AI assistance

CPQ can turn a plain-English description into a working draft: a workflow, a business rule, a custom field or a formula. The drafting is done by a language model, but nothing reaches you unchecked — every generated artefact is run through the same validators the product itself uses, and anything that still fails after two attempts at correction is thrown away rather than handed to you.

Where you meet it

The workflow designer

Open a saved workflow under Workflows and press AI in the toolbar — the button is only on the toolbar of a workflow that already exists. A chat panel opens on the right: describe what you want — “quote approval with three tiers and email notifications”, “email the owner when a contract is about to expire” — and the assistant replies in the panel and draws the workflow onto the canvas. The canvas is sent with every message, so the assistant is always looking at the workflow you have open. Nothing is saved until you save it, so a draft you do not like can be discarded by leaving without saving.

While the panel is still empty it also offers an Analyze this workflow shortcut for the workflow already on the canvas, which comes back with strengths, gaps (each with a severity and a suggested fix) and optimisations as a written review. It is advice only — it changes nothing.

The assistant can answer in two ways: with a complete workflow, or with an incremental edit to the one you have. Only a complete workflow is drawn onto the canvas. An incremental edit is validated and acknowledged in the chat but leaves the canvas as it was, so if you ask for a small change and nothing moves, ask again for the whole workflow — “rebuild this workflow with an approval step before the email” — and it will be applied.

Generate from inside the designer, not from the list

The AI Generate bar on the Workflows list drafts and validates a workflow, reports success, and then opens an empty designer — the draft it just made is not carried across. Use the AI panel inside the designer instead.

Starting from scratch takes one extra step, because the AI button is not shown on an unsaved workflow. Name the new workflow, drop a trigger node on the canvas — a workflow will not save without one — and save. The designer reopens on the saved workflow with AI in the toolbar, and a complete workflow from the assistant replaces whatever placeholder you started with.

Business rules

Open a rule under Rules — an existing one or a new one — and press AI Generate. Describe the rule (“enterprise customers with orders over $50K get 20% max discount”) and CPQ fills the rule’s conditions and actions into the builder for you to review, adjust and save. Your description is also kept as the rule’s description. Everything it writes is ordinary rule content: you can edit every condition and action afterwards, and nothing takes effect until you save.

This is the business-rules module. It is not the pricing-rule builder described in Products & pricing, which has no AI drafting.

Custom fields, formulas and objects

Under Admin → Custom Fields, pick a custom object and the field builder shows a Describe the field and let AI draft it box. Describe the field you want (“track customer satisfaction from 1 to 10”) and the suggestion is merged into the field form — type, name, label, help text, required flag, length, range, precision and picklist values — ready for you to correct before saving. This works on custom objects only; on a standard object the builder tells you so.

Admin → Schema Builder → Field Builder offers the same idea as a wizard step — the field wizard is opened from Schema Builder, not from the Admin menu. Switch the first step from Manual Selection to AI Suggest, describe what you need, and CPQ proposes a handful of fields; Accept carries one into the wizard’s configure step. The suggestions are drafted from your description — the assistant is not told which object you picked, so treat them as a starting shape rather than advice about that object.

The same wizard has an AI Generate button on a formula field. It is of limited use today: the dialog sends no list of your fields with the request, so any expression that names one of your fields is rejected as an unknown field reference and the dialog reports a failure. Write formula expressions by hand — see the formula manager — and use the AI box only for arithmetic over constants.

Admin → Schema Builder has an AI Suggest dialog that proposes whole objects from a business requirement, each with fields and relationships. Read it as a design sketch: the Create All button at the foot of the dialog reports success but creates nothing. Build the objects you like from the sketch by hand.

Every draft is checked before you see it

This is the part that matters most. A language model will happily produce a workflow with two triggers, an edge pointing at a node that does not exist, or a rule whose operator does not exist. CPQ does not show you any of that.

Every generated artefact is validated by the same rules the runtime enforces. If the validation finds errors, the diagnostics are sent back to the model with an instruction to fix them, and it tries again — up to two corrective rounds. Then one of three things is true:

Verdict	What it means	What you get
Green	Validated clean	The draft, applied as normal
Amber	Valid, with advisory warnings	The draft, plus a note listing the warnings
Red	Errors survived both repair rounds	Nothing. The request fails and the diagnostics are shown

A red verdict is a refusal, not a silent failure: the workflow list and the rule form show “AI generation failed” with the first few problems spelled out, and the chat panel says the workflow did not pass validation and was not applied. You never end up with a broken artefact on the canvas because the model had a bad turn. When a draft did need correcting, the designer says so — “validated after 1 repair round”.

What is actually checked:

• Workflows — exactly one trigger; every node of a type the designer knows, with an id of its own; every edge joining nodes that exist; no self-loops and no cycles; condition branches that lead somewhere; and complete configuration on Send Email and Update Record steps. Unlabelled nodes, unreachable nodes, a condition missing an explicit yes or no branch, and references to fields outside the catalogue are warnings — they come back amber rather than blocking.
• Business rules — at least one condition; every condition naming a field and using a supported operator (equal, notEqual, greaterThan, greaterThanOrEqual, lessThan, lessThanOrEqual, contains, in, notIn, isEmpty, isNotEmpty); a value present unless the operator is isEmpty or isNotEmpty; a list where in and notIn expect one; and actions drawn from the rule module’s own action vocabulary.
• Formulas — parsed by the same expression parser the formula manager uses, with every field reference resolved and the return type checked.
• Field and object suggestions — these are filtered rather than repaired. Suggestions with an invalid field type, a name that is not lower-case snake_case, or a picklist offered without any values are dropped before the list reaches you. Duplicates are not caught: the builder does not send the fields you already have, so the assistant can re-suggest one of them — check the name against your object before you accept a suggestion.

The conversation is remembered

The workflow chat is a real conversation: it keeps its history on the server, in CPQ’s own database, so it survives a restart of the service and follow-up messages like “now add a rejection branch” have the context they need.

Two things to know about it. The panel’s on-screen transcript is not stored — reload the page and the messages disappear from view although the assistant still remembers them. And there is one conversation per user rather than one per workflow, with no way to start a fresh one from the panel, so a long-running thread carries across the workflows you open. Because the canvas is re-sent with every message, the assistant still works on the right workflow; it is the conversational history that persists.

What the assistant knows about your data

It is given the standard field catalogue for the object type you are working on — quote, customer, contract, product, order, subscription, invoice or opportunity — so it drafts conditions against real field paths such as quote.total_amount or customer.tier, and a reference to anything outside that catalogue comes back as a warning.

That is the extent of it. The assistant does not see your approval thresholds, your pricing rules, your workflow run history or any of your records, so give it the numbers you want it to use rather than expecting it to look them up: say “over $50,000” rather than “over our approval threshold”.

Every generation is recorded

Each AI request is written to an audit trail in CPQ’s database: who asked, which feature they were using, the prompt, the resulting artefact, the verdict, how many repair rounds it took, the time it took and the tokens it consumed. Failed and rejected generations are recorded too, with the reason. There is no screen for this yet — an administrator reads it from the database — but it means AI use is accountable after the fact, and it is the record to check when someone asks where a rule came from.

Limits and setup

The workflow and rule assistants cap a prompt at 2,000 characters, strip control characters, and reject outright any blatant attempt to override the assistant’s instructions (“ignore previous instructions”, “reveal your system prompt”). They are also rate-limited per user — twenty requests a minute by default — and the chat panel tells you plainly when you have hit the limit. The field, formula and object suggestion boxes are lighter: they carry no prompt guard and no separate rate limit of their own.

AI is a deployment-time setting: CPQ reaches the model through the WorkSquares AI gateway, and until an administrator configures a gateway key the AI buttons will fail. Keys can be issued per company and are stored encrypted; usage is attributed to the company and the user who asked, which is what makes the audit trail above meaningful.

Pricing analysis

CPQ’s API carries a deal adviser as well — score a deal against similar historical deals, recommend a price with a floor and a ceiling, flag anomalies across a batch of deals, and suggest cross-sell — and the deal analysis goes through the same verify-and-repair gate. Nothing in the web app calls it today, so it is available to integrations rather than as a screen.

The Android sales app has an AI Pricing Analysis screen, opened from a quote. It does not currently produce a usable answer: the screen and the API it calls do not agree on the shape of the response, so the analysis renders as zeros with an “Unknown” risk level. What sits behind that endpoint is in any case a scoring heuristic computed from the quote’s own figures, not the language model. See Mobile.

Connects to

• Products & pricing — pricing rules and the waterfall, which are configured by hand.
• Quotes & approvals — the approval processes a generated workflow will typically be built around.
• Mobile — the Android apps, including the pricing analysis screen above.