Why does adding more agents sometimes reduce productivity? This question appeared on the desks of many corporate technology teams toward the end of 2025. AI agent configurations that delivered impressive pilot results were falling short of expectations six months after moving into production environments. The problem was rarely model quality. In most cases the RAG architecture had been set up correctly as well. The real issue was that the process handed to the agent had never been clearly defined by the humans running it in the first place. An ambiguous process does not get resolved through automation — it gets reproduced at greater speed and scale. This is precisely why the weakest link in the human-agent-process triangle is most often not the technical layer but the process itself.Consider a mid-sized automotive supplier in Bursa — 312 employees, parts shipped to five different OEMs, with exports accounting for roughly 57 percent of total revenue. In late 2025 the procurement department deployed an AI agent: it would compare supplier quotes, generate price-deviation alerts, and automate matching against GRN records. Within three months the team noticed that 46 percent of the agent’s alerts were being routed back to manual review, lengthening average approval time rather than shortening it. When they investigated, they found that the price-deviation threshold had never been defined in writing. A senior procurement specialist had been interpreting it intuitively for years. The agent applied a different assumption each time. The problem was not in the agent — it was in the undefined threshold.This pattern points to a straightforward principle: before handing a process to an agent, that process needs to be documented, tested, and bounded with explicit exception rules. The term for this is process maturity, and it is measurable. Does a process map exist? Are exception conditions written down? How many times in the past six months did human intervention occur, and at exactly which step? If those three questions cannot be answered, the process is not ready for an agent. Automating an immature process does not eliminate manual chaos — it converts it into numerical speed. For Turkish exporters, the EU AI Act’s requirements in force since 2025 make this directly relevant: decisions that cannot be documented are no longer merely an operational risk but a compliance exposure toward EU customers.What about the human side? Most change management frameworks reduce employee adaptation to a training checklist. The picture I see in practice is different. Resistance to agents rarely stems from not understanding the technology; it comes from accountability boundaries becoming unclear. In the Bursa factory, the procurement specialist’s real question was: if the agent accepts a wrong quote, who is responsible? When that question went unanswered, the specialist started re-verifying every agent output — which extended the timeline instead of compressing it. The accountability map — specifying which decisions remain with the agent, which require escalation to a human, and which demand explicit human approval before action — must be drawn before the technical architecture is designed. Without it a project can work technically while failing organizationally. That distinction matters more than the model version chosen.On the question of SLM (Small Language Models) versus cloud API: many Turkish SMEs shifted during 2025-2026 toward on-premise SLMs rather than large hosted models, and the driver was not primarily cost. Supplier pricing data, customer-specific terms, and production process parameters are commercially sensitive. Sending that data to an external API carries both KVKK and trade-secret risk. A practical rule of thumb: use cloud API for high-frequency, low-sensitivity tasks that do not involve proprietary data; use an on-premise SLM or at minimum a contractually isolated environment for tasks that involve the firm’s core operational knowledge. Do not make this decision by looking at the cost table. Make it by looking at the data flow diagram. Which data goes where, who can access it, how is it logged? One measurable indicator: of all records processed by agents each week, what percentage fall into sensitive data categories, and how many of those passed through an external API endpoint?Now for the limit of this argument, because presenting the human-agent-process triangle as a universal template would itself be misleading. Firms that have achieved process maturity, drawn accountability maps, and made clear data sovereignty decisions can still fail to extract expected value from agent integration — if the process in question needs redesign rather than automation. Some steps accelerate when automated but are fundamentally unnecessary: speeding them up adds no value. In the Bursa case, the GRN-to-supplier-quote matching step completed 73 percent faster with the agent in place. But when the team realised that this step was followed by an approval layer that served no operational purpose, most of the gain evaporated. The agent was doing the right task at the wrong place in the process. Sequence is not negotiable: process design first, then automation.Managing the human-agent-process triangle is an operations design problem, not a technology project. If you can define what you expect the agent to produce, list its exception boundaries, and identify who intervenes when it makes a mistake, that process is ready. If those three questions yield uncertain answers, the work ahead of you is managerial rather than technical. Three concrete steps for Monday morning: first, document the process you intend to hand over, including every exception that occurred in the past three months; second, put the agent-human accountability boundary for each decision point into a single-page table; third, define your sensitive data categories and map which step routes data to which system. Without those three documents in place, spending money on technical infrastructure is the equivalent of installing a roof on a foundation that has not been poured yet. The procurement specialist in Bursa is now asking exactly these questions — and that moment, not the go-live date, is when the project actually began.
This article was originally published in Turkish by Gökhan MERCANOĞLU on September 2, 2026. The English edition has been reviewed and edited by the author.