For two years the AI question was whether it works. BCG's 2026 Applied AI Index, published in September, suggests that question is settled: almost half of companies now generate value from AI, and 41% are in the scaling stage.
The more interesting number is the gap. BCG reports that its most mature 7.5% of companies, which it calls "future-built," generate roughly 2.4 times the top-line growth of laggards, and they do it through enterprise-wide programs.
The Experiment Phase Is Over
The numbers reward AI tied to real workflows, with owners, targets, and measurement. If your AI work is still "a few people trying ChatGPT," you're in the group the leaders are pulling away from.
What the Leaders Have in Common
BCG's finding is that the leaders run enterprise-wide programs. Translated to a smaller company, here's our read of what that looks like in practice:
- One named owner per automated workflow, accountable for the result
- A baseline measured before automating (hours, cost, error rate), so ROI is provable
- A few workflows done deeply instead of many tools adopted shallowly
- AI built into the systems people already use, with no separate tab to remember
- Review checkpoints wherever a mistake is expensive
Leaders treat AI as an operating change with owners and metrics, and budget size matters less than that discipline.
Agents Got Good, and Still Miss 1 in 3
Stanford's 2026 AI Index reports that AI agents jumped from roughly 12% to about 66% task success on OSWorld, a benchmark of real computer tasks across operating systems. That's a remarkable leap. It also means agents still fail roughly one in three attempts on structured benchmarks.
Read it both ways. Agents are now capable enough to own real workflow steps, and not reliable enough to run unsupervised on anything expensive to get wrong. Approval gates, logging, and test sets matter more as agents improve.
The Small-Business Version
You don't need an enterprise program. You need the same discipline at a smaller scale: pick the two or three workflows where hours leak the most, name an owner, measure the baseline, automate with approval gates, and review the numbers every month.
A 90-Day Plan
Days 1 to 30: List your workflows, measure hours and error rates, and pick the top three by time lost.
Days 31 to 60: Automate the first one, with human approval on any consequential step.
Days 61 to 90: Compare results to the baseline, tune, and start the second workflow.
Where to Start
Start with the baseline. If you can't say how many hours a workflow costs you today, you can't prove what automating it saved, and that's the difference between AI that gets renewed in the budget and AI that quietly gets cut.
Want your AI work to show up in the numbers?
We'll help you pick the right first workflows, set the baseline, and build with the controls that make results repeatable.
Book a Free CallSources: BCG, The Formula for Agentic AI Value (September 2026); Stanford HAI, 2026 AI Index Report.