AI agents for a small team: what works, what doesn't
There is a measurable gap between who is using AI at work and who isn't, and it runs along company size. We sell an agent platform, so treat this accordingly — but the useful version of this post is the one that tells you what still doesn't work, because a six-person company cannot afford a failed rollout and deserves better than a feature list.
The gap is real, and it's about size
The US Census Bureau's Business Trends and Outlook Survey, published 26 May 2026 and covering December 2025 to May 2026, found overall business AI use "hovered between 17% and 20%". Split by size, though: 37% of firms with 250+ employees reported using AI, 32% of firms with 100–249 — and under 20% of firms with fewer than 20 people. The line that matters most is the trend: AI use increased among firms with at least 20 employees but didn't change significantly among firms with fewer than 20. The gap isn't closing on its own.
One caveat almost nobody reports, which you should know before trusting any chart on this: in November 2025 the Census changed the question from AI use "in producing goods or services" to use "in any business function." Numbers either side of that change are not comparable, so a trend line drawn across it is partly an artifact of the wording.
About that "95% of AI pilots fail" statistic
You have seen it, and it is almost always quoted wrong. It comes from The GenAI Divide: State of AI in Business 2025, an MIT-affiliated Project NANDA working paper — self-described as preliminary, not peer-reviewed. The 95% describes organisations getting no measurable return, not pilots failing. The paper's own funnel: 60% of organisations investigated AI tools, 20% reached a pilot, 5% reached production — which makes roughly a quarter of the organisations that actually ran a pilot clear the bar, not one in twenty.
Two more things worth knowing before you weigh it. The bar was demanding by design: deployment beyond pilot, measurable KPIs, and ROI measured six months after the pilot — a short window in which to book a return. And the sample was 52 interviews and 153 leaders; the widely-repeated "150 interviews and 350 employees" came from press coverage and was copied outward from there.
The genuinely useful finding is the one nobody quotes. Buried in the same paper, generic off-the-shelf chatbots moved from pilot to implementation at around 83%. The simple tools worked. The bespoke enterprise builds are what produced the headline. That is close to the opposite of the lesson the statistic is usually deployed to teach, and it is a much better guide for a small company: buy the boring thing and use it, rather than commissioning something impressive.
The obvious first use case is the one that backfires
If you asked a room of owners where to start, most would say a customer-facing support bot. The evidence says start almost anywhere else. Qualtrics surveyed more than 20,000 consumers across 14 countries in Q3 2025 and found nearly one in five who had used AI for customer service got no benefit from it — "a failure rate almost four times higher than for AI use in general". Consumers rated AI customer service among the worst applications tested for convenience, time saved and usefulness; only one scored lower.
The reading isn't that customer-facing AI can't work — it is that it is the hardest place to start and the most expensive place to be wrong, because the person who experiences the failure is your customer. Internal work fails privately. Start where a bad day costs you an hour instead of an account.
What works reliably for a small team today
- Repetitive written work with a known shape. Triaging an inbox, drafting the reply you write forty times a month, summarising a long thread into a decision. This is the strongest category and it is not close.
- Being reachable when you aren't. Answering the phone after hours, taking a message, replying to a WhatsApp enquiry at 9pm. The value here is time-shifted rather than clever.
- Recurring work on a schedule. The Monday report, the price check, the follow-up nobody remembers. Reliable precisely because the task is the same every time.
- Reading a lot and reporting a little. Research, competitor checks, going through documents you'd never get around to.
What still doesn't
- Anything needing judgement it can't check. A model can be confidently wrong, and it will not know. Work with a verifiable output — a number that must reconcile, a test that must pass — is far safer than work whose only reviewer is you, later.
- Very long tasks, without losing detail. When a task outgrows the model's context, older parts get summarised and that summarisation is lossy. It is a design constraint, not a bug awaiting a fix.
- Reaching anything you didn't connect. An agent has no ambient access to your business. It reads the public web, and everything else — your inbox, your calendar, your records — only after you connect it.
- One-shot big builds. "Build me the whole system" produces a real first pass and then needs steering. Anyone demonstrating one perfect attempt is demonstrating a demo.
The longer limitations list goes further, including the gates we deliberately keep shut.
What actually decides whether it works for you
Not the model, and not the platform. Two things: picking a workflow that is genuinely repetitive rather than genuinely impressive, and deciding what must never happen without a person saying yes. Small teams that start with the boring recurring job — the one that eats a morning a week — get a result in days. Teams that start with their hardest, most customer-visible problem are the ones who end up quoting the failure statistics.
The cheapest way to find out which you are is one workflow for one week. The free plan needs no card, and if you'd rather not pick and configure it alone, we'll do the first one with you — including telling you if we think agents are the wrong tool for it.
See it for yourself — free, $2 of credits once you confirm your email, no card.
Start free