Insight · AI Operations
66% Productivity Gains. 85% Plan to Customize Agents. The Skills Gap Is the Real Barrier.
Deloitte and ServiceNow's 2026 Workflow Automation Outlook reports 66% productivity gains from AI and names the skills gap, not cost, as the biggest integration barrier. The gap is not a hiring problem.
The AI skills gap is the distance between what employees, principals, and operations leads would need to know to design, deploy, and operate AI agents inside their actual workflows, and what they currently know. It shows up as failed pilots, underused subscriptions, and agent builds that never quite reach production.
Deloitte and ServiceNow's 2026 Workflow Automation Outlook is two data points stacked in tension. 66% of organizations report measurable productivity gains from AI. 85% expect to customize their own agents over the next 18 months. The third data point sits in the middle and explains both: the biggest barrier to AI integration is not cost, not capability, not vendor lock-in. It is the AI skills gap.
What "skills gap" actually means in 2026
The phrase has been used so often it has lost edge. In the Deloitte / ServiceNow framing, it points at something specific.
The gap is not "people do not know what ChatGPT is." Most people in a small firm have used ChatGPT for a year. The gap is downstream of that. The gap is what you need to know to take a workflow that currently lives in three systems and design an agent that runs it end to end.
Concretely. You need to know enough about your firm's actual data layer to know what the agent can read from. You need to know enough about the failure modes of agents to know what guardrails to put around it. You need to know enough about how the team works to know which decisions to keep with the human and which to route to the model. You need to know enough about cost, governance, and the kill switch to know how to operate the agent once it ships.
None of this is taught in an AI workshop. All of it is learned by building something real, with someone who has built something real, against the firm's actual operations.
Why hiring does not solve it
The intuitive response to a skills gap is to hire for it. Most small firms cannot, and the ones that can usually should not.
The market for the skill set in question, an operator who can map workflows, design agent governance, and ship a working build, is currently priced at the upper end of mid-career consulting. Hiring full-time is expensive. The work is also lumpy. The first build takes weeks. The maintenance takes hours. A full-time hire is overbuilt for the load.
The deeper problem is that the skill set transfers poorly into a firm by hiring. A new senior person inherits a data layer they did not consolidate, a team they did not train, and a set of workflows they did not design. The first six months are catch-up. The build comes after.
The Deloitte data implies, and the field confirms, that the firms closing the gap are doing it through engagements that bring the skill set into the firm long enough to make the build, transfer the operating model, and then leave the firm able to extend it on its own.
Why training alone does not solve it either
The next intuitive move is to train the existing team. Run an AI workshop. Send the operations lead to a course. Buy a curriculum.
This produces enthusiastic team members who now know more about AI in general and still do not know what to do with their specific workflows. The course taught the principles. The firm needs the build.
There is a sequencing issue. Training works when the learner has a real problem to apply it to, a real artifact to learn against, and someone to ask when the artifact misbehaves. Course-style training delivers the principles in the wrong order, with no artifact, and no one to ask.
The pattern that works is the inverse. Build a small first agent against a real workflow, side by side with someone who has built agents before. The principal and the operations lead participate in the build. They see the decisions. They see the failure modes. By the end of the engagement, the firm has the agent and the skill at the same time.
This is what the Deloitte / ServiceNow data is gesturing at when it says 85% plan to customize agents. The customization is the skill transfer. Firms cannot get there from a course.
The Radiant Work read on the 85% customization number
The 85% figure deserves a closer look. Most firms in the survey are not yet customizing agents. They expect to in the next 18 months.
What "customization" means in practice will vary. For some, it is configuring an off-the-shelf agent (a vertical AI product) for the firm's specific workflows. For others, it is building a custom agent against the firm's source of truth, with the firm's voice files, the firm's playbook, and the firm's escalation rules.
The first kind of customization is cheaper but capped. The vertical product can be tuned, not redesigned. The output is constrained by what the vendor chose to expose.
The second kind is the one the Deloitte data suggests captures more value. A custom agent built against the firm's actual data layer can do things no vertical product can. It can also be governed, monitored, and rolled back by the firm's own discipline, not by the vendor's.
For most service businesses in the 5 to 50 person band, the answer is some of both. Vertical products for standardized tasks. Custom agents for the workflows that touch the firm's actual differentiated work.
How the skills gap looks inside a service firm right now
In practical terms, the gap shows up in three places.
Discovery. The principal can imagine what an AI-enhanced discovery process would look like. The principal cannot, on their own, design the data schema, write the agent's instructions, build the review surface, and handle the failure modes.
Deployment. The firm has bought an AI tool. The tool is technically deployed. Two of the team use it consistently. Four do not, partly because the workflow around it is unclear, partly because the data the tool runs on is inconsistent.
Operation. An agent the firm builds shipped two months ago. It worked at first. It is drifting. Nobody on the team knows whether the drift is the agent, the data, the prompt, or the underlying model changing under it.
Each of these is a skill issue. None of them is solved by adding another tool.
What the Radiant Work engagement actually transfers
The Radiant Work operations audit and the sprints that follow are designed around the skill-transfer problem. Three things happen during a typical engagement.
One. The firm gets a working build. An agent or a workflow that does specific work the firm needed done. The build is the proof.
Two. The principal and operations lead participate in the build. They see the design decisions, the failure modes, the trade-offs. By the end, they could not have built this from scratch, but they can extend it, debug it, and operate it.
Three. The firm gets a documented operating model: the source of truth schema, the agent log, the policy, the rollback path. The artifacts that make the build hold.
The combination is the skills transfer the Deloitte data says most firms need. See how we work for the methodology.
What to do next
The 66% productivity gain number is the carrot. The skills gap number is the gating constraint. A small firm closes the gap not with a hire and not with a course, but with one engagement that ships a real build and transfers the operating model at the same time.
If you want to know what that engagement would look like for your specific firm, schedule a conversation.
Frequently asked questions
What is the AI skills gap?
The distance between what a firm would need to know to design, deploy, and operate AI agents inside its workflows, and what its current team knows. It is the biggest reported barrier to AI integration in Deloitte and ServiceNow's 2026 Workflow Automation Outlook.
Why does the skills gap matter more than the cost gap?
Tools are cheap. Subscriptions are cheap. The skill to translate workflows into agent designs that hold up in production is expensive and rare. Most failed pilots fail on skill, not budget.
Can a small firm close the skills gap by hiring?
Rarely the right move. The market for the skill set is priced at upper-mid-career consulting, the work is lumpy, and a new hire inherits a data layer they did not build. Engagements that transfer the skill while shipping a build usually beat full-time hires for firms under 50 people.
Can the skills gap be closed by training alone?
No. Courses teach principles in the wrong order, with no real artifact, and no one to ask. The pattern that works is building a small first agent side by side with someone who has built them before. The principal and ops lead learn by participating.
What is the difference between configuring a vertical AI product and building a custom agent?
A vertical product can be tuned but not redesigned. A custom agent runs against the firm's own source of truth, with the firm's voice files and rules, and is governed by the firm's discipline. Most firms need some of both.
The Work Behind the Work
The skills gap does not close with a hire or a course. It closes with a build you keep.
Take the first step toward a business that runs with clarity and momentum.