The AI readiness gap is your people, not your processes
A Deloitte survey of 501 US executives — every one of them already piloting AI agents — found 74% expect to redesign nearly half their processes around agents, while only 5% call those processes highly prepared. Workforce readiness scored lowest of all. What that means for the order you do things in.
Most organisations planning to hand work to AI agents have not funded the part that decides whether it lands: the people who will have to work alongside them. Readiness surveys keep finding the same shape — ambitious plans, unprepared processes, and the workforce scoring lowest of every category measured.
What the survey actually found
Deloitte's 2026 survey AI Agents are Only the Beginning: Deloitte Survey Examines the AI Readiness Gap (published 12 August 2026) put four numbers next to each other:
- 74% expect nearly half their business processes to be redesigned or rebuilt around AI agents.
- 15% have scaled orchestrated, multi-agent adoption.
- Only 5% say those processes are highly prepared.
- Workforce readiness was the least-prepared area of all, at 25% — and about half said they are not adequately investing in AI-related workforce transformation.
Read the caveats before you quote it, because they change how much weight it carries:
- The sample is 501 senior-manager-to-C-suite respondents in the United States, fielded April to June 2026. It is not European data and should never be presented as such.
- Deloitte sells agentic-AI advisory. This is a consultancy publishing the size of a gap it is paid to close. Worth naming plainly.
- Every respondent was already piloting agentic AI. That self-selection makes the sample keener than the market, which makes the 5% *stronger*, not weaker — but it also means it cannot be read as "5% of all companies."
Why "the people are the least-ready part" is the line that matters
The interesting thing is not that readiness is low. It is *which* readiness is lowest. Process design, tooling and data all scored ahead of the workforce. So the plan is machines doing a large share of the work, and the line nobody funded is the one that decides whether any of it is used.
This is the same failure mode that produced a decade of underused software, arriving one layer up. A tool nobody adopts is an invoice. A redesigned process nobody adopts is an invoice plus disruption.
How do you get employees to actually use the AI tools you bought?
You embed the tool in work they already do, using their own real tasks, and you make a named person accountable for the workflow. What does not work is mandating usage or booking a generic course and counting attendance.
Four things that reliably move the number:
- Role-based sessions on real workflows. Not "here is the tool" — here is your Tuesday reconciliation, done with it. The transfer problem is the whole problem.
- Champions drawn from existing staff. A colleague two desks away who already does it well beats an external trainer who leaves on Friday. Find the people who quietly figured it out and give them the time and standing to spread it.
- Remove the old path. As long as the manual route still works and is faster for someone who has not learned the new one, they will take it. This is a management decision, not a training one.
- Measure usage, not attendance. Weekly active users on real work is the number. Completed training is not adoption and never was.
The sequencing this implies
If the workforce is the least-prepared layer, then automation-first is the wrong order. Our own rule is blunt about it: no automation without adoption first, and no workshops without an audit first. You establish what is actually owned and actually used, you get people using it, and only then do you automate around them. Reversed, you get an expensive process redesign resting on a layer that was never prepared.
What this is not
It is not an argument for buying AI training. Training is a delivery mechanism, not an outcome — and a completed course sitting in an LMS proves nothing about Tuesday. The outcome worth paying for is people who use the tools you already own, on their real work, with the hours it gives back visible in a number.
It is also not another mandate to hand down. The end users we work with are not resisting AI out of principle. They are resisting extra homework attached to a job that is already full. Relief is the pitch that works, because it is the one that is true.
The plan is machines doing the work. The line nobody funded is the one that decides whether any of it lands.