Few terms have made the journey from research niche to boardroom slide as quickly as "Agentic AI". Gartner lists multi-agent systems among the strategic technology trends for 2026, 89 percent of executives expect agentic AI to be standard within three years – and at the same time, the same analyst firm predicts that over 40 percent of agentic AI projects will be canceled by the end of 2027. Both numbers are correct. Reading them together tells you more about the topic than any vendor deck.
Definition: what separates an agent from a chatbot
A chatbot answers requests. An agent pursues goals: it breaks a task into steps, accesses systems through tools – APIs, databases, file systems –, evaluates intermediate results, and corrects its plan. Autonomy is not a switch but a spectrum:
| Level | What the system does | Example | Autonomy |
|---|---|---|---|
| Chatbot / assistant | answers requests from knowledge or documents | FAQ bot, RAG search across product data | none – purely reactive |
| Tool assistant | executes individual defined actions | booking an appointment, creating a ticket | low, confirmed per action |
| Task agent | plans and completes scoped tasks in multiple steps | checking incoming invoices, extracting data, posting to the ERP | medium, within guardrails |
| Orchestrated agents | multiple specialized agents share a workflow | quote creation: data retrieval, calculation, document, approval draft | high, with defined human gates |
Forrester describes the development in software as a turning point in 2026: from individual coding assistants to orchestrated agents across the entire development lifecycle – with "compound gains" instead of isolated islands of productivity. The same logic applies to business processes: the value is not created in the individual agent, but in cleanly scoped, monitored chains.
The uncomfortable number: why 40 percent of projects fail
Gartner's cancellation forecast cites three causes: escalating costs, unclear business value, insufficient risk management. From our project experience, this translates as follows: failed agentic projects almost always start with the technology instead of the process. An agent gets built because agents are what gets built right now – without a defined success metric, without a process owner, without a plan for the case that the agent gets it wrong. The State of the CIO 2026 confirms the pattern: 32 percent of CIOs name undefined ROI metrics as the biggest hurdle, 31 percent an unclear AI strategy.
Where agentic AI realistically delivers value for mid-sized companies today
- Document workflows: incoming invoices, contracts, forms – extract, validate, post to systems, with an approval gate.
- Knowledge access with RAG: making product data, documentation, and legacy systems searchable – as an assistant for sales, support, or editorial teams.
- Process automation: multi-step workflows between CRM, ERP, and email, orchestrated with n8n plus LLM steps, for example.
- Development processes: agents for tests, migrations, and boilerplate – under senior review, as we practice it ourselves.
What is notably absent from this list: the autonomous universal agent that "simply takes care of everything". The viable cases are tightly scoped, measurable, and have defined handover points to humans.
What decision-makers should check before starting
- Success metric first: what is the measurable target – cycle time, error rate, cost per transaction? No number, no project.
- Human-in-the-loop as architecture: at which points does a human decide? These gates belong in the design, not in the rework.
- Start small, orchestrate later: one process, one agent, four weeks to proof of value – then scale.
- Build in governance: tool approvals, data access, EU AI Act classification, and logging from day one.
- Plan the exit: agent logic belongs in separate, replaceable layers – models and vendors change faster than contracts run.
Our assessment
Agentic AI is not a hype term, but it is a maturity test: it rewards organizations that know their processes, define metrics, and anchor control in the architecture – and punishes everyone else with the cancellation statistics. How we build agentic systems with RAG, MCP, and n8n, and which guardrails apply, is covered on our page on AI-assisted software development. For the technical backend perspective: Supabase as an AI backend.
