Glossary · applied AI
AI agent, forward-deployed engineer, MCP, evals, agent washing, the EU AI Act and more. Each of the 53 terms is defined in a few sentences, from the point of view of a decision-maker who wants to understand what they are buying.
In brief
Djtal, the operational AI specialist for Swiss companies and an Authorized Zoho Partner since 2017, defines 53 business AI terms in seven families: fundamentals, agents and agentic AI, ERP/CRM and process integration, deployment and forward-deployed engineering, governance, risk and compliance, measurement and operation, models and ecosystem. Each definition gives a Swiss executive the essentials before investing.
The definitions
The building blocks every AI project works with.
An LLM (large language model) is an AI model trained on vast bodies of text to understand and generate language. Claude, ChatGPT and Gemini are LLMs. In a business, an LLM is the component that reads, writes and reasons. It becomes useful once it is connected to your data and your tools.
Generative AI is AI that produces content (text, images, code) from an instruction. It has put AI in everyone's hands. Producing text and running a process remain two different jobs, however. The second step, the AI that acts inside your systems, is agentic AI, and most of the business value is won or lost at that step.
A prompt is the instruction given to an AI model to get a result. The quality of the prompt (context, constraints, expected format) determines the quality of the answer. In a business, effective prompts are written, tested and version-controlled, like code.
A token is the unit into which AI models split text, and the unit they bill by. A single word in French often counts as two or three. Every call is paid for in tokens read and written. Understanding tokens means understanding your bill, and why a badly designed agent that re-reads everything at each step costs several times what it should.
A context window is the amount of information an AI model can hold in mind at one time: your documents, the current conversation, its instructions. It is counted in tokens and limits what an agent can process in one pass. A serious AI project is sized around that constraint: which documents go in, which are summarised, which are archived.
A reasoning model is an AI model that works through internal steps before it answers. On complex cases it decides better, at a higher cost and with a longer response time. The trade-off is made task by task. Sorting an email does not need the same engine as analysing a contract.
RAG (retrieval-augmented generation) is a technique that connects an LLM to your own documents. Before answering, the model retrieves the relevant passages from your knowledge base and draws on them. RAG grounds answers in your real data and reduces fabricated answers, without retraining the model. It is often the economical alternative to fine-tuning.
Fine-tuning is the further training of an existing AI model on domain-specific data, to specialise it. Often RAG is enough, and it costs less to ground a model in your data. Fine-tuning is justified when tone, format or very specific knowledge demand it.
A hallucination is a false AI answer, stated with confidence and invented by the model for lack of reliable information. Hallucinations are the central risk of business AI projects. They are reduced by RAG, verifiable sources and human oversight. Some always remain, which is why guardrails exist.
The difference between an assistant, a copilot and an agent decides what you buy.
An AI agent is software that carries out a task from start to finish on its own. It consults data, decides on an action and executes it in your tools (ERP, CRM, email), under human oversight.
AI agents for your business →Agentic AI is the generation of AI able to take real actions in live systems (ERP, CRM, email), beyond producing text. An agentic system chains perception, decision and action within a defined scope.
Agentic AI explained →An AI assistant is a tool that handles your one-off requests: it drafts, summarises, searches, then waits for the next one. It augments the person who uses it, without running a process itself. It is the level below the copilot and the agent, and the safest starting point for a team discovering AI.
An AI copilot is a tool that assists a professional inside the software they work in: it suggests the reply to a ticket, the accounting entry, the line of code. The person keeps control of every approval. Between the assistant (which handles one-off requests) and the agent (which runs a process), the copilot is the intermediate level that most providers sell.
A chatbot is a program that holds a conversation following predefined scripts, historically rule-based and today backed by language models. It stays confined to dialogue. It informs, without handling the case. Confusing a chatbot with an AI agent at the point of purchase means paying an agent's price for a question-and-answer desk.
Agent orchestration is the coordination of several specialised AI agents that share out a piece of work, each within its own scope. A team of agents (sales, finance, support) collaborates under supervision, where a single generalist agent soon reaches its limits. Djtal runs nine in-house, and they pass information to one another through a shared state.
A multi-agent system is a setup in which several AI agents cooperate on the same job, each within its own scope. It is powerful when the subtasks are independent and counterproductive when adopted to follow a trend, because every agent added brings more coordination, more cost and more points of failure.
The level of autonomy is what an AI agent is allowed to do on its own: read, propose, execute with approval, execute and report back. The design rule that protects you is to grant the minimum level of autonomy that gets the work done (‘least agency’), then widen it on evidence.
Human oversight is the practice of placing a human approval at the sensitive steps of an AI agent's work. At Djtal, this checkpoint is set down in black and white when each agent is scoped. The arrangement, known as human-in-the-loop, reserves decisions with real stakes (a message to a client, an accounting entry) for a person and automates the rest.
Human-on-the-loop is a level of oversight in which a person monitors the agent's activity and can step in, without approving each action. It sits one level above human-in-the-loop and is reserved for well-established processes where the agent's track record justifies trust, with an audit log and an emergency stop in place.
Tool calling is the mechanism by which an AI model acts. It invokes a specific function (create an invoice, update a record, send an email) with controlled parameters. This is how an agent acts in your ERP or your CRM, with rights limited tool by tool.
Guardrails are the technical limits set on an AI agent before it goes live: lists of permitted actions, spending limits, data boundaries, approval points. They are defined at scoping, written into the configuration and tested like everything else.
Agent washing is the relabelling of a chatbot or a conventional automation as an ‘AI agent’. Gartner estimates that about 130 providers offer real agentic AI, out of thousands that claim it. Three questions tell them apart. Does the tool act in your systems? Does it decide according to context? Does it work without being prompted at every step?
Where AI meets your existing systems.
An ERP (enterprise resource planning) system is the software that centralises a company's management processes: finance, sales, purchasing, stock, HR. Zoho, Odoo, Microsoft Dynamics and SAP are ERPs. Djtal applies AI to a company's existing ERP, whichever it is, with no change of system.
The complete ERP with Djtal →A CRM (customer relationship management) system is the software that brings together a company's contacts, opportunities and customer interactions. Zoho CRM, Salesforce and HubSpot are examples. Connecting an AI agent to the existing CRM (qualifying a contact, following up, keeping records up to date) is one of the most profitable uses of AI in business.
CRM and AI →MCP (Model Context Protocol) is an open standard that lets an AI model connect, in a controlled way, to external tools and data: CRM, files, business APIs. MCP standardises how an agent reaches your systems. One convention covers all your tools, where each bespoke integration used to need its own maintenance.
A2A (Agent2Agent) is an open protocol that lets AI agents talk to one another (delegate tasks, pass on results), including across different providers. It complements MCP. MCP connects an agent to your tools, and A2A connects agents to each other. Together they form the standard plumbing of agentic systems.
Zia is the native AI built into the Zoho ecosystem: predictions, anomaly detection, OCR, language processing and a conversational assistant, directly in the CRM and the other Zoho applications. An Authorized Zoho Partner since 2017, Djtal activates it where it pays off most, starting with prospect scoring and call transcription. Zia saves you from assembling a separate AI infrastructure.
Our Zoho expertise →ERP-native AI is the AI that the vendor builds into its own platform: Zia at Zoho, Copilot at Microsoft, Joule at SAP, Einstein at Salesforce. It is often already included in your licence. The rational approach starts there. Switch on what you already pay for, measure what is missing, then add dedicated agents.
Process automation is the use of software to carry out repetitive tasks that used to be manual: data entry, reminders, reconciliations, notifications. Combined with AI, it also handles cases that call for judgement, such as reading a document and sorting it. Administration is the first target for automation in most companies.
RPA (robotic process automation) is automation that works by replaying recorded actions: click here, copy there. Effective on stable screens, it breaks at the slightest variation, such as a field that has moved or a label that has changed. It is the automation of the days before language models. It keeps its place where everything is fixed and leaves the rest to agentic automation.
Agentic process automation (APA) is handing a process over to an agent that reads, judges and acts according to context, where RPA replays a script and intelligent process automation (IPA) handles semi-structured cases. The RPA, IPA and APA framework helps you choose. The more a process varies, the more judgement and oversight it needs.
Intelligent document processing (IDP) is the automated reading of incoming invoices, contracts and letters, with their data extracted and posted into your systems. It is the most common entry-level use case in business: high volume, clear rules, a measurable gain within the first weeks. OCR is its ancestor. Current models also read free-form documents.
Intelligent document management →Taking AI from the demonstration to a system that runs is Djtal's trade.
A forward-deployed engineer (FDE) is an engineer sent to the client to build the AI solution inside the real environment of its systems and teams, until the client can run it alone. The model comes from Palantir. AWS, Microsoft and OpenAI have been investing heavily in it since 2025. Djtal practises it in French-speaking Switzerland, with engineers embedded at the client's site.
Forward-deployed engineering at Djtal →Proof of concept (POC), pilot and production are the three steps of an AI project. The POC proves feasibility, the pilot tests under real conditions and production runs every day without assistance. According to MIT, 95% of generative AI pilots deliver no measurable return, and the last step is the one that kills them. Pilots clear it through evals, guardrails and operation, far more than through the demo.
Why AI projects fail →Build versus buy is the choice between building your AI solution and buying it. MIT found that a purchased tool reaches production about twice as often as an in-house build. The third way, often the best for a mid-sized Swiss company, is to strengthen what you already have, by connecting agents to the ERP and CRM already in place.
Customer Zero is the practice of a company being its own first customer. Djtal runs its own sales, finance and communications with named AI agents. On 3 July 2026, its communications agent published a LinkedIn post on the company page, end to end, on its own.
The AI-native company →An agentic harness is the tooling that surrounds an agent in production: evaluation sets, guardrails, logs, recovery procedures. It is the durable part of the system. When the model or the agent changes, the harness is kept and reused from one process to the next.
What makes AI defensible in front of a board, an auditor or a regulator.
The Swiss Federal Act on Data Protection (FADP) is the law that governs the processing of personal data in Switzerland, alongside the European GDPR. It is Djtal's first filter before connecting an AI to a client process. Which personal data does the agent touch, on what basis, and with what protective measures?
Security and compliance →Data sovereignty is control over where, and under which law, your data is processed and stored. For a Swiss company the question arises with every AI component: where do the prompts go, where do the documents live, which legislation applies in a dispute. It is dealt with through architecture (chosen hosting, mapped transfers), well before the contract clauses.
The EU AI Act is the European regulation on artificial intelligence. It classifies systems by level of risk and imposes obligations that tighten with the level of risk, up to an outright ban. A Swiss company is affected as soon as its system or its clients touch the EU market. The timetable for application is still changing, so have your specific case checked before you invest. The dates circulating online go out of date quickly.
A high-risk AI system, within the meaning of the EU AI Act, is one that touches sensitive areas: recruitment, credit, education, infrastructure, justice. It triggers the heaviest obligations, which are documentation, human oversight, logs and risk management. The first question for a project is whether your use case falls into one of these categories. The answer changes the budget and the timetable.
AI governance is the framework that answers four questions: which agents and tools are authorised, on which data, with which rights and under what oversight. It fits in a few practical pages, made up of a register of uses, approval points, an audit log and a periodic review.
Shadow AI is the use of AI by your employees outside any framework: personal accounts, client data pasted into a consumer tool, results reused unchecked. The risk is already inside the walls of most companies. The response that works combines a short written framework, approved tools and training. A ban on its own only moves the usage elsewhere.
Traceability, for an AI agent, is the requirement that every action leaves a replayable trace of what the agent read, decided and executed, and why. The audit trail is the record an auditor, a regulator and you yourself will ask for on the day a result surprises you.
A kill switch is the control that stops an AI agent immediately. It is a planned, tested switch, within reach of the right people, and it is designed before the agent goes live. Knowing that you can stop the agent cleanly changes the deployment, since you are willing to entrust more to an agent once taking back control is assured.
An agent with no measurement is judged by anecdote.
Evals are the tests of an AI system: a set of real cases with the expected results, replayed at each change of model, prompt or scope. They are the equivalent of software tests, applied to non-deterministic outputs. Without evals, every model update is made without a safety net.
Agent observability is the ability to see what your AI agents do: actions taken, cost per run, cases referred to a person, failures. It starts free of charge, with logs, a simple dashboard and a weekly review, well before any specialist tool.
Cost per task is the cost of one run of an AI agent (tokens, calls, licences, supervision included), set against the cost of the same work done by hand. It is the question to put to any AI agent provider. An agent that looks profitable in a demonstration can turn out to run at a loss at real volume. The calculation is made before going live, then monitored continuously.
AI ROI is the return on investment from AI, measured against indicators defined before the project: hours given back to teams, shorter lead times, errors avoided, AI-assisted revenue. The MIT finding on pilots with no measurable return (see Proof of concept, pilot, production) stems first from this: no indicator was set at the start.
The AI strategy audit →The escalation rate is the share of cases that an AI agent refers to a person. It is the most telling KPI of an agent in production. Too high, and the agent adds nothing. At zero, it is probably deciding things it should refer. A rate that climbs week after week signals a drift to investigate.
The names that come up in every AI discussion, and how to choose between them.
A frontier model is one of the most capable models available at any given time: Claude (Anthropic), GPT (OpenAI), Gemini (Google). The ranking shifts several times a year, which makes ‘the best model’ an architecture decision to review regularly. A well-designed project can change model without being rewritten.
An open model (open weights) is a model you can download and run on the infrastructure of your choice: Llama, Mistral, DeepSeek. Its strengths are control over where processing happens and costs at volume. Its trade-offs are that you take on its operation and that, depending on the task, its performance lags behind frontier models. It is justified case by case, often for sovereignty reasons.
GEO (generative engine optimisation) is the practice of getting your company cited by AI engines (ChatGPT, Perplexity, Gemini) when a buyer asks them for a recommendation. It complements SEO. Your Google ranking still matters, and the citation inside the answer becomes the unit that counts. Djtal applies it to its own website and measures the effects, query by query.
Further reading
These terms take on their meaning in real projects. Three ways in to continue.
Get started
Tell us which one, and which process you picture it in. We will reply in the same vocabulary as this page, with no obligation.
Last updated: