ArkangeAccording to McKinsey State of AI 2025 (1,491 respondents, 101 countries), 71% of companies use AI but only 1% reach operational maturity. The main cause is a taxonomy failure. Most leaders confuse “AI agent” and “chatbot”, launch a conversational assistant on every topic, and watch six months later as usage falls back to zero.
This article maps the three main families of AI agents that can be deployed in a company, their operational decision criteria, and the recommendation adapted to each context.
Why talk about three families of AI agents rather than a “universal chatbot”?
An AI agent is defined by its trigger far more than by its interface. A conversational agent waits to be talked to. A scheduled agent wakes up on a calendar. An event-driven agent reacts to a business signal (incoming email, uploaded document, new ticket). Confusing these three families means committing to spend on AI nobody uses.
The Stanford Digital Economy Lab study from April 2026 (Brynjolfsson et al., “The Enterprise AI Playbook. Lessons from 51 Successful Deployments”) shows that high-ROI deployments share one thing in common. A clearly identified trigger and a tight business scope. Not a single mention of a “generalist enterprise chatbot” among the 51 cases studied.
In 2025, about two thirds of projects started in “a chatbot for everyone” mode were abandoned within 90 days. The projects that survive and scale start with a specific family, on a specific friction.
What is a conversational AI agent and when to choose it?
A conversational AI agent is an agent whose trigger is a human question, formulated in natural language, in a chat interface. It answers, reasons, sometimes executes an action, always in reaction to an intention expressed by a user.
The typical use case of the conversational agent
The conversational agent makes sense when the user knows what to ask but does not know where to look. A typical case is the conversational assistant that answers trainers’ questions about Qualiopi frameworks, OPCO funding rules and regulatory requirements. The trigger is always a question. The user keeps the initiative.
Other fitting contexts.
- Internal HR support (questions on the collective agreement, leave, expense reports)
- Document search in a technical knowledge base
- Sales assistance in pre-sales on a complex product catalog
The limits of the conversational agent
The conversational agent suffers from a structural problem. If nobody remembers it exists, nobody uses it. McKinsey State of AI 2025 documents an average usage rate of 18% on internal chatbots 6 months after deployment. A generalist conversational agent reachable from the intranet regularly ends up with a handful of active users three months later. When the same logic is switched to event-driven mode (triggered from the mailbox when a client message arrives with an attachment), adoption takes off on the same target within weeks.
What is a scheduled AI agent and when to use it?
A scheduled AI agent is an agent whose trigger is temporal. Every day at 7 a.m., every Monday morning, every month end. It runs without human intervention, produces a structured deliverable and pushes it to the relevant recipients. No questions, no chat interface.
The typical use case of the scheduled agent
The scheduled agent is formidable for repetitive, high-value business rituals. A typical case is the agent that generates every Monday at 6 a.m. a synthesis of the files due within the week, per employee, with a priority level. The employee opens their screen and finds their day already framed. No question asked to the AI, no prompt to write.
Other fitting contexts.
- A daily sector press review
- A weekly sales report
- Monthly compliance alerts
- Committee summaries prepared the evening before at 10 p.m.
The limits of the scheduled agent
A scheduled agent becomes useless as soon as the input data is no longer available at the planned time. A recurring case is the logistics synthesis scheduled at 5 a.m. that fails two days out of five because the ERP has not finished its overnight batch. The workaround is to switch to event-driven logic (trigger when the ERP signals end of processing). Temporal planning assumes a stable infrastructure. Without it, the organization industrializes a false positive.
The a16z Enterprise CIO Survey 2025 (100 CIOs surveyed) notes that scheduled agents account for 34% of deployments with documented ROI, the second category after event-driven.
What is an event-driven AI agent and for which processes?
An event-driven AI agent is an agent whose trigger is a business event captured in real time. An email received, a document uploaded, a ticket opened, a status changed in the CRM, an electronic signature validated. The agent reacts, processes and returns the result inside the existing flow. The user receives work already done.
The typical use case of the event-driven agent
The event-driven agent excels on high-volume processes with an identifiable trigger. A typical case is the agent that fires on every document upload in a DMS, classifies the document, extracts the structured data, pre-fills the production tool and notifies the employee when there is an anomaly. The observable order of magnitude for this kind of usage runs around 10 to 15 minutes saved per document, on volumes of several thousand events per month.
Other examples.
- Automatic qualification of an incoming lead in a CRM
- Compliance check of a contract as soon as it lands in an e-signature tool
- Generation of a standard reply when a support ticket arrives with a recurring reason
On the public sector side, an event-driven agent can process incoming mail requests and propose a draft reply to the case handler before the file is opened.
The limits of the event-driven agent
The event-driven agent requires clean integration with the source systems (email, DMS, CRM, ERP). It is the most expensive family to set up initially and the most ROI-positive at 100 days. A 10/30/100 rhythm remains a good reference. A properly scoped event-driven agent usually delivers its first useful processing in 10 days, reaches stable usage at 30 days and measurable ROI at 100 days. If the source system’s API is unstable or undocumented, the project derails.
Stanford Digital Economy Lab (April 2026) counts 27 of the 51 high-ROI deployments studied as event-driven agents. It is the family that produces the most value in a mature company.
Which criterion to use to choose between conversational, scheduled and event-driven?
The main criterion is the nature of the business trigger, before the technology or the budget. Three questions are enough to decide in 80% of cases.
Criterion 1. Who takes the initiative?
If the user knows they have a need and formulates a question, it is conversational. If the clock takes the initiative at a fixed time, it is scheduled. If a business event (email, document, status) takes the initiative, it is event-driven. This question alone eliminates 50% of false leads.
Criterion 2. What is the frequency of the need?
- A sporadic, unpredictable need calls for a conversational agent
- A regular, calendar-based need calls for a scheduled agent
- A need triggered by a steady event flow (10+ per day) calls for an event-driven agent
Frequency also determines ROI. An event-driven agent on 3,200 monthly events does not carry the same return as a conversational agent used 12 times a month.
Criterion 3. What is the organization’s current AI maturity?
On an AI maturity grid of 5 levels (from N1 Curious to N5 Scaler), N1-N2 organizations are better off starting with a targeted conversational agent (low integration cost, fast learning). N3-N4 organizations should prioritize event-driven agents (demonstrable ROI, possible scaling). N5 organizations orchestrate all three families in parallel inside a coherent agentic architecture.
In a mature industrial organization, the three families can coexist. Conversational for technical document search, scheduled for weekly quality reviews, event-driven for supplier non-conformance processing. This kind of orchestration assumes an N4-N5 level.
Which family to deploy first depending on your context?
Start with the family that matches the most painful business trigger in your organization.
If you are an SMB or mid-cap at N1-N2 on the maturity grid
Start with a conversational agent targeted on a known documentation friction. HR assistance on the collective agreement, search in the technical knowledge base, pre-sales commercial support. Low cost, fast learning curve, strong political signal. Avoid the “chatbot for everyone” trap. Target one business function, one typical question, one defined audience.
If you are at N3 with already 1-2 agents in production
Move to event-driven on your highest-volume repetitive process. That is where ROI becomes the most readable. On quality document management in an industrial environment, this switch regularly produces savings of about 1 FTE per processed scope. The conversational to event-driven shift is the most profitable maturity jump.
If you run business rituals with strong regularity
The scheduled agent is your simplest entry point. Reviews, reports, compliance alerts. On weekly regulatory summaries in a banking or financial environment, a scheduled agent produces its first deliverable in under 10 days, because there is no real-time integration to build.
The thesis on the three families of AI agents
The most expensive mistake in 2026 is choosing the wrong agent family. 90% of French companies use AI without a governance framework (Usine Digitale, 2025). The result is predictable, chatbots deployed with great noise then abandoned in silence.
The conversational / scheduled / event-driven taxonomy works as an investment filter. The choice of the family is made on the business trigger and on the organization’s maturity. Naming correctly what you deploy is the first step of a durable adoption.
