Three reports, one shift
McKinsey's 2025 outlook treated agentic AI separately from generative AI. A generative model provides an output for a person to use; an agent can carry out an action. Teams need to decide who owns those actions and how to review them.
Bain's 2025 report focused on the economic consequences. Hyperscaler capex for AI infrastructure crossed half a trillion dollars in commitments for the 2025-26 cycle. Software industry margins came under pressure as the cost of inference shifted from a marginal expense to a structural one. The companies positioned to win were either the infrastructure providers themselves or the operators with workflows efficient enough to make agent deployment economical.
Deloitte's 2025 outlook surfaced the second-order question almost nobody was asking: where does accountability live when a software agent makes a decision that's wrong? Their answer (and it's the right one) is that it has to live somewhere it always lived: with a human owner of the workflow the agent runs on. Companies that haven't named that owner can't deploy agents safely.
What an agent actually requires
The 2025 reports converge on a deceptively simple list. To deploy an agent against a workflow, you need:
- A workflow that's been documented well enough that a human could write the runbook from scratch.
- Inputs and outputs that are machine-addressable: APIs, structured data, defined schemas.
- Decision logic with explicit thresholds, not 'use judgment,' but 'if X, then Y, else escalate.'
- An owner who can read the agent's logs and intervene before drift becomes loss.
- An evaluation regimen: how often does someone check that the agent is still doing the work better than the human alternative.
"The agent is the easy part. The five preconditions that have to be true for an agent to be safe and economical: that's the consulting engagement."
Where the gap shows up financially
Bain's 2025 numbers tell a sharp story. Agent pilots are running everywhere; agents in production with measurable ROI are concentrated in a small number of operators: Bain's estimate is roughly 15%.
The financial gap between the two cohorts widens fast. The 15% are absorbing the cost of inference because they've collapsed the human work it replaces. The other 85% are paying for the inference and still paying for the human work, because the workflow wasn't redesigned to actually hand the work off.
This is the same operating-model gap we wrote about in 2021, 2022, 2023, and 2024: just with a more expensive failure mode.