One job, clean data, a person on standby: what the big agent surveys say small firms should copy

By Jason Sibley

Most advice on AI agents comes from firms with IT departments and a big budget line for change. This latest survey is no different. Still, the habits it describes cost very little and suit a firm of ten as well as a firm of ten thousand. If you're unsure what an agent is for, read what AI agents are good for in a small business first.

What the survey is

On 27 October 2026 Salesforce published a blog post, 4 Takeaways From the State of Agentic AI Report for Small Business. It summarises Salesforce's first State of Agentic AI in the Enterprise report. I've worked from that blog summary, not the full report.

Before using its numbers, bear these points in mind:

• Salesforce sells AI agents (its product is called Agentforce). It has a stake in the answer.
• The figures are self-reported by 2,025 agentic AI decision-makers across 20 countries.
• Only 21% of them run small or medium businesses. Roughly four in five respondents are bigger organisations.

This is one vendor's survey of mostly larger firms that have already bought in. It isn't a verdict on what works for a ten-person shop. I still find it useful because the lessons concern habits rather than budgets. Treat the numbers as hints, not promises.

Lesson 1: pick one tightly bounded job

Salesforce says organisations with agents live in production reached a return on investment in about 8 months on average. That is self-reported, and "return" means whatever each respondent counted as one.

More interesting is what predicted a fast, autonomous rollout. Two things tied for first place, each cited by 36% of respondents: clean, accessible data where the agent takes action, and a tightly bounded use case.

That means a small, boring job with a clear start and finish. Salesforce's examples are routing support tickets or qualifying leads. Nobody starts with "run my customer service". If your last pilot sprawled, why small business AI pilots fail covers the usual reasons.

Lesson 2: clean the data for that job, not everything

Owners hear "clean data" and picture a six-month tidy-up. The survey figures suggest otherwise:

• Only 31% fully unified their data before deploying agents.
• 69% of the rest still reached ROI from incomplete or fragmented data, according to Salesforce.
• 63% hit agent errors tied to stale, missing or wrong data.
• Of those 63%, 79% also had data connectivity issues.

You don't need a perfect system, but bad data causes problems. Tidy the data your one job touches: the price list, the list of services, the enquiry inbox. Leave the rest of the filing cabinet alone.

The connectivity figure matters too. If the AI can't reach the information it needs where the work happens, it will guess. A guess delivered confidently to a customer is worse than no answer.

Lesson 3: keep a person on standby

Only 36% of respondents run agents in production with full autonomy. Most don't. Salesforce reports that 81% saw more human re-entry into workflows as deployments expanded. People got pulled back in as the agents took on more. And 89% say they're rethinking their workforce models to account for agents.

Governance involves a trade-off. Respondents with above-average governance took longer to reach ROI (9.3 months against 7.2), but caught problems proactively 68% of the time against 55%. Read the headline "8 months" loosely. Some firms were slower because they were being careful.

There were regrets too. 43% had to stall or scale back an AI initiative. The top regret, at 39%, was not budgeting for AI fluency training and reskilling upfront. Next, at 38%, was not setting clearer governance before scaling.

Salesforce makes a useful point here. For a small business, it says, governance can be as simple as deciding who reviews what an agent does and where it escalates to a human. For a starting framework, there's a simple AI policy for small teams.

What that looks like in a ten-person firm

This section is my opinion, not the survey's.

Pick one job from a list like this:

• Sorting incoming enquiries into a few categories, so the right person sees them first.
• Drafting a first reply that a human edits before it goes anywhere.
• Spotting quotes that have had no response and drafting the chaser.

None of these jobs sends money, changes a contract or speaks to a customer unchecked.

Then do three things:

• [ ] Name one person who checks the output. A name, not "the team". If the data includes personal details, ask whoever advises you on data protection before you start.
• [ ] Draw the handover line in writing. Mine would be: anything involving a complaint, a refund, a price change or an upset customer goes straight to a human.
• [ ] Tidy only the data that job reads, such as the current price list or the standard answers.

Run it for a month and note what the checker has to fix. If the fixes are small and getting smaller, widen the job a little. If not, you've lost a month, not a quarter. One AI workflow for 30 days sets out how to run that month. To build it with simple automation first, try Zapier and Make first automations. Once you have more than one tool, an AI tools register stops you losing track of what's running.

A one-page agent brief

Before building anything, write the job down on one page. You can ask an AI to draft it with you. Copy this prompt and answer its questions honestly:

I run a small business in the UK. Please help me write a one-page brief for one AI task.

Ask me questions one at a time, then produce a brief with these headings:
1. The job: what the AI does, in one or two plain sentences.
2. The data it uses: which documents, lists or systems it reads, and anything it must not see.
3. What it must never do: for example, promise prices, make refunds, or reply to complaints.
4. Who checks: the named person, and how often they review its work.
5. When it hands over: the exact situations where a human takes over.

Use British English and plain words. Do not invent facts about my business, my customers, my tools or my numbers. If I haven't told you something, write [OWNER TO FILL IN] in that spot instead of guessing. At the end, list every gap you found so I can fill them in.

If you can't fill in "who checks" and "when it hands over", the job isn't ready to hand to an AI yet. Better to find that out on paper.

A quiet word at the end

At My AI Apprentice we run AI sessions for small businesses, including sessions on picking that first sensible job. If that would be useful, you're welcome to get in touch. No pressure either way.