Glossary
AI services terminology, in plain English.
Twenty-five terms you will hear from any managed AI services provider — and what they actually mean when stripped of marketing.
- AI agent
- A piece of software that takes a goal, uses a large language model and external tools to plan and execute work, and produces a finished artifact. Different from a chatbot in that an agent runs multi-step workflows — searches, reads, writes, calls APIs — and hands you a deliverable, not a conversation. See: AI agents vs virtual assistants.
- AI-native services company
- A services firm built around AI agents from day one, where humans focus on judgment and quality while agents do the volume work. Distinct from a traditional services company that started using AI tools later. See the long form: What is an AI-native services company.
- Managed AI services
- A subscription model where the provider operates the AI agents on your behalf, supervises them with senior humans, and delivers finished work — instead of selling you software you have to run yourself. See: Build vs managed AI agents.
- Operator gate
- The design pattern where every AI-generated artifact passes through a senior human reviewer before it reaches the client. The agent carries the load; the operator carries the accountability. Without an operator gate, you are buying the AI itself, not a service.
- Agent runtime
- The infrastructure that runs AI agents in production: model routing, tool integrations, memory, evaluation hooks, logging, error recovery. The plumbing under the agent that turns a model into a worker. Most companies underestimate how much of an AI service is the runtime versus the model.
- Productized services
- Services packaged with fixed scope, fixed price, fixed delivery cadence — sold like software rather than billed by the hour. Removes scope creep and gives buyers predictable spend. See: Productized services vs custom agency.
- Evaluation layer (evals)
- Automated checks that verify an agent's output against criteria — facts, schema, brand voice, safety — before the artifact reaches a human reviewer. Catches most errors at zero marginal cost. See: AI agent evaluations explained.
- Multi-model routing
- An agent system that picks between different LLMs (Claude, GPT, Gemini) per task based on which model wins for that workload on quality and cost. Avoids single-vendor lock-in. Worth asking any AI services provider if they do this or are locked to one model.
- Senior operator
- A named human with deep domain expertise who owns the quality of every artifact an AI agent produces for a specific client. Reviews, edits, signs, and is reachable for escalation. "Named" matters — generic "our team" answers should be a red flag.
- Replayable runs
- Agent executions stored with full state — prompts, tool calls, intermediate outputs — so any run can be re-played later for debugging, audit, or improvement. Foundational to AI service transparency. If a provider cannot replay a run from three weeks ago, their logs are too thin.
- EU data residency
- A commitment that client data is stored and processed inside the European Union, never moved to non-EU jurisdictions without explicit consent. Required by some GDPR-compliant procurement workflows. See: AI agents and EU data residency.
- Tenant isolation
- Architecture where each client's data, agents, and runtime state are kept separate from every other client's. Prevents data leakage between accounts even within the same provider. Standard for enterprise SaaS; should be standard for managed AI services too.
- DPA (Data Processing Agreement)
- A legal agreement under GDPR Article 28 that defines how a vendor processes personal data on behalf of a client — scope, sub-processors, security measures, retention, deletion. See: AI agents GDPR compliance.
- Zero-training agreement
- A contractual commitment that your data will not be used to train the AI provider's foundation models. Standard with paid Claude, GPT, Gemini API tiers; less common with consumer-grade tools. Worth getting in writing for any sensitive workload.
- Client portal
- In the context of managed AI services: a live interface where the client can see the plan, agent logs, every artifact, costs, and operator approvals — in real time, not in a monthly status PDF. The honest version of "transparency" in AI services marketing.
- Weekly artifact
- The default unit of delivery in productized AI services. Something concrete — a published article, a reconciled month, a working code release — handed over once a week, signed by an operator. Forces tight scope and steady output.
- Sprint (services context)
- A fixed-time cycle (typically one or two weeks) during which a specific deliverable is scoped, executed, reviewed, and shipped. Borrowed from software engineering and adapted to services delivery. Distinct from an agile sprint in that the output is a finished artifact, not a code increment.
- Audit log
- An immutable record of every action taken inside an AI agent system: prompts, tools used, outputs, costs, operator reviews. Used for debugging, compliance, and forensic analysis. See: AI agents security checklist.
- ChatGPT alternative
- Often misused. ChatGPT is a tool; a "ChatGPT alternative" can mean either another raw LLM interface (Claude.ai, Gemini), or a managed AI service that finishes the job for you instead of just generating text. See: ChatGPT alternative for business.
- AI services company
- Any firm that delivers business outcomes using AI as the primary engine. Spectrum from "consulting with ChatGPT" (low end) to fully agent-operated delivery with operator review (AI-native end). See: How to choose an AI agents services company.
- B2B AI services
- AI-powered services aimed at businesses rather than consumers. Typically structured work — content, ops, finance, research, code — sold on subscription or project basis to teams of 5-500 people.
- AI-managed PPC
- Pay-per-click campaigns (Google Ads, Meta Ads) operated by AI agents under senior human supervision rather than by a traditional agency. Faster cadence, fixed fee instead of percent-of-spend. See: AI-managed PPC vs Google Ads agency.
- Outcome-based pricing
- A contract where payment is tied to achievement of a defined result (lead generated, article published, ROAS hit) rather than to hours worked. Aligns incentives but requires very tight scope definition. Rare in services because it transfers all the risk to the vendor.
- Definition of done (DoD)
- A written specification of what counts as "finished" for a deliverable — quality criteria, format, sign-off requirements. Critical for productized services to avoid scope drift. If a provider cannot give you a DoD in writing, they have not productized.
- Human-in-the-loop (HITL)
- Any AI workflow that requires human approval, review, or input at one or more steps. The opposite of fully autonomous AI. In B2B AI services, HITL at the gate is currently the only configuration safe enough to ship to clients.
Term not on the list?
Email us at info@logitelia.com — if it is genuinely a useful addition, we will define it here and link from your email.
info@logitelia.com