Picture the AI project list at a typical mid-market company this fall. Invoice matching in finance. A help desk assistant in IT. Contract summaries for legal. Every item saves time or cost, and every one is worth doing. Now ask the question a board will eventually ask: which of these makes us more money? On most lists, the honest answer is none of them.
Now ask the question a board will eventually ask: which of these makes us more money? On most lists, the honest answer is none of them.
This article is about the other column. The large surveys keep pointing to the same gap. Deloitte’s 2026 State of AI in the Enterprise, based on 3,235 leaders surveyed in the fall of 2025, found that 74 percent hope AI will grow revenue, while 20 percent say it already has. PwC’s 2026 CEO survey found that 56 percent of 4,454 CEOs reported no significant financial benefit from AI so far. McKinsey’s 2026 State of AI survey of 1,719 respondents found that where companies do report revenue gains from AI, they most often credit marketing and sales. All three are self-reported and weighted toward large companies, but together they say something useful: revenue is where AI is hoped for most and delivered least, and it is a specific function that is delivering it.
Why the Revenue Side Gets Skipped
In our experience, revenue uses get skipped for an understandable reason. A cost saving is easy to see: the invoice took four minutes and now takes one. A revenue gain is hard to see, because sales rise and fall for a dozen reasons at once, and nobody can say for sure that the new tool caused the difference. So budgets drift toward what can be measured.
The fix is not to give up on measurement. It is to design the measurement before you start. Four of the five uses below come with evidence from a controlled study, the fifth is labeled as a correlation, and every one can be run inside a single quarter as a test with a comparison group, which is the discipline we described in The Control Group Is Gone. Here are five uses, ranked by how directly the evidence applies to a mid-market business.
Price Every Quote With a Second Opinion
The best evidence for mid-market B2B comes from an unglamorous place. A study published in Marketing Science in 2024 followed a US metals service center that cuts aluminum and steel for machine shops. For eight business days, quotes from its 17 sales reps were randomly assigned, and on some of them the rep saw a model’s recommended price before quoting. Where the recommendation was shown, gross profit per line rose 11 percent, and the gain came from more quotes being accepted, not from charging more.
The detail that matters most is that the reps followed the recommendation only about one time in five. The model did not replace their judgment; it gave them a second opinion, and they used it when it made sense. It is one firm and a short window, so treat 11 percent as a signal and not a forecast. But if your business quotes custom work, your quote history is already the training data, and a randomized pilot on a share of quotes is a quarter’s work.
The model did not replace their judgment; it gave them a second opinion, and they used it when it made sense.
Answer Leads While They Are Still Warm
The most cited number in sales operations comes from a 2011 Harvard Business Review analysis of 1.25 million leads at 42 companies, which found that firms trying to contact a lead within an hour were nearly seven times as likely to qualify it as firms that waited even one more hour. The same study audited 2,241 companies and found the average response time, among those that answered within 30 days, was 42 hours.
Read that carefully. It is a correlation from 2011, the outcome was a qualifying conversation and not a sale, and one of the authors led a sales software company that likely supplied the data. Fast responders may simply be better run. Still, the gap it describes is easy to close with AI: triage incoming inquiries, draft a first reply for a person to approve, and route the lead to the right rep in minutes rather than days. We know of no rigorous study tying quote or response speed to win rates in B2B, so the honest framing is that this is a cheap test of a plausible idea, not a proven result. For manufacturers, our piece on preparing for AI buyer agents makes the same point about quoting at the speed the buyer expects.
Coach the Middle of Your Sales Team
Field experiments on AI sales coaching point to a pattern that surprises most sales leaders. In three randomized experiments at two financial technology firms, reported in the Journal of Marketing, the reps who gained most from AI coaching were in the middle of the pack. The bottom performers were overwhelmed by the feedback, the top performers resisted it, and a combination of AI and human coaching beat either one alone. The press summary we could read did not give effect sizes, so we can describe the shape of the finding but not its size.
The practical reading, and it is our judgment, is to aim AI coaching at the reps who are close to good: call summaries, a note on what the best calls did differently, and a short weekly review with a manager. Do not expect it to rescue a struggling hire or to impress your top producer.
Keep the Customers Who Will Respond
Most retention programs target the customers most likely to leave. The research says that is often the wrong list. In work published in the Journal of Marketing Research in 2018, Eva Ascarza found that targeting the customers most likely to respond to a retention offer, rather than those at highest risk, reduced churn by 4.1 to 8.7 percentage points in field data from two companies, and that targeting the highest-risk customers sometimes increased churn. One plausible reading is that reaching out to a customer who was quietly thinking of leaving can remind them to go.
That changes what you build. A churn-risk score is not enough. What works is a test: offer the intervention to a random part of the at-risk list, hold back a comparison group, and learn who actually responds before you scale. Our easy button case is a reminder of what retention is worth: the firm reported a six-point gain in retention, worth $3.2 million a year, from better matching of deals to buyers.
Let Service Make the Next Sale Easier
The largest field study of generative AI at work comes from customer service. Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, in the Quarterly Journal of Economics in 2025, studied 5,172 support agents at a software company that sells to small and mid-sized businesses. An AI assistant raised issues resolved per hour by 15 percent on average and by about 30 percent for less experienced agents, and customer sentiment in the conversations improved.
The study did not measure revenue or renewals, so we will not claim it did. Our view is that service is where a mid-market company keeps the customers its sales team won, and that faster, better service is upstream of renewal. If you run this, measure renewal and expansion in the accounts served by AI-assisted agents against a comparison group, not just handle time. A broader set of randomized experiments on an online retail platform, published as a working paper in 2025 and revised this year, found that four of seven AI uses raised sales by up to 16.3 percent, mostly through higher conversion. That was consumer retail, not B2B, but it is consistent with the idea that customer-facing AI can show up in the top line.
Run It as a Test, Not a Rollout
Each of the five uses above has the same shape: a short pilot, a comparison group, and a number agreed in advance. That is the Prove step of Project Vista, where every funded project carries the result it was approved to produce and ends in a verdict to scale it, rework it, or stop.
It is also worth remembering where revenue gains come from in practice. In our regional brokerage case, the firm reported $5 million in attributed new revenue and $3 million in efficiency gains over 24 months on a $1.5 million investment, and there was not a single AI model in it. The gain came from data the sales team could trust. Every use in this article sits on the same foundation; if your CRM cannot tell you which quotes won and why, start there, as we argued in A Single-pane View Has Never Been Easier.
What Can Go Wrong
Customers notice. In a 2019 Marketing Science experiment with more than 6,200 customers of a Chinese financial services firm, an undisclosed AI caller sold as well as skilled reps, but disclosing it was a bot before the conversation cut purchase rates by more than 79 percent. That was an early voice bot, not today’s tools, but it is a warning: decide deliberately where AI speaks to customers directly and where it works behind your people, and check the disclosure rules in the markets you serve.
AI also fails in ways that look like success. The well-known Harvard and BCG study of 758 consultants, published this year in Organization Science, found that on tasks inside AI’s capabilities, people finished more work, faster, and better, but on a task just outside them, they were about 19 percentage points less likely to get the right answer. Those were consulting tasks, not sales, but the lesson carries: a pricing recommendation or a drafted reply needs a person who knows when to ignore it.
AI also fails in ways that look like success.
What This Will Not Do
None of these studies was run at a company exactly like yours. The strongest B2B result comes from one metals distributor over eight days. Much of the causal evidence comes from telecom, consumer retail, and Chinese financial services. The service study measured productivity and sentiment, not revenue. The lead-response figure is an old correlation with a likely vendor connection, and the surveys are self-reported and weighted toward large companies. Our ranking of the five uses, and our suggestions for running them, are our judgment, built on that evidence. What the evidence supports is a set of tests worth running, not a set of results to expect.
Where to Start
Pick one use where you already have the data, most often quoting or retention, and design the test before you buy anything: who gets the new process, who does not, and what number decides the outcome. If you would like an outside view of where your data can support a revenue test, our independent assessment starts there. If you would rather talk it through first, you can reach us through the contact form on our About page.
