Processes Before Prompts - why agentic CRM will be won on the process layer
I have similar conversations once a week. A board member, a CEO, sometimes a CIO ask very similar question: „Jakub, we need to do agentic AI. We’ve looked at three platforms. Which one would you recommend?” And every time my answer disappoints them, because I don’t start with the platform, nor with a model. I ask them to tell me about the process the agent is supposed to automate.
Eight times out of ten the answer is like, “Well, we have different processes, what do you mean?“
That pause is where most agent projects die before they arrive - in the gap between what a company thinks its process is and what its process actually is.
This is my first post on this Substack and the title is also the thesis. Processes Before Prompts. I’ve been saying it to clients for some time, when we started do help with intelligent automation and later with building agentic capabilities. Time to write it down because the decisions made over the next 12 months will determine whether your firms come out of the agentic era as what Microsoft calls Frontier Firms or as the next generation of cautionary case studies titled “We tried AI, it didn’t work”.
Important numbers about agentic AI
According to MIT1, mid-2025: 95% of corporate generative-AI pilots produce no measurable P&L impact.
McKinsey’s „The state of AI in 2025: Agents, innovation, and transformation“2 shows that 88% organizations report regular AI use in at least one business function and 39% report EBIT impact at the enterprise level, whilst only 6% of organizations qualify as “high performers” who see significant enterprise-wide value.
BCG in their “The Widening AI Value Gap”3: 60% companies are reaping hardly any material value, reporting minimal revenue and cost gains despite substantial investment.
IBM’s CEO study “Rewiring the C-suite: The fast track to 2030”4: 16% of AI initiatives have been scaled enterprise-wide and only 25% have delivered the expected ROI.
Gartner forecasts that by the end of 2027, more than 40% of agentic AI projects will be cancelled, killed by escalating costs, unclear business value, and lack of risk controls.
So, it looks like a pattern.
It is the same disease that was causing CRM failures
Technology was never the bottleneck. Processes were.
In 2026, we are about to make exactly the same mistake like many companies did with treating CRM systems as software, not strategy. They license a platform, contract a partner to „implement it“, instruct users to “use“ the system, and 6 months later wonder why adoption is flat and the pipeline forecast is still being managed in Excels.
This time, however, the mistake will be more expensive, faster and with dramatically higher consequences. Because an agentic CRM does not behave like a traditional CRM. A traditional CRM is a passive system of record - it lets you store data. An agentic CRM is a system of action, which executes actions on this data. And it executes whatever process it was handed. If you haven’t handed it one, it improvises off the mess.
As Bill Gates once said, automation applied to an inefficient operation will magnify the inefficiency.
The agentic enterprise, what Microsoft has rebranded as the Frontier Firm, is not a software upgrade. It is a deliberate and controlled re-design of a process.
Processes before prompts
Processes Before Prompts means that before any company deploys an AI agent into production, three things have to happen and they have to be owned by the business and not by IT.
The first is that the process has to be discovered, mapped, and named. Not the imagined process from the playbook, bu the actual process, observed in the data. Process-mining tools such as Celonis, Microsoft’s own Power Automate Process Mining, SAP Signavio exist because most companies do not know how their work actually flows. Wil van der Aalst, Chief Scientist at Celonis, ran a tutorial at the 28th European Conference on Artificial Intelligence 2025 titled “No Enterprise AI Without Process Intelligence! Using Process Mining as the Lens to Address Performance and Compliance Problems”. The argument is that an agent without process context is an autonomous actor in a building with no floor plan.
The second is that the process must have a human owner with the authority and the budget to redesign it. Not a “champion“ and not a “stakeholder” - an owner. McKinsey is clear on this, stating that of the 25 organisational attributes they tested, the one most strongly correlated with measurable EBIT impact from generative AI is workflow redesign. AI high performers are more than 3x more likely to fundamentally redesign work than the rest. Direct CEO involvement in AI governance is the second strongest variable. Implementing AI is not a technical project. It is a leadership project, and the one of the causes of failure is treating it as a technical one.
The third is that the process must be instrumented for measurement before the agent is deployed. You cannot improve what you cannot see. You cannot trust an agent whose outputs you have nothing to compare to.
Bill McDermott of ServiceNow opened his Knowledge 2026 keynote on Tuesday with a story about an AI agent hitting a credential error and deleting the entire production database in 9 seconds at one of the startups. McDermott's conclusion is concise:
Governance isn't a feature, it's the whole ball game. Because without it, your whole company can come down.
Here we arrive at the three main pillars: discovery, ownership and telemetry.
Specifically about agentic CRM
I run a firm that has been implementing enterprise CRM systems since 2008 and now builds agentic CRMs on the Microsoft stack. Our voice agent connected with MCP servers is already in the future - it can take meeting notes in natural voice, link opportunities to quotes, generate tasks based on meetings and customer overlook - all permission-aware, all anchored in the actual CRM record. We did not build that without first defining with the customer what a qualified opportunity means, where tasks are stored and who owns the conversion.
Every week we educate our clients and potential clients that giving every salesperson an AI agent, when they are not even consistent when a specific sales stage should switch, is not a strategy. It is a delegation of strategy to chaos, potantially with a Microsoft Copilot or Salesforce Agentforce logos on top.
Building an agent is just new programming, a technical task. A fairly easy piece for experienced developers and AI practitioners. Aligning on the process and building a helpful agent is the hard piece.
Salesforce’s assessed agent performance honestly in their own paper from 2025. On their CRMArena-Pro benchmark frontier LLMs scored 58% on single-step CRM tasks and 35% on multi-step ones. 35% on multi-step workflows is, as one analyst put it, a non-starter for enterprise. And that was in laboratory conditions without the technical debt of a real organization.
The European (and Polish) angle
I am writing this from Warsaw and most of the leaders and our clients reading this run businesses across the world. The Polish and European context shows the urgency. According to Eurostat only 8,4% of Polish enterprises had adopted any form of AI in 2024. It is the third-lowest figure in the European Union. And yet the Polish edition of Microsoft’s Work Trend Index5 reports that 84% of Polish business leaders plan to deploy AI agents within the next 12 to 18 months, slightly above the global average. We are, in other words, about to leap from one of the lowest AI adoption rates in Europe to mass deployment of autonomous systems in 12 months, on top of companies that have not done the process work.
Add the EU AI Act, whose enforcement against high-risk systems begins on 2026-08-02 and Article 22 of the GDPR, which already prohibits significant decisions taken solely by automated means without meaningful human oversight. Add the overlapping NIS2 obligations across regulated sectors.
Companies that introduce agents on undocumented processes will probably spend 2027 paying lawyers what they should have spent in 2026 paying process designers.
What to do next?
Stop implementing AI agents, because “everyone does that”. Instead, ask each business unit owner two questions:
In which of our processes do we want an agent to operate?
Can we describe that process in enough detail that a new hire could perform it on day one?
If the answer to the second question is no, you do not have a process.
Designate a single accountable owner per process, not per agent. Agents will multiply, retire, and be replaced by better agents with better model next quarter. Processes are durable and their owners outlive tools. That is the lesson of twenty five years of my CRM history applied to the next decade of agentic history.
Treat the AI Act not as a compliance enforcement but as a discipline-forcing one. Article 14, with its requirement for “human oversight” of high-risk systems, read correctly, is an instruction: know your process well enough to know where the human belongs in it. This could mean that companies that take this seriously will end up with better agentic AI than their American competitors whose board members pay for AI with their credit cards without proper process redesign first…
Conclusion
Microsoft’s 2026 Work Trend Index summarises the moment in a sentence I have already quoted at several boards: “the workers are ready; their organisations are not”. Jared Spataro, who runs Microsoft’s AI at Work efforts is right - the constraint on the agentic enterprise is no longer what the AI can do. It is how we have structured the work around it.
The companies that come through the next 18 months will not be the ones with the best prompts, or the most agents, or the largest token budgets. They will be the ones whose CEOs and COOs decided, before the rest of the market did, that the answer to the question “which model should we pick?” was always “how does the process really work?”.
Processes Before Prompts. Welcome to my Substack. Subscribe if you’d like to argue.
The GenAI Divide: State of AI in Business 2025, July 2025.
The state of AI in 2025: Agents, innovation, and transformation, https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai, November 2025
The Widening AI Value Gap, https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap, September 2025
Rewiring the C-suite: The fast track to 2030, https://www.ibm.com/thought-leadership/institute-business-value/en-us/c-suite-study/ceo
Work Trend Index, https://www.microsoft.com/en-us/worklab/work-trend-index
And the next big thing will be data quality. But I'm sure You will cover that also.