Generative, predictive, agentic: the three kinds of AI in plain English
When a quoting tool says it "uses AI", it could mean three quite different things. Here is how to tell them apart, where each one fits in the quoting process, and what to watch for.
"AI" has become a single word for several different technologies. One reads and writes text. Another estimates what is likely from your historical data. A third carries out tasks across your systems with limited supervision. In quoting, all three have a job to do: reading the buyer's RFQ and drafting the answers, judging whether an enquiry is worth the effort, and chasing supplier prices and buyer replies. They behave differently, carry different risks and need different controls.
Knowing which is which makes it much easier to judge any quoting tool, including ours. This explainer takes the three types in turn, using a quote as the running example. Agentic comes last, because it is built on top of generative and is easier to understand once that is in place.
Can you write or summarise this?ProducesText, summaries, answers from documentsIn a quoteReading the RFQ and drafting the answers
What is likely to happen next?ProducesScores, estimates, classifications, alertsIn a quoteJudging how well an enquiry fits
Can you go and get this done?ProducesCompleted tasks and actions in your systemsIn a quoteChasing prices and replies, filling in return files
The quickest way to tell them apart is by the question each one answers, and the part of a quote it handles.
Generative AI: it reads and writes
Generative AI is the type most people have now used. ChatGPT, Claude, Gemini and Copilot are all generative: you give them a request in ordinary language and they produce new content in return.
The important thing to understand is how it works. A generative model has been trained on a very large amount of text and has learned the patterns in it. When you ask a question, it does not look up the answer in a database. It produces the most plausible response given everything it has seen. That is why it is so fluent, and also why it can be fluently wrong.
Where it turns up in quoting
- Reading a sixty-page RFQ, specification and drawing pack, and pulling out the requirements, deadlines and compliance clauses an engineer needs to see.
- Drafting the technical answers to a buyer's questionnaire from datasheets, test reports and past quotes.
- Answering an engineer's question, such as "have we supplied this material grade before?", from your own records, with a link to the source.
- Turning a vague enquiry email into a structured quote request, and drafting the clarification questions to send back.
- Writing the covering letter and the scope and exclusions section in your house style.
The pattern that works in quoting: the buyer's documents and your own knowledge in, a draft out, an engineer in the loop. Grounding the model on your own documents narrows what it can invent. The approval at the end is not optional.
Generative AI does most of the visible work. Parker reads the buyer's RFQ, specification and drawings and turns them into a structured list of requirements. It drafts technical answers from your Knowledge Hub, with a link to the source document behind each one, and writes clarification emails to the buyer.
Nothing goes out on Parker's word alone. An engineer reviews every answer, confirms or edits it, and the approved version is what the buyer sees.
What to watch
Generative models fill gaps with plausible-sounding invention, usually called hallucination. In a draft that is an inconvenience. In a submitted quote it is a commercial risk: an invented compliance claim, a test standard you do not hold or a material you do not stock becomes a commitment once the buyer accepts it. The practical fixes are to point the model at your own documents so it answers from them rather than from memory, to show where each answer came from, and to keep an engineer responsible for what is sent.
The other thing to watch is where the buyer's data goes. RFQs often contain drawings and specifications shared under NDA. Pasting them into a consumer AI account may mean they are used to train future models, which few buyers would accept. Business tools, properly configured, should not do this.
Predictive AI: it estimates what is likely
Predictive AI is the oldest of the three, and it has quietly run parts of business for decades. Credit scoring, demand forecasting and the spam filter on your email are all predictive. The idea is simple: take historical data, find the patterns in it, and use those patterns to estimate what is likely to happen next.
Unlike a generative model, a predictive model does not produce prose. It produces a score, a probability or a category, and it is usually built to answer one narrow question. In quoting, the useful questions are narrow ones: how closely does this enquiry match work we have won before, and is there anything about it that has caused trouble on similar jobs?
Where it turns up in quoting
- Assessing how well a new enquiry fits your products, capabilities and the jobs you have won before, so the team can decide early whether to bid.
- Recognising enquiries that resemble past losses: the same kind of buyer or product, lost on price or lead time.
- Flagging specifications that tend to lead to scope changes or disputes later, based on what happened on similar jobs.
- Highlighting a quote that looks unusually high or low against similar past quotes, so someone takes a second look before it goes.
Predictive work in Mainspring Quote is narrow and deliberate. When an enquiry arrives, Parker's initial read assesses how well it fits your products and past work, and flags the gaps and risks worth a look before you commit time to it. It draws on your quote history, including which bids were won and lost, so the read gets sharper as that history grows.
Prices are not predicted. They are calculated from your own costing models and rulebook, so the number on a quote can always be traced back to how it was worked out.
What to watch
A predictive read is only as good as the history behind it. Most businesses record their wins well and their losses poorly, and a loss with no reason attached teaches very little. Recording the outcome of every quote, and why it was lost where you know, is what makes the read useful. It also assumes the future will resemble the past: a new product line, a new market or a supply shock breaks the pattern, and the model will not know. That is why it should advise the team, never set the price or make the bid decision on its own.
Agentic AI: it carries out tasks
Agentic AI is the newest of the three and the most talked about. An agent is a generative model that has been given a goal, access to tools, and permission to take a sequence of steps on its own. Instead of handing you back a draft, it reads the inbox, looks up the record, decides what to do next, does it, checks the result and moves on.
The word "agent" gets stretched a long way in sales material, so a useful test is this: does the system decide its own next step, and does it act inside your systems (send, update, create, request) rather than return text for a person to act on? If the answer to both is yes, it is agentic. If not, it is generative AI with a nicer interface.
Where it turns up in quoting
- Watching email or the CRM for a buyer's reply, recognising a change of scope and preparing a revised quote for review.
- Requesting prices and lead times from suppliers for the bought-in items on a bill of materials, chasing late responses and collating the answers.
- Filling in the buyer's own spreadsheet, portal form or compliance return from the approved quote.
- Keeping the CRM up to date with quote status, revisions and the final outcome.
- Following up on a submitted quote at the right time, with a draft message held for sign-off.
An agent runs a loop. The approval gate and the log are what make it safe to let one work on your quotes.
Parker works as an agent across the steps of a quote. It watches for buyer replies in connected email and CRM systems, drafts revisions when the scope changes, requests and collects supplier prices, and fills in the buyer's own spreadsheet or return form.
The controls described below are built in. Parker prepares and recommends; sending a quote, an email or a price needs a person's approval, and every action is recorded in an audit trail.
What to watch
Because an agent acts, its mistakes compound. A misread change of scope early on becomes a wrong revision, a wrong supplier request and a wrong price later. In quoting, anything sent to a buyer or supplier is a commitment, so the controls matter more than the cleverness. Be clear about what the agent may do without asking (gather, collate, draft), what needs a person's approval (anything that leaves the business, and any price), and keep a log of every action it took. The sensible route is to let agents gather, prepare and recommend first, and widen what they do only once the pattern has proved itself.
How the three fit together
A quote is a natural example of all three working together, each handing on to the next:
- Generative reads the enquiry and the buyer's documents and turns them into a structured list of requirements.
- Predictive gives an initial read on fit and risk, so the team can decide whether to bid.
- Agentic requests prices from suppliers for the bought-in items and collates the replies.
- Generative drafts the technical answers and covering letter from your own knowledge.
- An engineer reviews, edits and approves.
- Agentic fills in the buyer's return file, then watches for their reply and prepares any revision.
- The outcome, won or lost, is recorded and feeds the next predictive read.
| Generative | Predictive | Agentic | |
|---|---|---|---|
| In a quote | Reading RFQs, drafting answers and emails | Judging fit and risk against past quotes | Supplier prices, buyer replies, return files, CRM updates |
| Needs | Your documents, organised in one place | A quote history with outcomes recorded | Connections to email, CRM and suppliers, and clear rules |
| Main risk | An invented claim in a submitted quote | Patchy records, especially of losses | An action that goes out before anyone checks it |
| Control | Sources shown; an engineer approves | Advises only; never sets the price | Approval gates and a full audit trail |
The common thread is that all three depend on your information being somewhere the model can reach it. In Mainspring Quote that is the Knowledge Hub for your documents, connections to your email, CRM and suppliers, a rulebook for the way you quote, and a quote history that records how each one ended. The model itself is increasingly the interchangeable part. Get the knowledge, the connections and the rules right, and the best model for each job can be used, and replaced as more capable ones arrive.
What to ask of any AI quoting tool
Knowing the three types turns a vague claim that a product "uses AI" into questions you can check:
- Which of the three is it using, and at which step of the quote?
- Are its answers grounded in your own documents, with a source an engineer can check?
- Can anything reach a buyer or supplier without a person approving it?
- Does it learn from the quotes you lose as well as the ones you win?
- Is every action logged, so you can see who approved what and when?
A good quoting tool should have a clear answer to each of those. Mainspring Quote was built around them: generative drafting grounded in your Knowledge Hub, a predictive read that learns from your quote history, and agentic steps that stop for an engineer's approval.
See Parker at work.
Book a demo with our team to see what Parker can do with your own quote requests.