Plain English
Every term this site uses, defined.
Thirty-five terms, grouped by what you are trying to understand rather than alphabetically, because most people reading a glossary are following an argument rather than looking up a single word. If we use a word anywhere on this site and it is not explained here, that is a gap and we would like to know.
AI and models
The vocabulary of the underlying technology, without the marketing.
- Agentic AI
- AI that can plan a sequence of steps and take actions toward a goal, rather than answering one question at a time. AutoServa is agentic: it reads a case, decides what to check next, acts, and stops when it hits a boundary, without a person prompting each step.
- Large language model
- The kind of AI model underneath the operator. It reads and produces language, which is why it can work with contracts, invoices, tickets and emails rather than only with structured database fields.
- Frontier model
- The most capable model currently available from a given provider, for example GPT-5 from OpenAI or Claude Opus 5 from Anthropic. The frontier moves every few months, which is why AutoServa is built so the model can be swapped.
- Open-weight model
- A model whose parameters are published, so it can be downloaded and run on hardware you own instead of being reached through a provider's cloud. This is what makes a fully local deployment possible.
- Fine-tuning
- Further training a general model on a specific, narrower body of material, in this case your documents, decisions and vocabulary, so its answers reflect your organisation rather than a general internet average.
- Context window
- How much material a model can consider at once. It matters because a long contract or a case with a lot of history has to fit inside it for the model to reason across the whole thing.
- Inference
- The act of the model actually producing an answer for one case. It is the per-case running cost of a deployment, which is why case volume affects the ongoing bill.
- Hallucination
- When a model states something false with the same confidence as something true. Fine-tuning on your own material and requiring evidence for every claim is how AutoServa keeps this from reaching a live decision unchecked.
- Latency
- How long the answer takes to arrive. A real selection criterion: an overnight reconciliation can use a slower, more thorough model, a first response on a service desk cannot.
How AutoServa works
The concepts specific to this service. These are the words you will hear most in a working session.
- Operator
- The tuned agentic AI that works one of your queues. Named that way deliberately: it occupies the position a member of operations staff occupies, between your business and its systems.
- Decision framework
- A written structure describing the inputs a decision depends on, the thresholds that change the answer, the constraints that cannot be crossed, who approves what, and when a case must escalate. It is the central deliverable of a decision audit.
- Threshold
- The specific value at which an answer changes: a tolerance percentage, a spending limit, an age of debt, a risk score. Most organisations have these, and most have never written all of them down in one place.
- Escalation rule
- The explicit statement of when a case stops being routine and must go to a person. Defined per workflow during Codify, not inferred by the model.
- Precedent
- What was decided last time on a similar case, kept searchable, so the hundredth decision is consistent with the first rather than depending on who happens to be handling it.
- Standards
- Your templates, standard terms, accounting treatments, service levels and tolerances. What makes an output recognisably yours rather than a plausible generic answer.
- Vocabulary
- Your entity names, cost centres, product codes, project references and internal shorthand. A general model guesses at these; a tuned one reads your documents the way your staff read them.
- Queue
- The specific stream of repeating work an engagement targets. AutoServa is deployed onto one queue at a time, deliberately.
- Shadow running
- A period where the operator works the live queue and produces answers, but nothing it produces is acted on. Used to measure accuracy against real cases before it is trusted with anything.
- Disagreement log
- The record of every case where the operator's answer differed from the human's. Reviewed weekly rather than silently resolved, because the disagreements are where the framework is still wrong.
Governance and control
The vocabulary your compliance and risk colleagues will use.
- Workflow Guard
- AutoServa's governance layer. It checks every action the operator would take against approval chains, spending ceilings, access scope, segregation of duties, retention rules and regulatory obligations, before that action executes.
- Evidence trail
- The record attached to every decision: what was read, which rule in the framework applied, what threshold was hit, what precedent it followed, and who approved the release. It is what makes the operator auditable rather than a black box.
- Human in the loop
- A design principle where a person must review or approve before an action is final. AutoServa keeps a human in the loop on anything irreversible; routine, in-policy steps proceed without a fresh approval each time.
- Segregation of duties
- A control keeping the party who prepares a transaction separate from the party who approves it, so no single actor can complete a sensitive action alone. Workflow Guard applies it to the operator too.
- Least privilege
- Granting only the access a role actually needs. The operator starts read-only and gains write access to specific systems only once it has been tested against your own history.
- Approval chain
- The defined sequence of who must sign off on what, and above which threshold. The operator prepares a case and stops at the same point a junior member of staff would stop.
- Data residency
- A requirement that data be stored and processed in a particular country or region. Where a regulator requires it and in-region cloud hosting is not sufficient, a local deployment is the answer.
- Air-gapped
- A system with no network connection to the outside internet. The strictest form of local deployment, used where policy requires that no data can leave the premises under any circumstance.
- Blast radius
- How far the consequences of a mistake reach. Because the evidence trail records every action, the blast radius of an incident can be established from the record rather than estimated.
Engagement and delivery
What the commercial and project vocabulary actually means here.
- Decision audit
- The first stage of an engagement: two to three weeks mapping how one workflow really decides, ending in a written decision framework that is yours to keep regardless of what happens next.
- Codify
- The stage where captured knowledge becomes an explicit decision framework. The point at which tacit judgement becomes a document your organisation owns.
- Process owner
- The person accountable for a queue, who arbitrates when the team disagrees about how a case should be decided and signs off the framework.
- Supervised live run
- The operator working real cases with a person holding release on the output. The stage between shadow running and any autonomy, and the stage most deployments spend the longest in.
- Baseline
- How the queue performed before the engagement started, agreed during the audit. Results are measured against it rather than against an industry benchmark.
- Key-person risk
- The exposure created when one individual holds knowledge nobody else has. Consolidating that judgement into a framework the business owns is the main thing an engagement is for.
- Model migration
- Moving a live deployment onto a newer or better-suited model. Because the framework, precedent and evidence trail sit outside the model, this does not mean rebuilding the deployment.
Still a word doing too much work?
Vendors in this market lean hard on vocabulary, and a term that sounds precise is often hiding the interesting question. Ask us what one of them actually means for your queue and we will answer without the jargon.