Salesforce and Deloitte conducted a survey among 1,050 IT decision-makers within enterprise organizations. IT decision-makers appreciate the value of AI, but at the same time, they see problems. In 2026, AI adoption will continue to increase, and agent-to-agent communication will become the norm. This will lead to greater complexity and, in some organizations, chaos, although it can add value.
In recent years, AI has played a major role in business. It is seen as the way to improve organizations and make them more efficient. At the same time, it is a technology that is still in its early stages and will improve every year. For many organizations, AI adoption is both necessary and a pitfall, requiring very clear frameworks.
The average enterprise organization (with more than 1,000 employees) currently uses 12 AI agents, and this research projects that number will rise to 20 by 2027. At the same time, the IT decision-makers in this study state that the major challenges to better utilize AI are far from being resolved.
The AI problems in 2026
The average number of applications in organizations has risen from 897 to 957. Of these, only 27 percent are integrated with each other. Among IT decision-makers, 35 percent believe that integrating applications and data will hinder their AI ambitions.
In addition, 40 percent say they are dealing with outdated IT infrastructure that prevents the organization from using data for AI purposes. The well-known data silo problem, discussed for years, is still very much a reality.
However, we are not there yet; 42 percent are still busy mapping out the risks. Consider how the organization remains compliant and whether security is properly arranged with the current AI implementations. To make things even harder, 41 percent say that their organization currently lacks sufficient AI expertise to develop AI agents and AI processes. There is therefore a great need for more knowledge within organizations.
Shadow AI and AI governance are a major problem
There will still be IT decision-makers who prefer to postpone adopting AI because it’s hard, it involves a lot of work, and there are almost too many things to consider. Not only the investment and acquiring sufficient knowledge, but also the necessary integrations, replacing legacy applications with more API-driven applications, and above all, managing all the different AI solutions. AI governance is a major problem for many organizations.
Realistically, doing nothing is no longer an option. There is no turning back. This study shows that 83 percent of organizations already use AI and AI agents in virtually all teams. About 36 percent work with ready-made SaaS agents, 34 percent with integrated agents in platforms, and 30 percent with self-built AI solutions.
Employees have embraced AI tools en masse to assist them in their work and take work off their hands. If organizations are too passive in 2026 and offer no or insufficient AI tooling to their employees, they will have to deal with widespread Shadow AI. Employees will simply create free accounts on ChatGPT, Claude, Gemini, Perplexity, etc. This is not without risk, as many of these free accounts use the data provided to further train and improve their models. This means that trade secrets or confidential information could suddenly end up in a model. That is about the last thing you want as an organization.
That is why 86 percent of IT decision-makers believe that AI agents will add more problems than value without proper integration. AI adoption among employees remains high, but integration and data silos persist. If governance is also lacking, you can talk about serious problems and challenges.
In 2026, agents will also collaborate much more frequently
In addition, agent communication is now becoming a reality. Over the past year, there has been a lot of talk about MCP and A2A, protocols that allow agents to communicate with each other. But more and more agents that are now becoming available support and use them. Agents will soon be able to easily exchange information and transfer tasks to each other to achieve much better results.
Currently, 50 percent of AI agents in organizations still work as a silo. This means that no context or data from external systems is added. The need for context is now clear to many organizations. 96 percent of IT decision-makers understand that success depends on seamless integration.
This puts renewed pressure on data silos and integrations. These still hinder agent-to-agent communication, but more and more software will offer MCP integrations or more extensive APIs, enabling data integration.
The solution of agent management platforms
We see in the market that many large platform providers also recognize the governance issues and are trying to address them. Workday, ServiceNow, and Salesforce have introduced solutions to help integrate data silos and manage AI agents.
ServiceNow, for example, has the AI Control Tower, which actively monitors all agents on the NOW platform. What is still missing, however, is the ability to manage agents on third-party platforms. Then there is Workday, which wants to treat AI agents as employees in an HR database. At Workday, agents are part of the business teams and therefore have the same rights. In theory, this is not a bad idea, but here too, the ability to manage agents on third-party platforms is still lacking. For data integration, Workday recently acquired the iPaaS platform Pipedream.
MuleSoft is one of the first to manage all AI agents
Finally, Salesforce has had MuleSoft for years, an iPaaS platform, and recently added the MuleSoft Agent Fabric to it. They have opted for a kind of man-in-the-middle approach, whereby all data integrations and agent-2-agent communication takes place via the MuleSoft platform. With this man-in-the-middle approach, MuleSoft can offer governance and compliance on all data flows between applications and agents. For example, it can detect deviating data flows or simply block data flows that are not permitted. If a lot of personal information is being shared, for example, MuleSoft can simply block it. This is something many organizations still seek, as only 54 percent of enterprises have an AI governance framework in place. To be fair, MuleSoft does not yet manage agents on third-party platforms; it cannot disable a Workday agent, but it does provide control and insight into AI agents’ data exchange.
Ultimately, the problem is that there is currently no protocol for managing agents on third-party platforms. MuleSoft addresses this by acting as an intermediary. If you would like to know more, we had an extensive discussion about it earlier at Dreamforce. Watch the video below:
What will really be needed in 2026?
For IT decision-makers wondering what they really need to do in 2026, doing nothing is definitely not the right answer, as your competitors who do invest in AI will quickly overtake you. On the other hand, you don’t have to go all-in and blow your entire IT budget on it.
Start small, pick the low-hanging fruit
The most important thing is to make a plan that enables smaller, simple AI applications to yield significant gains. Redesigning your entire infrastructure for AI is a mega-project that will take years. You need to start now, so start small. Putting the three or five most frequently asked questions to your customer service or HR team into an AI agent can take a huge workload off those teams. There are now several case studies showing that this has reduced the number of tickets by as much as 50-60 percent. AI can also be used for sales reports or planning, which currently takes employees many hours each week. There is also a lot of low-hanging fruit and potential cost savings here.
Engage in conversation with your employees
Depending on the size of the organization, it is wise to engage in conversation with your employees. This can also be done with a survey if that works well within your organization. Questions you should definitely ask are:
- What do they think about AI in the workplace?
- What AI tools are you currently using?
- What kind of AI tools would you like to see to automate (repetitive) tasks?
- If the organization already offers AI tools, what do people think of those tools?
This will give you quick insight into how employees feel about AI, how they use it, where there is still low-hanging fruit, and what shadow IT is already being used.
Suppose a large number of people respond that they use ChatGPT, but as an organization you do not offer it because you have chosen to offer Gemini or CoPilot. This means employees are feeding the company’s data to OpenAI in the free version. The survey also shows that 49 percent of organizations see shadow AI as a major challenge.
Finally, simply offering OpenAI, Claude, Gemini, or CoPilot is not enough. Users need guidance and training. Adoption follows a good explanation, not simply making it available. With SaaS applications, AI functions are often already more focused on a specific function, but explanation and training are still necessary.
Innovation is outpacing adoption
At the end of the day, organizations are currently struggling to keep up. The adoption of AI is lagging behind the pace of innovation. AI is improving and becoming more widely applicable every year. By 2026, agent-to-agent communication will become the norm. While organizations are still searching for ways to roll out AI in a safe and effective manner, with the right governance. It is therefore not surprising that 86 percent of IT decision-makers are concerned that AI agents will add more problems than value without proper integration. However, with the right steps, it can add enormous value. For many organizations, this will still be too early in 2026, but certainly not for everyone.