Alright, the AI "novelty phase" is over. Call it.

Time of death: Now. (Maybe even a few days ago.)

Everyone has a “copilot,” and almost everyone’s disappointed by how little changed. Across orgs, the same pattern shows up: adoption is easy but impact is hard.

Survey after survey (like our recent Big AI at Work Study) now estimates that only a scanty slice of enterprises have AI truly embedded in day-to-day workflows at scale (on the order of five percent) and when you ask why, the answers aren’t “models” or “legal.”

The blocker is simpler: most tools don’t learn, don’t remember, and don’t fit the work.

Meanwhile, employees are sneakily crossing the gap without permission.

Even where companies haven’t bought official LLM seats, workers use personal tools every day.

In one large sample, staff at over ninety percent of companies reported regular use of consumer AI, while only about forty percent of companies had purchased enterprise subscriptions. That “shadow AI” economy explains why AI feels real at the edge even as official programs stall.

When AI Agent Actually Work

When deployments do break through, the timelines and tactics sure are telling.

Mid-market teams can move from pilot to production in roughly a quarter; and enterprises often take three. And the efforts that survive aren’t heroic internal builds — they’re “buy-partner-customize” motions, where pilots built with strategic partners are about twice as likely to make it to full deployment and see higher end-user adoption. (Make sure to pay your Customer Success team well, folks!)

What about the ROI?

The ROI also isn’t where most budgets point.

Executives tend to over-weigh sales and marketing because outcomes are easy to display, but some of the sharpest paybacks are showing up in back-office work: eliminating $2M to $10M annually in BPO contracts, cutting agency spend around thirty percent, and shaving seven figures off outsourced risk checks—often without changing headcount.

Architecturally, the winners are converging on the same shape.

Agents live inside the system of record, not beside it; they’re observable and auditable; and they operate as a small ensemble across a process rather than as one clever bot. Leaders are normalizing multi-agent deployments—ten or more purpose-built agents across functions isn’t unusual—and they start where exceptions rule the day: finance, supply chain, case and entitlement ops.

The Case of Sweep in These Findings

If you’re Sweep (that's us) that’s our argument: production value appears when AI is boring in the best way — embedded in Salesforce metadata, explainable end-to-end, governed like any other service, and designed to learn from operational experience without turning into a black box.

Buyers are already evolving their evaluation rubrics toward those criteria: context beyond raw data, orchestration across tools and systems, and controls that include access governance, observability, outcome evaluation, and escalation.

How to measure it (the only scoreboard that matters)

Don’t count prompts for Pete's sake. Count outcomes that touch the business: time-to-rep on routed leads, misroute rate, SLA adherence in case ops, rework avoided in approvals, and external spend you can shut off (BPO, agencies, outsourced checks). These are the lines that showed real movement in the field (e.g., faster qualification, measurable retention lift on follow-ups, and hard-dollar back-office savings) and they’re exactly where embedded agents, not bolt-ons, tend to shine their brightest.

A quick vignette (change the names and ship)

A revenue ops team inherits a classic mess: leads stalling in queues, handoffs slipping, and execs convinced “AI” will fix it.

Instead of another sidebar copilot, they embed a Sweep agent inside the routing flow itself. In week one, it starts tracing decisions — every rule, every queue, every exception — surfacing the three patterns behind 80% of misroutes.

By week two, the admin toggles safe fixes behind RBAC with a rollback switch; by week three, time-to-rep is down, escalations are cleaner, and no one had to retrain the org.

Procurement notices they can pause a chunk of external triage spend next quarter.

The win sticks because it’s observable, auditable, and lives where the work lives—not in a demo tab. (This is also why partner-led pilots, with real operational fit, graduate to production more than twice as often.)

The bottom line is...

The next phase is going to be about systems that fit your workflows and get better over time.

If your vendor can’t show memory, governance, and visible outcomes inside your core processes, it won’t make it past the pilot wall.

And if you start with a back-office flow you can measure, the business will notice — because those kind of savings can't hide.