The agent market has moved past the demo phase, but not into the fully autonomous future many decks promised. Capgemini's 2025 report on agentic AI says 14 percent of surveyed organizations have already implemented AI agents at partial or full scale, 23 percent are piloting, and 61 percent are preparing or exploring deployment. That looks like fast adoption. The more useful signal for builders is the brake pedal: only 27 percent of organizations say they trust fully autonomous AI agents, down from 43 percent a year earlier. In other words, buyers want agents, but they do not want black-box coworkers roaming across production systems.
Source Card
PDF Rise of agentic AI - CapgeminiThe report is useful because it separates adoption enthusiasm from operational readiness. It surveyed 1,500 leaders across 14 countries and highlights a gap that agent builders should design for: organizations are experimenting broadly, but most still lack mature data foundations, infrastructure, governance, and confidence in high-autonomy systems.
capgemini.com
The builder takeaway is not that agents are overhyped. It is that the product surface is changing. An agent product can no longer be sold as a smarter chatbot plus tools. It has to be packaged as a controlled execution system with clear scope, recoverable state, visible reasoning traces where appropriate, approval gates, policy enforcement, and measurable handoff quality. The winning infrastructure will make autonomy adjustable, not absolute. Enterprises will buy systems that can start as narrow task runners, graduate into supervised team members, and prove reliability before they receive more permissions.
Key Takeaways
- Enterprise agent adoption is broadening, but high autonomy remains rare. Capgemini reports only 15 percent of business processes are expected to run at semi-autonomous to fully autonomous levels in the next 12 months.
- Trust is declining as buyers learn where agents fail. That creates demand for evaluation, observability, permissions, audit logs, and human approval workflows.
- Data readiness is a hard scaling limit. If the agent cannot access current, governed, high-quality context, better reasoning models will not fix the product.
- Intermediate autonomy is commercially valuable. Builders should optimize for supervised execution, not wait for fully autonomous agents to become acceptable.
- The agent stack is becoming organizational infrastructure. Deployment now touches workflow design, workforce anxiety, security, compliance, and operating model change.
The autonomy ladder is the product
Capgemini defines AI agents as software connected to a business environment, operating within boundaries, making decisions, and acting toward goals with varying levels of human involvement. That definition matters because it moves the conversation away from model capability alone. A useful enterprise agent is not just a model call. It is a loop: perceive context, plan a step, call a tool, inspect the result, update state, decide whether to continue, ask for help, or stop. Each loop creates risk. The product question is how much loop authority the system should receive in a given workflow.
| Signal | Why it matters |
|---|---|
| Only 2 percent of surveyed organizations report full-scale implementation | Builders should expect pilots, phased rollouts, and proof-heavy procurement rather than instant enterprise-wide deployment. |
| Trust in fully autonomous agents fell to 27 percent | Autonomy controls, rollback, and explainability are buying criteria, not compliance decorations. |
| Fewer than one in five organizations report high data readiness | Retrieval, data contracts, catalog quality, permissions, and freshness checks are part of the agent product, even if they feel like plumbing. |
| More than 80 percent lack mature AI infrastructure | There is room for agent platforms, observability layers, evaluation harnesses, orchestration tools, and managed runtime services. |
| Most near-term adoption is expected in customer service, IT, and sales | The first durable markets are workflows with high volume, measurable outcomes, and existing escalation paths. |
Builder note
Do not treat autonomy as a single product setting. Break it into permissions: read access, write access, external communication, financial authority, customer-facing action, system configuration, and ability to spawn follow-up tasks. A support agent might draft replies automatically, update a CRM only after confidence checks, and refund an order only below a policy threshold. This is more work than a simple agent mode toggle, but it maps to how enterprises actually allocate risk.
Where agent infrastructure is becoming mandatory
The report's data-readiness and infrastructure findings should make agent founders cautious about promises that depend on clean enterprise context. Agents degrade quickly when they retrieve stale policies, miss permissions, confuse similar records, or inherit fragmented workflow state from legacy systems. In a chatbot, that may produce a bad answer. In an agent, it may trigger a bad action. The infrastructure need is a trusted context layer: governed connectors, identity-aware retrieval, event history, state storage, source attribution, and freshness metadata. Without that layer, tool use becomes a liability because the agent can act confidently on broken inputs.
The next agent platform will not be judged by how often it can act alone. It will be judged by how safely it knows when not to.
- Start with bounded workflows that already have human escalation, such as ticket triage, IT incident summaries, sales research, invoice exception handling, or internal knowledge routing.
- Define the agent's action budget before model selection: which tools it can call, which records it can modify, which customers it can contact, and which decisions require approval.
- Instrument every run with task objective, retrieved sources, tool calls, intermediate state, confidence markers, human interventions, and final outcome.
- Build an evaluation set from real failed work, not synthetic happy paths. Include missing data, contradictory policies, ambiguous user intent, tool timeouts, permission denials, and adversarial instructions.
- Ship a human handoff path that preserves context. If the agent fails, the human operator should see what the agent tried, what evidence it used, and why it stopped.
This is also a workforce product problem. Capgemini reports rising employee anxiety around agents and notes that many organizations expect agents to become members of human-supervised teams over the next three years. Builders often underweight this because they focus on task completion metrics. Operators care about adoption friction. If an agent quietly changes a workflow, employees may route around it, double-check everything, or refuse to trust its outputs. A better rollout pattern is explicit collaboration design: name the agent's role, publish its limitations, show its evidence, invite corrections, and track whether humans accept, edit, reject, or escalate its work. Acceptance rate is not just a UX metric. It is a trust metric.
Failure modes to design for now
- Tool overreach: the agent completes the wrong action because the tool schema allows too much. Fix with scoped tools, constrained parameters, dry-run previews, and policy checks.
- Context poisoning: the agent retrieves untrusted or outdated content and treats it as authoritative. Fix with source ranking, freshness labels, permission-aware retrieval, and exclusion rules for low-trust repositories.
- Silent drift: workflow behavior changes as prompts, models, data, or tools are updated. Fix with regression suites, versioned prompts, canary releases, and run-level comparison.
- Human approval theater: a reviewer rubber-stamps actions because the approval screen lacks evidence. Fix by showing the decision basis, changed fields, risk flags, and alternative actions.
- Unclear accountability: teams cannot tell whether a failure came from the model, tool, data source, policy layer, or user instruction. Fix with traceable execution and ownership boundaries across the stack.
The uncertain part is how fast buyers will move from supervised agents to higher-autonomy systems. Capgemini projects that the share of business processes operating at semi-autonomous to fully autonomous levels could rise from 15 percent in the next 12 months to 25 percent by 2028. That is meaningful, but it is not a takeover curve. It suggests a long middle period where agents act like junior operators inside guarded workflows. For founders, that means the best commercial wedge may be boring but durable: reduce handle time, improve compliance, cut rework, increase throughput, and make expert employees less interrupt-driven.
Adoption guidance
If you are building or buying agent infrastructure, ask for proof at the workflow boundary. Can the system show which data it used, which policy it applied, which tool changed state, and which human approved the action? Can it replay a failed run? Can it pause when confidence drops or policy conflicts appear? Can permissions be narrowed per role, region, customer tier, or transaction amount? These are the questions that separate a demo agent from an operational agent.
- Capgemini Research Institute, Rise of agentic AI: How trust is the key to human-AI collaboration, 2025.
- Source URL: https://www.capgemini.com/wp-content/uploads/2025/07/Final-Web-Version-Report-AI-Agents.pdf
- Reported survey base: 1,500 leaders at director level and above across 14 countries, with adoption figures reconfirmed for a majority subset of respondents according to the report.