous prospecting and sending, we found no controlled evidence of revenue effect, and searches for "ai sales agents" fell 45% year on year (DataForSEO, retrieved 2026-11-09). Keep a human send gate until you have your own reply-rate data.

Sources and evidence

1. The state of AI in 2025: Agents, innovation, and transformation (as reported by ppc.land) — McKinsey & Company (QuantumBlack) (2025), report B ppc.land/most-companies-still-pilot-ai-programs-despite-widespread-ado (retrieved 2026-11-09)
2. The state of AI in 2025: EBIT impact and high performers (as reported by IDM) — McKinsey & Company (QuantumBlack) (2025), report B idm.net.au/node/15415 (retrieved 2026-11-09)
3. PwC's 28th Annual Global CEO Survey (press release) — PwC (2025), report B www.pwc.com/gx/en/news-room/press-releases/2025/pwc-2025-global-ceo-su (retrieved 2026-11-09)
4. The State of Generative AI in the Enterprise: Generating a new future (Q4 report) — Deloitte (2025), report A www.deloitte.com/us/en/about/press-room/state-of-generative-ai.html (retrieved 2026-11-09)
5. Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 (press release; analyst estimate, headline verified via search only, page returned 403) — Gartner (2025), report C www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts (retrieved 2026-11-09)
6. Generative AI at Work — Brynjolfsson, Erik; Li, Danielle; Raymond, Lindsey (2025), observational A doi.org/10.1093/qje/qjae044 (retrieved 2026-11-09)
7. Experimental evidence on the productivity effects of generative artificial intelligence — Noy, Shakked; Zhang, Whitney (2023), RCT A doi.org/10.1126/science.adh2586 (retrieved 2026-11-09)
8. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua, Fabrizio; McFowland III, Edward; Mollick, Ethan R.; Lifshitz-Assaf, Hila; Kellogg, Katherine; Rajendran, Saran; Krayer, Lisa; Candelon, François (2023), RCT B doi.org/10.2139/ssrn.4573321 (retrieved 2026-11-09)
9. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot — Peng, Sida; Kalliamvakou, Eirini; Cihon, Peter; Demirer, Mert (2023), RCT C doi.org/10.48550/arXiv.2302.06590 (retrieved 2026-11-09)
10. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — Becker, Joel; Rush, Nate; Barnes, Elizabeth; Rein, David (METR) (2025), RCT B doi.org/10.48550/arXiv.2507.09089 (retrieved 2026-11-09)
11. When combinations of humans and AI are useful: A systematic review and meta-analysis — Vaccaro, Michelle; Almaatouq, Abdullah; Malone, Thomas (2024), meta-analysis A doi.org/10.1038/s41562-024-02024-1 (retrieved 2026-11-09)
12. τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains — Yao, Shunyu; Shinn, Noah; Razavi, Pedram; Narasimhan, Karthik (2024), report B doi.org/10.48550/arXiv.2406.12045 (retrieved 2026-11-09)
13. AI Act | Shaping Europe's digital future (Regulation (EU) 2024/1689, application timeline incl. 2026 Omnibus amendment) — European Commission (2026), report A digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai (retrieved 2026-11-09)
14. EU AI Act, Article 5 (prohibited practices) and Annex III (high-risk systems incl. employment) — Future of Life Institute, artificialintelligenceact.eu (2024), report B artificialintelligenceact.eu/article/5/ (retrieved 2026-11-09)
15. GDPR Article 22: Automated individual decision-making, including profiling — gdpr-info.eu (Intersoft Consulting) (2016), report B gdpr-info.eu/art-22-gdpr/ (retrieved 2026-11-09)
16. Zapier pricing — Zapier (2026), vendor page A zapier.com/pricing (retrieved 2026-11-09)
17. Fathom pricing — Fathom (2026), vendor page A fathom.ai/pricing (retrieved 2026-11-09)