3 min read
Most businesses have tried AI by now. Far fewer can say what it returned. The difference is rarely the model — it is whether anyone picked a specific job, measured it before and after, and wrote down who owns it.
The numbers, and what they measure
37%of respondents attribute at least some EBIT impact to their organization’s use of AI — roughly the same share as a year earlier — even though nearly nine in ten report using AI in at least one business function. McKinsey, The state of AI in 2026, 1,719 respondents, surveyed May–June 2026.
The small-business picture is earlier still. The US Census Bureau’s Business Trends and Outlook Survey found overall AI use across US businesses hovering between 17% and 20% from December 2025 to May 2026, with 37% of firms of 250 or more employees using it — and no significant change among firms with fewer than 20.
The two surveys measure different things. McKinsey surveys managers and executives at organizations worldwide about use in any function; the Census Bureau asks a representative sample of US businesses, and broadened its question in November 2025. What they agree on is the shape: use is widespread in large companies, patchy in small ones, and measured return lags use everywhere.
Why the gap exists
Because the order is usually backwards. A tool gets bought, then someone looks for a problem it solves. A pilot runs without a baseline, so nobody can say whether it worked. The tool is still on the card; the problem it was meant to fix is still there.
A roadmap that starts small on purpose
- Map how work actually moves. Not the org chart — the real path a lead, an order or a booking takes, including the inbox nobody owns and the spreadsheet only one person understands. The leaks show up on the map.
- Pick one workflow. Score candidates on three things: how often it happens, how much it costs when it goes wrong, and how easy it is to undo. High volume, real pain, reversible: start there.
- Baseline it before you touch it. Response time, hours per week, error rate, whatever the job is measured by. Two weeks of honest numbers is enough. Without them, nothing afterwards can be called a result.
- Build the simplest thing that works. Often that is a plain automation with one language-model step in it — reading, sorting or drafting — and a person approving anything that leaves the building. It is rarely an agent.
- Write the policy before client data goes anywhere. Which tools, which data, who reviews what. A few pages is enough. See AI governance.
- Hand it over documented, then measure at 30, 60 and 90 days. If it paid for itself, pick the next workflow. If it didn’t, you spent one workflow finding out — which is far cheaper than a platform finding out.
The first workflow we build for almost every business is lead routing: every inquiry lands in one place, assigned to one person, acknowledged within minutes, followed up automatically. It is unglamorous, easy to measure, and it usually pays for the next one.
Questions to ask before you buy anything
- Which system does this read from, and which does it write to?
- What does it cost per month at our volume, not at the demo’s?
- What happens when it is wrong, and who finds out?
- Can we export everything — prompts, workflows, data — if we leave?
Where we say no
- Fine-tuning a model for a ten-person company.
- An agent that “runs the business.”
- Anything that touches client data before there is a written policy on where it goes.
Sources: McKinsey, The state of AI in 2026: On the road to ROI, August 2026 · US Census Bureau, Large Firms With at Least 20 Employees Biggest AI Users, May 2026 · Anthropic, Building effective agents, December 2024.
Definitions: AI integration, workflow automation, human-in-the-loop, AI governance. Found First’s AI implementation work starts with a Readiness Assessment — see AI implementation and automation.