74% of leaders expect AI agents to redesign half their work. 21% are ready.

The report is AI agents are only the beginning: The path to agentic transformation, published August 12, 2026. Deloitte surveyed 501 senior managers and C-suite executives in the US between April and June, every one of them at an organization at least piloting agentic AI, and backed the survey with 20 executive interviews. That sample choice matters: these are the people who already signed the pilot budget, reporting what happened next, not a panel reacting to headlines.

The expectation side is loud. 74% of leaders expect close to half of their business processes to be redesigned around AI agents within four years. 61% expect most agents to run largely autonomously, with humans moved into oversight roles. Nearly two thirds say agentic AI already has them reevaluating their business model.

Then the reality side. 21% say their business processes are prepared for agentic adoption, and inside that number only 5% call themselves highly prepared. 15% have reached scaled, orchestrated multi-agent adoption. 42% are still testing small numbers of agents. As HR Dive put it in its August 20 write-up, only one in five organizations is prepared to move toward the autonomous agents most of them expect to arrive.

Every outlet that covered this report restated the percentages and stopped. CIO Dive's read was that full-scale adoption remains three or four years away. True, and not the useful part.

The useful part is that a 53-point gap between expectation and preparation is not a prediction problem. It is a work backlog. The same survey names exactly what the work is, and puts a percentage on each piece. Read that way, this stops being another adoption stat and becomes the most specific to-do list any of these companies has been handed.

Here is the context none of the coverage supplied, and it is what makes the readiness gap urgent instead of academic.

Buying and launching agents used to be the bottleneck. It is not anymore. Salesforce's 2026 Agentic Enterprise Index, built from aggregated usage across businesses running agents between February 2025 and April 2026, found the average organization went from 5 activated agents to 13 across that window. The time to create and activate an agent fell 53%, to under two days. Actions per account grew at a 31% compound monthly rate.

Put the two studies side by side and the constraint has visibly moved. Agents are multiplying inside companies at consumer-software speed, while the processes they are supposed to run on are ready in one company out of five. When the tooling gets that cheap and that fast, the bottleneck does not disappear. It relocates to the buyer's own operations. OpenAI's Codex study measured the same wall from inside its own customer base: 97.9% agent adoption among its staff, 17.3% among its organizational subscribers. The 42% of Deloitte's sample stuck in testing are stuck for a mundane reason: a pilot that touches an unprepared process has nowhere to go.

Deloitte asked what is actually blocking scale, and three answers came back within five points of each other. 72% cite the lack of a unified, accessible data foundation. 70% struggle with agent trustworthiness and governance. 67% face integration complexity and cost.

Survey write-ups treat numbers like these as atmosphere. They are better read as a project plan with the priorities already voted on by 501 people who tried. And the order matters, because each barrier compounds the next: an agent without governed data produces answers you cannot trust, ungoverned trust decisions stall every approval that follows, and integration priced late is the surprise that kills the budget.

The 72% barrier goes first because it silently caps everything downstream. A human running a process routes around bad data without noticing: they know the price list in the shared drive is stale, so they check with Priya before quoting. An agent does not know about Priya. It reads what it is given and acts on it, at volume.

Fixing this is less grand than "data transformation" sounds. For one process, it means naming the systems of record, killing the duplicate copies, and writing down which field wins when two disagree. Call it days of unglamorous work per process. Nobody needs a two-year platform program for one workflow. The mistake we see is doing it nowhere because doing it everywhere is unaffordable. You do not need an enterprise data foundation to run one agent well. You need the data foundation for that process.

The 70% barrier is usually misread as a policy-writing exercise. Companies have policies. What they do not have is policies an agent can execute: which actions run without sign-off, which queue for a named human, what the agent must log, and who reviews the log.

The distinction is the whole game. A policy PDF constrains nobody at 2am. An approval rule wired into the workflow does. Deloitte's own respondents get this at some level: 61% expect humans to move into oversight roles, and 75% believe human-agent collaboration produces more value than pure automation. Yet fewer than half have defined an operating model for that collaboration. Believing in oversight without building the oversight mechanism is how you end up in the 70%. And an approval table you have never tested is a policy PDF with extra steps, which is the case for running agent evals before an agent gets any authority worth worrying about.

The 67% barrier is the least discussed and the most predictable. An agent that cannot touch your CRM, your ticketing system, and your ERP is a chatbot with opinions. Getting it those connections costs real engineering money, and the cost arrives after the demo impressed everyone, which is the worst possible moment to discover it.

So price it first. Before a pilot starts, list every system the agent must read or write, check what each exposes by API, and get a number on the gaps. If the integration bill exceeds the value of the process, you want that answer before the pilot, not after three months of sunk demo. This is also the honest argument for starting with small, bounded builds: our starter builds run $1,500 to $2,500 fixed precisely because a first agent should be small enough that its integration cost is knowable in advance.

The evidence on what separates the scaled minority from everyone else is consistent across three independent research groups, and none of it says they bought better agents.

McKinsey's State of AI research finds that among the roughly 6% of organizations getting real bottom-line impact from AI, the single strongest differentiator is fundamentally redesigning workflows rather than bolting AI onto existing ones. MIT's Project NANDA, whose GenAI Divide report was covered by The Register, found that partnering with an external builder roughly doubled the success rate versus building internally, with the internal builds stalling because the tools never learned from or adapted to the workflows they landed in. And Deloitte's scaled 15% correlate with exactly the unglamorous traits above: clear operating models, governed data, priced integration.

The pattern in plain terms: the winners did the process work first and treated the agent as the last mile. The losers bought the agent first and hoped the process would sort itself out.

The failure numbers on skipping readiness are not projections anymore. They are measurements.

MIT's analysis of 300 public deployments found 95% of generative AI pilots delivered no measurable P&L impact. Gartner projects over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear value, and inadequate risk controls. The same Gartner note estimates that of the thousands of vendors selling "agentic AI," only about 130 are selling the real thing. The rest are rebranded chatbots and RPA, which means an unprepared buyer is not just likely to fail, but likely to fail on a product that was never going to work.

We wrote up what production failure actually looks like from our own build logs, and the categories rhyme with Deloitte's barriers almost line for line: stale data, missing approval paths, integrations that broke on a vendor's schema change.

Play the optimistic branch honestly. A company that spends a quarter fixing one process's data, writing executable approval rules, and pricing its integrations has built the on-ramp the 74% expectation actually requires. When it then deploys an agent, the Salesforce numbers work in its favor: activation in days, capacity compounding monthly, each additional agent cheaper to justify than the last because the groundwork is shared.

That company also gets something subtler: the redesign dividend. McKinsey's data says the EBIT impact lives in the redesign, not the deployment. Redesigning a process around an agent that works, with people moved to exceptions and oversight, is where the ten-times-output math stops being a slide and starts being a Tuesday. The four-year timeline the executives predicted is realistic for this company. It might even be conservative.

Now the branch most of the 80% are actually on. The agent gets deployed onto an unprepared process, because deploying is now the easy part. It reads stale data confidently. Nobody defined what it may do alone, so either it is throttled into uselessness by blanket approval requirements or it acts and somebody spends their week auditing it. The integration bill lands mid-pilot. The pilot joins MIT's 95%.

And here is the part that should actually worry an operator: the disruption arrives on schedule anyway. 43% of Deloitte's respondents expect heavy job and process disruption within 12 to 18 months, and half admit they are underinvesting in the workforce transition. A failed pilot does not pause the market. Your competitors in the 15% are compounding at 31% monthly action growth while your team learns to distrust the word "agent." The cost of buying first is not just the wasted spend. It is entering the redesign era with an organization that has been taught agents do not work.

The work plan falls straight out of the barrier data, and none of it requires a big-firm engagement to start.

Pick one process that repeats weekly and annoys everyone. Map who touches it and which systems it crosses. Fix that process's data: systems of record, dead copies deleted, one owner per field. Write the approval rules an agent would need, in a table, not a memo. Get integration quotes for the two or three systems involved. Then, and only then, pilot a custom agent on it with humans handling exceptions.

This is also exactly the shape of ai automation consulting engagement to buy if you want help. Our 10x audit is this sequence run against your operations: it names the process, the data fixes, the approval rules, and the integration costs before anyone builds anything, and it is free. Where a firm proposes agents before it has asked about your data, you are looking at a future member of Gartner's 40%. And if the audit shows your data is genuinely a swamp, the honest advice is to fix that first and buy agents later, from us or from anyone.