The “93%” figure does not mean 93% of companies already run autonomous agents or have proved business value. Salesforce and MuleSoft’s 2025 research reported that 93% of surveyed enterprise IT leaders had implemented—or planned to implement—AI agents within two years. The same research found major integration problems, including difficulty connecting AI to existing systems and missed delivery goals. The practical lesson is clear: enterprise enthusiasm is high, but production value depends on data access, safe permissions, reliable integrations, governance and measurable workflows.
What the 93% statistic actually measured
The figure came from Salesforce/MuleSoft’s 2025 Connectivity Benchmark research, described as a survey of 1,050 enterprise IT leaders worldwide. It referred to respondents that had implemented or planned to implement AI agents within a two-year period—not to 93% of organizations already operating autonomous software successfully. VentureBeat’s report also attributed two related findings to the research: 80% identified data integration as a major AI challenge, and 29% missed delivery goals in 2024.
MuleSoft separately reported that 95% of IT leaders had difficulty connecting AI to existing systems and that enterprises averaged 897 applications. Developers spent an average 39% of their time designing, building and testing custom integrations. Those are survey averages, not a forecast for every company. MuleSoft’s 2025 report does not turn deployment plans into independently audited production results.
Do not combine the different 93% claims
| Statistic | Scope and meaning |
|---|---|
| 93% | Global enterprise IT leaders that had implemented or planned to implement AI agents within two years in the 2025 research. |
| 97% | APAC enterprise IT leaders that had implemented or planned to implement agents within two years; a different regional release and population. |
| 93% | APAC IT leaders reporting data silos as a business challenge; not an adoption rate. |
| 93% | CIOs in Salesforce’s 2026 research saying workplace adoption depends on integrating agents into everyday work; a different survey and wording. |
“AI agent” is also not a standardized category. A vendor survey may include copilots, workflow automation, assistants that retrieve information, and more autonomous systems. A read-only assistant is materially less complex and risky than an agent that changes records or coordinates transactions.
#1 Best Overall
Why promising pilots stall before production
Disconnected data and applications
An agent needs current, permissioned information. Customer, order, inventory, employee, financial and operational records often sit in incompatible systems with different identifiers and definitions. An agent can produce fluent text while relying on stale or incomplete data, or it may lack the authority to take the requested action.
In an APAC-specific Salesforce release, organizations using agents averaged 1,130 applications versus 771 among organizations not using agents, and 98% of organizations using agents reported data silos as a challenge. The comparison is regional and should not be generalized to every enterprise. Salesforce’s APAC release also reported that 93% of APAC leaders planned or had implemented agents, a figure distinct from the global 2025 statistic.
Integration is operational engineering, not just a connector purchase
- Expose legacy data through dependable APIs or controlled alternatives.
- Map inconsistent fields, identities and business definitions.
- Propagate user and service-account permissions correctly.
- Orchestrate multi-step work with timeouts, retries, idempotency and rollback.
- Monitor tool calls, failures, latency and data quality as applications change.
- Assign owners for integration maintenance and version upgrades.
A platform can accelerate this work, but it cannot remove the need for semantic mapping, authorization design, transaction consistency and process ownership.
Legacy systems limit what an agent can do
Separate four capability levels before approving a use case:
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitches- Knowledge access: retrieving a record or policy.
- Decision support: recommending an action for a person.
- Transaction execution: changing a record, issuing a refund or triggering a process.
- Autonomous orchestration: coordinating several systems with limited human intervention.
Risk, testing burden and recovery requirements rise sharply at each level. A mainframe with batch-only processing or a system requiring manual approval may support an answer but not safe autonomous execution.
Governance and security gaps
An agent is a non-human identity that can read data, invoke tools and potentially make consequential changes. Required controls include least-privilege access, explicit tool allowlists, secrets management, data-loss-prevention rules, approval gates, audit logs, prompt-injection defenses, separation of duties, human escalation, adversarial testing and rollback procedures.
Rank #3
MuleSoft’s 2026 research says only 54% of organizations reported a centralized governance framework for AI agents and that 50% of agents operated in isolation. These definitions and survey responses should be validated against your own architecture before they are used as a benchmark. MuleSoft’s 2026 benchmark also reported that 95% of organizations faced integration challenges and 96% said agent success depended heavily on seamless integration.
Skills, ownership and economics
Production deployment requires enterprise architecture, API engineering, identity and access management, security, compliance, model evaluation, workflow design and change management—not only data science. Salesforce’s 2026 CIO research found 94% of CIOs saying agents increase the need to expand skills and 81% saying they require closer collaboration with functions such as HR, finance and sales. That research is separate from the 2025 Connectivity Benchmark.
Recommended Free Tools
Track completed outcomes rather than conversations or model calls. Useful measures include task-completion rate, factual accuracy, tool-call success, escalation and exception rates, time saved, cost per completed task, customer-satisfaction change, security incidents and human-review workload. Include model, API, integration, storage, monitoring and supervision costs in the denominator.
What “production-ready” means
Before moving beyond a pilot, document all of the following:
- A specific process, business owner and accountable support team.
- Explicit authority boundaries and a list of prohibited actions.
- Reliable, current and permissioned source data.
- Tested APIs, identity propagation and failure handling.
- Acceptance thresholds for accuracy, latency, cost and escalation.
- Full logs of prompts, retrieved data, tool calls, approvals and outcomes.
- A fast human fallback for ambiguity or high-impact decisions.
- Rollback, compensation or recovery procedures for partial failures.
- Ongoing monitoring, red-team testing and a change-management process.
- A cost model that compares automation with human handling and error costs.
Choose a first use case with bounded risk
The strongest first deployments are high-volume, measurable and reversible. Suitable examples include customer-service triage, internal IT-help-desk resolution, employee-policy lookup, sales-research preparation, document classification, ticket summarization and routing, order-status inquiries, onboarding assistance, data-quality checks and routine report generation.
Use substantially more caution for healthcare, credit, insurance or employment decisions; legal conclusions; financial transactions; production-infrastructure changes; unrestricted refunds; security-control changes; and actions involving regulated or highly confidential data. A human approval gate may be mandatory even when the retrieval portion is automated.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Best Value
Build, buy or use a hybrid architecture?
| Approach | Usually fits when | Main trade-off |
|---|---|---|
| Buy a platform | You need prebuilt connectors, governance and vendor support, and the workflow matches the platform. | Faster delivery can mean ecosystem dependence, licensing complexity and less control. |
| Build internally | The workflow is a durable differentiator, APIs are strong and platform-engineering and security teams are mature. | More control and portability, but you own integration, evaluation, support and upgrades. |
| Hybrid | You buy a reasoning or agent layer while exposing enterprise actions through internal APIs and approval services. | Can preserve control over sensitive systems while creating more integration seams to operate. |
How major platforms map to existing estates
- Salesforce-centric: Agentforce, often with MuleSoft for external systems. See Agentforce, MuleSoft AI connectivity and MuleSoft for Agentforce. Exact pricing should be confirmed through Salesforce.
- Microsoft-centric: Copilot Studio and Azure AI services for organizations using Microsoft 365, Entra ID, Power Platform and Dynamics. See Copilot Studio and Azure AI Foundry. Pricing and capacity models can change.
- AWS-centric: Amazon Bedrock for engineering-led teams that want multiple models and AWS-native controls. See Amazon Bedrock; pricing is consumption- and model-dependent.
- Google Cloud-centric: Vertex AI Agent Builder for teams using Vertex AI, BigQuery and Google data services. See Vertex AI Agent Builder.
- Service-management-centric: ServiceNow AI agents when IT, employee or customer workflows already run in ServiceNow. See ServiceNow AI agents.
- RPA and legacy-UI-heavy: UiPath when desktop automation and systems without clean APIs are central. See UiPath agentic automation.
MuleSoft advertises a 30-day Anypoint Platform trial without a credit card or installation through its Agentforce page, but trial access is not evidence that a production architecture will be simple. Check the stated trial terms and obtain current commercial terms directly from the vendor.
Questions executives should ask before approving a rollout
- Which exact business outcome is being automated, and who owns it?
- What data and tools must the agent access, and are they permissioned and current?
- What is the worst plausible failure, and can the action be reversed?
- Where is human approval required, and how quickly can a person intervene?
- How will every retrieval, decision, tool call and exception be audited?
- What is the cost per completed task after integration and review labor?
- Can the workflow move to another model or platform without rebuilding core business logic?
- What evidence will trigger expansion, redesign or shutdown?
Salesforce and MuleSoft benefit commercially from an integration-centered interpretation because they sell CRM, integration, automation and agent products. That interest does not invalidate the findings, but it makes attribution, methodology, independent measurement and a clear denominator essential. “Plans to deploy” are weaker evidence than reliable production outcomes, and integration is a major bottleneck—not the only one.
The Bottom Line
Bottom line: The 93% claim is best read as evidence of enterprise intent, not proof of widespread autonomous-agent success. Organizations should invest only after proving that a narrowly defined process has reliable data, controlled permissions, observable integrations, human fallback and economics that remain positive after errors and supervision.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
