Companies now run 13 AI agents on average and stand up a new one in two days. That's faster than most construction back offices can vet a single new vendor

The average company now runs 13 AI agents, up from 5 about 15 months ago, and can stand up a new one in under two days. That's the headline from Salesforce's latest Agentic Enterprise Index — and the report's own conclusion is the part that matters for construction: the bottleneck isn't building the agent anymore, it's deciding who's allowed to and what it can touch.

What did Salesforce actually measure?

The index draws on two data sources: aggregated usage data pulled from Salesforce's own Agentforce platform between February 2025 and April 2026, plus a proprietary survey of 4,689 respondents across the US, UK, France, Canada, Australia, Spain, and Italy, fielded in May 2026. That combination makes it one of the larger deployment-plus-survey studies of enterprise AI agents published so far this year.

The headline numbers: the average number of active agents per organization rose from 5 to 13, and average time to deploy a new agent fell 53%, down to about 1.9 days. Salesforce also found a split in how industries build agents — consumer-facing sectors lean into high-volume, narrow, task-specific bots (answer this, route that), while manufacturing, financial services, and healthcare build fewer but more capable agents that chain multiple steps together across systems.

Why does deployment speed matter more than the agent count?

Thirteen agents isn't the number worth sitting with — 1.9 days is. That's roughly the time it takes to get a same-day RFI answered, not to stand up a new piece of software that can read data, write to a system, and act without someone watching every step. Salesforce's own read on this: the constraint has moved from "can we build it" to "where should it work, what data can it touch, and what does human sign-off look like" — governance questions, not technical ones.

That's the finding a construction ops director should flag. Most GCs don't have 13 agents running yet, but the tools that let a non-technical employee wire one together — Copilot Studio, Agentforce, Claude's agent builder, a low-code workflow app — are already sitting on desktops across the industry. The two-day number means the jump from "one person's convenience script" to "several agents touching live project data" can happen faster than IT or ops leadership would notice.

What's the construction parallel?

Salesforce's bifurcation finding maps cleanly onto construction work. A retail chatbot answering "where's my order" is the consumer-facing, high-volume pattern. A submittal-review agent that has to read a spec section, pull the matching cut sheet, check a compliance requirement, and flag a gap is the manufacturing pattern — multi-step, cross-system, judgment-dependent. That's the harder, slower-to-trust category the report says complex and regulated industries are building toward, and it's the category nearly every useful construction agent falls into: RFI triage, change-order drafting, daily-log summarization, schedule-impact flagging.

The second category is exactly where an ungoverned agent does the most damage — because it's touching pricing, spec compliance, or a document that goes to an owner or inspector, not routing a support ticket.

Who's accountable when an agent touches a submittal or a change order?

Right now, probably nobody in writing. Before letting anyone on staff build or buy an agent that reads or writes to a system of record, put three things in place:

• A named owner per agent. Not "the office" — one person accountable for what it does and who checks it.
• A write-access boundary. Can it draft a submittal transmittal for review, or can it actually update the log and notify the sub? Those are different risk levels and should require different approval.
• A review step before anything client- or inspector-facing goes out. Salesforce's 1.9-day deployment number is a feature for the vendor pitching you the tool. It's a warning if your firm doesn't have a faster review process to match it.

The limits here

Salesforce is grading its own platform's customers, so this isn't neutral third-party research, and construction isn't broken out as its own category. The 1.9-day figure is also an average across agent types; a narrow FAQ bot and a multi-system submittal agent don't take the same effort to stand up safely, even if the platform lets you click "deploy" on both in the same afternoon.

None of that changes the practical point: the tooling to build an agent that reads and writes construction documents is now cheap and fast enough that it will show up on a job whether or not ops leadership approved it first. Related read: how three AI agents on one shared task started sabotaging each other — the same ownership gap, at the next scale up.

The takeaway: before your next AI tool pilot, write down who owns it, what it's allowed to touch, and who signs off before it goes external — on paper, before the two-day deployment clock starts, not after.

Running more than one AI tool on your project already? Map what each one can read and write before you add a third.

Construction AI Brief publishes three times a week. Subscribe at constructionaibrief.com.