Insights

Should your scale-up do more with agentic AI? Start where the evidence is

By Maya Moufarek, founder of Forward Operating Partners · 11 November 2026

Most founders I speak to already have agents running somewhere, usually a support bot and a few automations someone built on a Friday. The question they bring me is whether to do more, and where.

My answer is to start where the published evidence is, and to be honest about where it isn't. Most guides on agentic AI list use cases by how easy they are to build. Very few sort them by whether anyone has shown they move a number. That's the list I've put together below, from the field experiments and large studies published so far.

Should our scale-up be doing more with agentic AI?

Yes, a scale-up should do more with agentic AI in the few places where published evidence shows agents move a business number, and it should start with customer support, where an AI agent helps a person. High-volume screening also has strong evidence, while B2B sales, demand generation and retention have little or none yet.

So "more" is the wrong frame for most scale-ups. You probably have enough agents. The useful decision is which ones to keep and which to stop, and anything in between gets measured like the experiment it is.

The best long-run data comes from Denmark, where records for about 25,000 workers rule out effects on earnings or hours above 2% two years after adopting AI (BFI, 2025). Gains on single tasks are real, and they reach the P&L only when someone redesigns the work around them.

Where have AI agents been shown to work?

AI agents have the strongest published results in customer support, where an assistant helps a human agent resolve 14% more issues an hour and new agents 34% more, and in high-volume screening interviews, where AI interviews raised job offers 12% and starts 18%.

The support result comes from a study of 5,172 agents at a company selling business software to small and mid-sized firms, so it is B2B support (Brynjolfsson, Li and Raymond, NBER). The later QJE version of the paper reports 15%. New agents with two months of AI help performed like agents with more than six months' experience working without it (NBER). The most experienced people gained little.

The hiring result comes from 70,884 applications for entry-level customer service roles. Applicants interviewed by an AI voice agent were 12% more likely to get an offer, 18% more likely to start and 17% more likely to still be there after 30 days (Jabarian and Henkel). The starts and retention figures come from the 2025 draft of the paper.

The evidence for pre-sale chatbots and the finance close is weaker. A pre-sale chatbot on a large consumer marketplace lifted sales 16.3% and conversion 21.7%, though this is B2C, the figures were read from a conference paper summary, and two authors worked with the platform (Fang et al.). And across 79 small and mid-sized firms on one accounting platform, AI adoption went with a monthly close 7.5 days faster. That one is observational, from a summary of the study (Choi and Xie, Stanford GSB).

In every one of these studies the gains were biggest where a person stayed in the loop, and for the least experienced staff.

How long until AI agents show results?

AI agents show results in 2 to 8 weeks on support and front-of-funnel metrics and in 2 to 8 months on finance and engineering metrics, and they need at least one renewal cycle before churn, net revenue retention or margins can move.

No published study has yet shown churn, NRR or margins move because of generative AI. With tens of support agents and hundreds of accounts, you are too small to detect the 1 to 5% effects the big studies find. Pick a metric where you expect a large change, such as handling speed, throughput or how fast a new starter gets up to speed.

Where is there no evidence yet for AI agents?

There is no randomised evidence yet that AI agents improve B2B demand generation, B2B sales results, customer onboarding, churn or referrals, so a scale-up that starts there is running an experiment and should measure it as one.

The literature I reviewed found no randomised study of AI on MQLs, cost per lead or sourced pipeline, and none on B2B meetings booked, win rate, cycle length or revenue per rep. There is no independent study of an in-product agent changing customer activation. "AI reduces churn" research is almost all about prediction models.

Almost every rigorous study comes from a large platform: Alibaba, a big cross-border marketplace, Microsoft, Google, a Fortune 500 software company. None tests a B2B scale-up directly. Applying any of it to a 60-person company is an extrapolation, and I'd treat it as one.

Should we use AI agents in sales?

Use agents to prepare sales work and keep a person in front of the buyer: no randomised study shows AI agents improving B2B win rates, and a field experiment found that telling buyers they were speaking to a bot cut purchases by 79.7%.

That experiment predates generative AI, and the undisclosed bots sold as well as good human reps (Luo et al., Marketing Science, 2019). In a B2B relationship you will have to disclose, so that result counts. Let the agent research the account and draft the follow-up, and have a person send it and take the call.

Your buyers are meeting AI agents before they ever meet you, which I wrote about in how AI agents are changing the way buyers find companies.

Why do AI agents stall in smaller companies?

Smaller companies adopt agents quickly and rarely scale them: McKinsey's 2026 survey found 22% of firms under $1bn scaling agents, flat on the year, against 40% of larger firms (McKinsey, 2026), and the usual blockers are data readiness and the absence of a measured result.

That 22% sits beside very high adoption. Mercury's survey of 1,500 US founders found 95% had deployed agents, though Mercury is a bank selling to those founders (Mercury, 2026). In the mid-market, 91% of firms use generative AI and 25% say it is fully integrated into core operations (RSM).

Data readiness is the top barrier to scaling agents for US private companies, named by 72% of the 100 surveyed (Deloitte Private). The other blocker is measurement, and only 13% of SaaS companies track specific KPIs for AI impact (High Alpha).

I saw the same thing in a small survey I ran in April with founders across one VC firm's portfolio. Of the roughly 10 who replied, 90% wanted AI for internal workflow automation, about half had pilots running, and measuring whether AI was working came last on their list. It's a small sample, about AI in general, and it points the same way as the large surveys.

Where should a scale-up start with agentic AI?

A scale-up should start agentic AI with one high-volume workflow where published studies show agents work, keep a person approving anything a customer sees, and take a baseline on a number the board already reviews before building anything.

When I do this with a founder, I start by mapping how the work actually runs, on the company's own data, before automating any of it. Each candidate use case gets a go or no-go with the reasoning written down, and we write a separate list of what not to automate yet. The first deliverable is one agent working on real data. Any strategy document comes after it.

Of the company directors PwC surveyed, 82% rate the information they get linking AI to business performance as fair, poor or missing (PwC, 2026). So take the baseline before the agent goes live.

It is the same discipline as starting from the job the customer is trying to get done.

What should a scale-up not automate with AI yet?

A scale-up should not yet hand AI agents live B2B sales conversations, irreversible actions such as payments and contracts, emotional or high-stakes customer contacts, or executive forecasting, because the published evidence shows no gain or a loss in each.

In an agentic AI experiment on Taobao, ratings fell substantially on chats the AI handled alone (only the abstract is published) (arXiv 2605.14830). A company is liable for what its bot tells a customer, as Air Canada found in court in 2024 (Law360 Canada). And in one classroom study of nearly 300 managers, those who consulted ChatGPT forecast a share price less accurately than those who asked their peers (the effect size comes from secondary reports) (National CIO Review). Leave out any process nobody has mapped yet.

I run my own business on more than 20 agents, and I add them where the evidence says they earn their keep. That is the work we do at Forward Operating Partners: find where to act first, and prove it on a number your board already watches.

Which of your agents could you show a before-and-after number for today?