The fifteen or so words that decide whether what you are being sold will still run in a year
A buyer does not need to understand how any of this works. I build these systems and most of the vocabulary is irrelevant to the decision in front of you. What you need is a much shorter list: the words that come up in every sales conversation with every automation vendor, why each one matters to what you are buying, and what a real answer sounds like when you ask about it. This is that list, in the order a conversation actually goes. The definitions live in <a href="/glossary/">the glossary</a>, one click away from every term. This is the reading order and the reason to care.
Why the words matter more than the demo
A demo shows you the system on its best day, with the vendor driving. The vocabulary is how you find out what happens on an ordinary Tuesday, when the customer asks something the demo did not cover, when the CRM changes a field name, when the person who set it up has moved on.
Every term below is a place where a vague answer and a real answer sound almost the same to somebody hearing the word for the first time. Once you know what each word is doing in the sentence, the difference is obvious. That is the whole point of this post: not to make you technical, but to make the vague answer audible.
The words about what is being built
The first thing to settle is what kind of thing you are actually being sold, because three very different products get described with the same word. Most disappointment starts here, with a chatbot sold as an agent, or an agent sold to a business that needed a workflow.
- Workflow automation, chatbot, AI agent. These are three different things with three different failure modes and three different price shapes, and vendors use them interchangeably. Ask which one you are getting and why that one. If the answer is a pause, keep asking. Definitions: workflow automation, AI chatbot, AI agent. The longer version is in the guide on the difference.
- LLM. The model underneath. It matters far less than the vendor's pitch will suggest, and it matters in one specific way: which one, who runs it, and whether swapping it later would mean a rebuild. A vendor who cannot name the model or explain what happens when it changes has not thought about year two. Definition: LLM.
- Knowledge base and retrieval. The pair that decides whether the system answers as your business or as the average business in your sector. When a vendor says the system will "learn your business", ask what material it will read, who writes it, who keeps it current, and how a wrong answer gets traced back to the passage that caused it. Definitions: knowledge base, retrieval.
- Prompt engineering. Real, useful, and the least important of the four. If a proposal leans on it as the main deliverable, the vendor is selling instructions rather than a system. Instructions without material to answer from produce confident, generic, occasionally invented replies. Definition: prompt engineering.
The words about how it stays right
Anything that talks to customers will, at some point, meet a question it should not answer. The vendors worth hiring have thought about that moment before you ask. The others describe the system as if it never happens.
- Human in the loop. The question is not whether a person is involved but exactly where: reviewing everything, approving before it goes out, or stepping in only on exceptions. Ask the vendor to point at the place in the design where a person sits. "You can always override it" is not a place. Definition: human in the loop.
- Escalation rule. The written condition under which the system stops and gets a person, which person, on which channel, in what timeframe, including at 11pm on Saturday. This is the single term most often left implicit in a proposal, and in a regulated business it is a compliance control rather than a preference. Definition: escalation rule.
- Handoff. What travels with the customer when the system passes them to a person. A good one carries the whole conversation. A bad one is a notification with a name on it and a person starting from zero, which the customer experiences as being asked to repeat themselves. Ask to see one. Definition: handoff.
- Sandbox environment. A copy of the system, with test data, where changes are tried before they touch anything real. Its absence is how a small edit on a Friday afternoon becomes a weekend of wrong messages to real customers. Ask whether one exists and who has access to it. Definition: sandbox environment.
The words about what it plugs into
Almost every automation is a connection between things you already own. The words in this group decide whether that connection survives the first time one of those things changes, and whether anyone but the vendor understands it.
- API. Whether two systems can be connected at all, and on what terms. Ask the vendor which of your systems they will connect through a real API and which through a workaround, because workarounds are the part that breaks when the tool updates. Definition: API.
- Webhook. The difference between a system that reacts the moment something happens and one that checks every fifteen minutes. If the promise is "instant", this is the mechanism that makes it true. Definition: webhook.
- CRM integration. The connection that most projects live or die on. Ask which fields move, in which direction, what happens on a conflict, and what happens when your team adds a field next spring. Definition: CRM integration.
- System of record. When two systems disagree about a customer, which one wins. If nobody names it before the build, it gets argued about during an incident. A vendor who raises this before you do is thinking about your data rather than their demo. Definition: system of record.
- Integration debt. The accumulated cost of connections built quickly and understood by one person. It is not always wrong to take some on, but you should know you are taking it on. Ask what is documented, where, and what happens when the person who built it is not available. Definition: integration debt.
The words about your obligations
This group is the one buyers skip and regret. Every system that touches customer information inherits obligations that already applied to your business, and automating a process makes those obligations easier to meet or much easier to breach, depending on how it is built.
- Loi 25. Quebec's private-sector privacy law, and it applies to your business at every size. Any vendor working in Quebec should be able to say, unprompted, how the build handles consent, purpose and access requests. If they have never heard of it, that is your answer. Definition: Loi 25. Fuller treatment: the Law 25 and AI guide.
- Audit trail. The record of what was said to whom, what the system decided, and what a person approved. A well-built system produces a better trail than the manual process it replaces. Ask to see what one entry looks like. Definition: audit trail.
- Data retention. How long each kind of information is kept, decided per type rather than left at forever by default. Ask what the system keeps, for how long, and who can delete it. Definition: data retention.
- Data residency. Where the information physically lives, decided before anything is connected. Some of the tools in a typical stack are outside Canada by default, which is not automatically a problem, but it is a fact you should hear before you sign, not discover afterwards. Definition: data residency.
- Rate limiting. The cap that stops one script, or one runaway process, from consuming everything and running up a bill. Mostly a builder's concern, but if you are exposing anything to the public, ask whether it has one. Definition: rate limiting.
The words about whether it is working
The last group is the one that should have come up first, and rarely does. If the vendor has not asked which number this project is supposed to move, they are selling a system rather than an outcome. Pick the one that matches what you are buying and make it the measure.
- Speed to lead, and lead response time. Two names for the same clock, and the second is usually reported more generously than the first. Ask which window is being measured: business hours only, or the whole night the enquiry sat there. Definitions: speed to lead, lead response time.
- No-show rate. If you book appointments, this is the number a reminder and confirmation system exists to move. Ask what it is today, because most businesses do not know, and a vendor who has not asked has nothing to measure against. Definition: no-show rate.
- Review velocity. Not how many reviews you have but how steadily they arrive. The measure for any post-visit or post-job follow-up system, and one that exposes a system that ran for two weeks and stopped. Definition: review velocity.
What a vague answer sounds like, and what a real one sounds like
Five of the terms above, side by side. The vague answers are not lies. They are what somebody says when the decision has not been made yet, and the risk is that it will get made by default after you have signed.
| Term | A vague answer sounds like | A real answer sounds like |
|---|---|---|
| Escalation rule | "It knows when to bring in a human." | "If it cannot answer from the knowledge base, or the customer asks for a person, or the word complaint appears, it messages the on-call person on Slack within two minutes and tells the customer someone will reply by name." |
| Knowledge base | "It learns from your website and your documents." | "We will build it from your site, your intake notes and a session with whoever answers the phone. You will own the document. Here is who updates it and how often." |
| CRM integration | "It syncs with your CRM." | "New enquiries create a contact with these six fields, updates flow one way from the CRM back, and if a record already exists it updates rather than duplicates. Here is the list." |
| Data residency | "Everything is secure and compliant." | "The conversation logs live here, the model calls go there, and this is the one piece that leaves Canada. Here is what that means under Loi 25 and what we do about it." |
| Speed to lead | "Customers get an instant response." | "An acknowledgement that answers one thing and asks one thing, within a minute, at any hour. A person within thirty minutes during the day. We measure arrival to first human contact and report the median including weekends." |
The three words vendors overuse, and what to ask when you hear them
None of these words is dishonest. Each one has a real meaning that gets stretched because it sells. The fair response is not to flinch but to ask the question that turns the word back into something specific.
- Agentic. It genuinely describes systems that take steps toward a goal on their own, and some of the most useful things we build are exactly that. It also gets attached to anything with a model in it. Ask: which decisions does it make without a person, and which does it stop for? If the answer is "all of them", ask what the escalation rule is. If the answer is "none", you are buying a chatbot or a workflow, which may be exactly right, at a different price.
- Seamless. It means the connections are built and tested, which is a real thing worth paying for. It also gets used to skip the conversation about which connections exist. Ask: which of my systems, through which mechanism, and what happens when one of them changes? A vendor who answers with a list of APIs and webhooks means it. A vendor who repeats the word does not know yet.
- Enterprise-grade. Usually a shorthand for security and reliability, both of which matter and both of which are checkable. Ask: where does the data live, what is the audit trail, what is the retention policy, and is there a sandbox? Four specific answers make the word true. Four vague ones make it decoration.
Where this connects
Every term above has its full entry in the glossary, grouped by what the term is for rather than alphabetically, and written for someone who runs a business rather than someone who builds systems.
For the conversation itself, what to ask an automation vendor is the list of questions and the answers that should end the meeting, and reading an automation proposal covers what to look for once the words arrive on paper. If the first group of terms is where you are stuck, agents, chatbots and automations settles which one you actually need.
Where these words carry the most weight: professional services, where the escalation rule and the audit trail are the product; financial services and insurance, where residency and retention are not optional; and multi-location operators, where the system of record question decides whether ten sites report the same number.