Enterprise AI

Company-Wide AI Training Is Done. Why Haven't Workflows Changed?

McKinsey's 2025 survey: 88% of organizations use AI regularly, yet most haven't embedded it in workflows. After AI training, redesign one real workflow first.

Baize AIOctober 23, 2026

A completion certificate stacked in front of an orange gate, while ticket and release paths bypass the gate on the old track

The easiest way to report company AI training as done: everyone completes the course, certificates are issued, and people can open the tools. The next day, tickets, reviews, and releases follow the same path they followed before training. The course is finished. The workflow is unchanged.

McKinsey’s The state of AI in 2025 (published November 2025; surveyed June–July; n=1,993) found that 88% of organizations “regularly use AI” in at least one function. Most have not yet embedded AI deeply into workflows and processes, so they do not capture enterprise-level gains. In the same survey, the high performers that attribute meaningful EBIT to AI were nearly 2.8 times as likely as others to fundamentally redesign workflows when deploying AI. Training coverage and workflow redesign are two separate metrics.

Left: all-staff training certificates and unlocked tools. Right: tickets, reviews, and releases still running on the old SOP conveyor — Figure: All Staff Trained | Same Old SOP

BCG press release, June 3, 2026 (11,749 people across 14 markets): among white-collar workers who use AI regularly, 66% say they get “little or no guidance” on what to do with the time they save. A clear strategy lifts measurable business impact by about 25 points. Better tools without strategy and redesign move it by about 5 points. Companies that redesign workflows end to end report measurable business improvement 24 percentage points higher.

Deloitte’s State of AI in the Enterprise 2026 (August–September 2025; 3,235 leaders in 24 countries): 37% still use AI at a surface level, with “little or no change to existing processes.” The top talent strategy is “educating the broader workforce to raise AI fluency” (53%). “Restructuring or rethinking the organization around new AI-driven patterns” is at 30%. Companies run far more training than role and process changes.

Completion certificates don’t replace exit checks

The UK Government Digital Service (GDS) ran a cross-government Microsoft 365 Copilot experiment with about 20,000 people (report published 2025-06-02). Participants got licenses and training. They self-reported saving 26 minutes a day on average, and 17% saw no savings. The report states that it could not identify where the saved time went. It describes training as necessary. It does not show whether tickets, approvals, or releases changed.

Accenture is rolling out Copilot to about 743,000 people (Microsoft Source, 2026-04-27). CIO Tony Leraris said the value does not come from simply turning it on. The company tracks real usage in Outlook, Teams, and Word and builds drafting and its existing messaging standards into daily work. The figure that 97% report routine tasks running 15 times faster is company self-reported. Treat it as unaudited.

MIT Project NANDA calls the gap “High Adoption, Low Transformation”: ChatGPT and Copilot raise individual productivity, not P&L. Do not restate the report’s pilot failure rate as a “training failure rate.”

China’s policy targets deployment in real use cases

This section covers China’s regulatory context. Article 8 of China’s Interim Measures for the Management of Generative AI Services requires “necessary training” for data annotation staff. It does not require company-wide ChatGPT-style courses. We found no official penalty for failing to run company-wide AI training. In August 2025, China’s State Council named the “last mile” obstacles to deploying AI in real applications. In August 2026, the Ministry of Industry and Information Technology (MIIT) defined the work of AI application service providers as consulting and planning, delivery and implementation, operations management, and security governance. Both documents ask for AI deployed into operations through consulting and implementation work.

Buying accounts for everyone is still only the entry point; see Why Companies Shouldn’t Just Buy Individual AI Accounts. One person using Codex well does not mean Codex is part of the engineering process; see What Does It Take to Make Codex Part of Your Team’s Dev Workflow?. To write requirements, implementation, testing, and review into one process, see How Should a Small Team Design Its First AI Coding Workflow?.

How to use this: three changes after training

Counts as trained: People can open the tools and know which data must not be pasted into consumer AI apps.
Does not yet count as deployed: Ticket fields, review checklists, release evidence, and client delivery steps are the same as before training.
Must change before you scale: A table of what AI does on its own and what needs human approval; an SOP for one ticket from intake to close; a rewrite of one real workflow from this week, not a demo environment.

Your task: pick one real ticket path and change it until it runs.
When done, show me: old SOP vs. new SOP side by side, the exit checks, and how the same task is scored two weeks later.
Do not: hand in only attendance sheets and screenshots. Do not treat a demo environment as the production process.

Company AI training: FAQ

Everyone finished company AI training. Why do tickets, reviews, and releases look the same?

Completion shows that people can open the tools. It does not show that the SOP changed. The missing piece is consulting work that changes the ticket path. Another certificate does not supply it.

Training is done but workflows haven’t changed. Will another round of courses help?

Change one real path first. BCG reported in June 2026 that better tools, without strategy and redesign, barely move measurable impact.

Company-wide AI training or rewriting the SOP into tickets: which one counts as AI implementation?

The second. Knowing how to use the tools is the entry point. It counts as implemented when the approval boundary table is signed off and the same task is scored against the same exit checks two weeks later.