For the last couple of years, AI has already changed how work gets done. It has helped teams accelerate development, improve decision preparation, analyze complex information, and produce higher-quality work with less friction. But the human still remained the central operator. The emerging agentic AI era is different.
AI agents are becoming a visible product category and not just another class of productivity tools. They are moving toward work execution. The promise is no longer that AI will help me do my work. The promise is that AI agents can help deliver the work itself.
But that shift changes the enterprise challenge. A chatbot can be wrong and still be useful. An agent can be wrong and create damage. If AI drafts a bad message, you reject it. If an autonomous agent sends it, books the order, changes the record, or triggers the process, you own the consequences. The agentic era moves us from editing suggestions to governing execution. And execution requires more than intelligence. It requires trust.
That is why we are introducing Company Memory. It is a concept we have been working on to answer a simple but urgent question: what guardrails, context, and capabilities need to be in place before businesses can trust AI agents with high-stakes work?
The emerging agent stack is becoming standardized
Across the market, the basic architecture of agents is becoming clearer, as most systems converge on a similar pattern: a conversational interface, a memory layer, access to tools, and a goal-oriented loop.
This agentic loop is powerful. I see it in my own work and in the work of our organization, which has dramatically increased the velocity of its feature releases. But as more companies adopt similar agentic systems, similar orchestration patterns, and similar foundation models, the question becomes, if everyone has the same AI agents, what will make one company’s agents better, safer, or more valuable than another’s?
The answer won’t be AI models. They commoditize quickly, and businesses will keep switching them as the price/performance frontier moves ahead.
I am convinced that the real differentiation will come from what surrounds the model: context, constraints, tools, governance, and process knowledge. In other words, the harness that turns raw AI capability into trusted enterprise behavior. More specifically, it will come from the institutional memory that has shaped your company's success.
Out of the box, even a very capable AI agent is just like a brilliant intern on the first day of work. It may be fast, articulate, and impressive. But it does not yet understand how the company actually works.
The challenge is not only that agents need more information. It is that they need the right kind of company-specific knowledge, namely, current, governed, procedural, and usable at runtime. That is the gap we address with Company Memory.
From facts to procedural knowledge
What is the right kind of memory, then? Simply adding more data to an agent's context window certainly increases cost but not business performance. For me, a useful distinction is that between semantic memory and procedural memory.
Semantic memory is factual knowledge. A company’s suppliers, contracts, systems, customers, process names, policies, and documents. This matters, and I have seen enterprises making significant progress in utilizing this kind of knowledge.
But business execution depends just as much on procedural memory. Procedural memory is the knowledge of how things are done. How decisions are made, when exceptions apply, which trade-offs matter, what good judgment looks like, and how the company wants work to move through its systems.
Enterprise agents increasingly have access to facts. What they often lack is the procedural layer, the “how” and “why” behind company-specific business logic. If semantic memory is knowing the parts of a car, procedural memory is knowing how to drive it.
That is where SAP Signavio becomes the bridge. Imagine an urgent server replacement in one of your data centers. The agent finds the fastest available supplier and orders the server. The server arrives. So does the invoice for $25,000.
A generic assistant might suggest contacting finance, documenting the urgency, or explaining why the usual process was bypassed. That is helpful, but not enough. An agent powered by Company Memory can go much further. It can check whether a sourcing contract exists, whether a purchase requisition or purchase order was created, whether goods receipt has been posted, whether the amount triggers an approval threshold, and whether an exception workflow applies.
Most importantly, it can flag what is missing and guide the case back into the company’s actual procurement logic. That is the difference between an agent that knows procurement in general and an agent that understands how your company wants procurement to work.
Your processes are your institutional knowledge and competitive differentiation
I still see companies that misunderstand processes. They think of them as mere documentation or something that needs to be updated before an audit.
Yet, we know they can be so much more. Processes are the institutional knowledge of your company. They capture the rules, behaviors, preferences, decision patterns, guardrails, and exceptions that make your business work.
This implicit knowledge is often one of a company’s most valuable assets. It is also one of the hardest assets to make available to agentic AI. That’s why SAP Signavio’s Company Memory turns that institutional knowledge into something agents can use.
What Company Memory does
Company Memory is not a static knowledge base. It is a set of capabilities that help companies turn fragmented business knowledge into governed, agent-usable context. It works in three steps.
First, acquisition. Company Memory helps bring together knowledge from sources such as process models, decision models, documentation, process logs, SAP content, and other enterprise systems.
Second, synthesis. It helps translate that knowledge into agent-usable building blocks, based on our Process Atoms research. Think of Process Atoms as small, governable units of company logic that can shape business-relevant decisions. They are not just decision rules. They can include preferences, guidelines, escalation logic, historical behavior, exception patterns, operating conventions, and trade-offs.
Third, application. Company Memory provides the relevant subset of company knowledge to agents at runtime, so they can check their plans and actions against how the company wants work to be done. This becomes especially important as agents take on longer-running tasks and interact with more systems.
The governance layer is central. Company knowledge is always in flux. Policies change. Processes change. Regulatory requirements change. Business priorities change. Company Memory helps keep this knowledge current, coherent, and usable by agents. It can identify conflicts, duplicates, outdated guidance, and gaps in the logic agents rely on.
This is where compliance and behavioral control become practical. Company Memory can help define what an agent may do, what it should escalate, what it must not do, and what evidence it needs before acting. It can also support validator agents that inspect plans, outputs, and tool calls against approved company logic.
The goal is not to preload agents with every possible piece of information. The goal is to equip them with the right guidance and guardrails at the right moment.
From process excellence to agent excellence
SAP Signavio’s mission has always been to help companies run better processes and make business transformation easier. I know that this mission becomes even more relevant in the era of AI agents.
As agents take on more work, companies will need to better understand how agents behave. Their execution paths become traces for mining. This extends process observability into the agent era.
Scaling AI agents without trust is not viable. Even if AI agents perform well most of the time, the remaining failures can have disastrous consequences, or as they say, one bad apple can ruin the entire batch. Additionally, as AI models and agents become more widely deployed, differentiation shifts to your institutional knowledge and your ability to scale what makes your company unique and successful at the speed of AI.
For me, this is the exciting part, we are not just helping companies adopt agents. We are helping them make agents understand how the business actually works. The companies that win in the agent era will not simply be the companies with the most agents. They will be the companies whose agents best embody the knowledge, judgment, and operating logic that made the company successful in the first place.
With thanks to Gero Decker and Lukas N.P. Egger for their input, review, and the many discussions that helped sharpen this piece.