Businesses Are Planning AI Agents Faster Than They Can Run Them
Synergised Consulting Ltd
Last Updated: 12 November 2026
AI agent programmes stall because businesses approve the plan before they build the skill to run it. Two 2026 surveys, one of 1,224 professionals and one of 501 US leaders already piloting agents, measure the same gap from opposite ends: employees can use AI tools in general but cannot apply them to their own work, and organisations keep expanding agent deployments without ever reaching scaled, coordinated use. The bottleneck is capability, and capability is built one measured workflow at a time.
What the New Research Actually Measures
The new evidence is two large surveys arriving at the same conclusion from opposite directions. Acorn's 2026 State of Learning for AI Fluency Report surveyed 1,224 professionals across the C-suite, directors, managers and individual contributors in the United States, Canada and Australia in April 2026, and found that 58% of companies report employees who are proficient with AI tools in general but struggle to apply them meaningfully to the specific requirements of their role [1]. The same report found that 75% of employees say AI has made them less than 25% more efficient [1].
Deloitte's agentic AI survey, published in August 2026, measures the organisational end of the same gap. Of 501 senior leaders at US companies already piloting agents, 42% have tested or deployed agents and 43% have deployed them across more than one function, but only 15% have scaled, orchestrated, multi-agent adoption in place [2]. Read together, the two studies describe a capability ceiling: plenty of organisations can switch tools on, and very few have built the working practices that let agents carry real workflows.
Why both numbers come from committed adopters
Neither survey measures passive interest. The Acorn sample comes from organisations investing in AI training, and Deloitte restricted its sample to companies that had already cleared the first hurdle of funding and running a pilot. That makes the figures stronger, not weaker: the drop-off from 43% expanding to 15% scaled describes what happens after budget and executive attention have already been committed, which is exactly the position a new owner is in during the first hundred days.
Why Plans Outrun Capability
A plan can be approved in a meeting, but fluency is a habit, and the two move at different speeds. Adoption is easy to count: seats licensed, tools rolled out, a pilot launched. Outcome is hard to count, so it usually goes unmeasured, and the Acorn figure of 75% of employees reporting less than a 25% efficiency gain is what a tooling-first rollout looks like from the inside [1]. The infrastructure for measuring anything else does not exist in most organisations yet: 34% of companies have not defined AI competencies at the role level, and 47% have not included AI capability in formal performance reviews [1]. With no defined target, there is nothing to towards and nothing to measure against.
The perception gap sits on top of the measurement gap. Among executives, 77% believe their managers are prepared to guide AI skills development, while 91% of employees say their managers lack that preparation [1]. The further a person sits from the work, the more confident they tend to be that the rollout is working, which is why a capability problem is so often reported upward as a success.
Two teams, same budget, different results
Consider two illustrative teams of the same size given the same tooling budget. The first rolls the tool out to everyone, runs a training afternoon on prompting, and reports adoption at the next board meeting. The second picks one workflow, such as invoice coding, writes down how long it takes and how often it goes wrong today, and has one person rebuild that single workflow with the tool until the baseline moves. Six months later the rollout-first team has an adoption statistic while the workflow-first team has a measured result, plus a person who now knows how to do the same for the next workflow. That person, not the licence, is the asset.
What General Use Is, and What It Is Not
General proficiency means getting usable material out of a model. Role competence means changing how a specific piece of work actually gets done: knowing which steps of the job a model should handle, which steps stay with a person, and where the output gets checked before it reaches a customer. The 58% finding sits exactly in that distance [1], and closing it is a management task rather than a training one, because only someone who understands the role can define what good AI use in that role looks like.
This is why generic courses rarely move an efficiency number. Prompting can be taught in an afternoon, but deciding what the model should be allowed to do inside a real process requires the person who owns that process to be in the room. In our own discovery workshops, the list of what is worth automating almost never arrives with the owner; it gets built in the conversation, walking a normal week task by task, and the repetitive screen-based work separates itself from the tasks that genuinely need judgement. That conversation, not the software, is the first hundred days' most useful AI activity.
The first candidate workflow
The right first workflow is high volume, rule bounded and reversible: quoting, invoice coding, inbox triage or status reporting all qualify. The test is simple. If the agent gets it wrong, can a person catch it before it reaches a customer, a payment or a regulator? If yes, it is a good first candidate, and completing it end to end, even clumsily, teaches the team more than any plan would. If no, it belongs in a later phase with tighter controls, because a dramatic first demo on the most tangled workflow is how a promising programme acquires a reputation it never recovers from.
Sources
- [1] Acorn, "The 2026 State of Learning for AI Fluency Report", survey of 1,224 professionals across the US, Canada and Australia, conducted April 2026, released May 2026:
- [2] Deloitte, "AI Agents are Only the Beginning: Deloitte Survey Examines the AI Readiness Gap and Reveals How Enterprises Can Prepare for Agentic Success", press release, 12 August 2026, reporting a survey of 501 US senior leaders fielded April to June 2026: