What is AI Script Generation?
AI Script Generation builds a complete, editable call script from a short description of what you want the assistant to do. You provide a name, a call direction, who the assistant will be talking to, a voice, and a description — and optionally a website URL and supporting documents. Voxworks then generates the full package: a call flow with branching steps, Call Objectives, Guidelines, a populated Knowledge Base, and a voicemail message. Generation runs in the background with a live progress log, and the finished script opens in the standard script editor — it is a starting point to review and refine, not a locked artefact. The same generator powers the AI Receptionist creation flow, with a simpler form (see Generating an AI Receptionist below).Starting a generation
- Go to Call Scripts and select New script.
- In the Create New Script dialog, Generate with AI — “Describe your script and let AI generate it for you” — is already selected for every script type.
- Select Create. The Generate Script form opens.
Generation inputs
Select Generate to start. The button reads Generating… while the request is submitted.
Document limits
- Formats — PDF, DOCX, PPTX, TXT, Markdown, CSV, or JSON. Files in any other format are skipped with a warning under the dropzone.
- Count — up to 5 documents per generation.
- Size — up to 10 MB per file.
- Content — text is extracted from each file and the first 12,000 characters are used. Duplicate files (same name and size) are added only once.
Website research
When you provide a website URL, the generator researches the site before writing the script:- Page discovery — it starts from the URL you gave, then looks for high-value pages via common paths (about, contact, FAQ, help, support, services), the site’s sitemap, and links on the homepage. It visits up to 5 pages, prioritising the homepage plus one each of FAQ, about, services, contact, and pricing pages where they exist.
- Business grounding — from the visited pages it distils your positioning, products and services, brand tone, audience, and key facts (locations, hours, policies), and writes the script around them. The generator is instructed not to invent facts beyond what it found — where a detail is missing, the script falls back to offering human follow-up rather than guessing.
- Verbatim FAQs — every question-and-answer pair found on FAQ pages is copied into the script’s Knowledge Base word-for-word, one entry per question, with the caller’s question in the Question column. Answers are never summarised or paraphrased, so the assistant answers common questions using your site’s exact wording. Duplicate questions are removed. The wider business context the research finds lands as its own entries — Website: About, Website: Products and Services, Website: Audience, Website: Brand Tone, Website: Key Facts and Website: FAQ.
The progress log
Once submitted, the modal switches to Generating Script, with a progress bar and a streaming log of steps. The current step shows a spinner; when more than 5 steps have run, earlier entries collapse under a “completed steps” toggle you can expand. A typical run logs steps such as:- Script generation request received.
- Will research https://yourbusiness.com for business context.
- Researching https://yourbusiness.com for business context…
- Website research complete. FAQ items: 12.
- Identifying conversational structure and key pathways…
- Building script flow and branching logic…
- Sending to AI for generation…
- AI response received. Processing output…
- Validating script structure…
- Saving script to database…
- Script generation complete!
Setup needed notes
After a script generation completes, an automatic check pass verifies the generated content against your team’s real configuration — for example, matching generated tool steps to the tools actually available to your team. Anything that needs your attention is appended to the log as a Setup needed: entry, such as a required tool field you must fill in or a variable reference to replace or remove. Open the script in the editor and work through these notes before testing. Generated receptionists use a fixed, pre-verified call structure, so they skip this check.What gets generated
When the log reaches Generation Complete, you see a summary of the script’s type and contact, with buttons to Close or Open Script. Opening the script takes you straight to the script editor. The generated script includes:- Call flow — voice steps with branching conditions and a defined end of call, following the pathways in your description.
- Call Objectives — success criteria for the call.
- Guidelines — behavioural rules and safety boundaries.
- Knowledge Base — facts from your description, researched website content, verbatim FAQ entries, and uploaded documents.
- Voicemail message — a concise message for unanswered outbound calls.
- Default Voice — the voice you selected on the form.
{{contact.first_name}} (string, e.g. Jane).
The Data Collection tab starts empty — add Data Collection fields yourself if the script should capture structured data.
Always review the generated script step by step and run a test call before using it in production.
Generating an AI Receptionist
The receptionist form is shorter because receptionists always answer inbound calls from callers — there is no Call Direction or Contact Type to set:
The modal titles read Generate Receptionist and Generating Receptionist, and the log follows receptionist-specific steps (building the greeting, instructions, and message flow). The generated receptionist has a consistent call structure:
- A greeting spoken word-for-word at the start of every call.
- A reception hub that works out what the caller needs and directs the conversation.
- A take-a-message branch that captures the caller’s name, number, and reason for calling.
- An offer-callback branch for callers who ask to speak with a person — newly generated receptionists have no transfer destination configured yet, so they offer a call back instead until you set one up.
- A wrap-up step that confirms next steps and ends the call.
Failures and retries
Generation retries automatically. Each job is attempted up to 3 times:- If an attempt fails — for example, the AI returns output that fails validation — the log shows “Error: (reason) - will retry.” and the job re-runs. Completed website research is kept between attempts, so retries do not re-crawl your site.
- If the final attempt fails, the log shows “Error: (reason) - max attempts reached.” and the modal returns to the input form so you can adjust your inputs and generate again.
- If the request cannot be submitted at all, the error appears in the log immediately (for example, “Error: Could not reach the server. Please try again.”).
AI recommendations
Once a script exists, Voxworks keeps reading it and suggesting improvements. Suggestions are reviewed in the script editor, one at a time, and nothing changes until you accept a suggestion and save.Turning the review on
The wand button in the editor toolbar switches AI recommendations on and off. Its tooltip reads Turn on AI recommendations, or when suggestions are waiting, “N AI recommendations waiting” — and a red count sits on the button (it reads 99+ above ninety-nine). With the mode on, every line a suggestion touches is marked across all the editor tabs, and a small floating toolbar headed AI recommendations appears. Drag it anywhere on screen.Reviewing a suggestion
The toolbar shows one suggestion at a time — its title and the reason for it — with the change itself marked in place:- The old wording sits on a red row, struck through.
- The new wording sits on a green row.
- Within a rewritten line, the exact words that changed are highlighted, unless most of the line was rewritten.
Progress lines tell you where you are: All reviewed. Save the script to keep the change., All reviewed — nothing kept., No active suggestions., or “No active suggestions — N waiting on other tabs.” when the remaining suggestions are on a tab you do not have open.
Nothing changes until you save
Accepting or dismissing changes nothing on its own. Your decisions live in your working copy, exactly like any other edit, and are written to a version only when the script save succeeds. If a save fails, your pending decisions are still there to try again — and every dismissal can be undone.Regenerate — review this script now
Reviews run automatically each night. To ask for one immediately, use Regenerate on the toolbar (tooltip: Ask the AI to read this script now and suggest improvements). Save your changes first — with unsaved edits the button is blocked, with the tooltip Save your changes before running a review. While a review runs the toolbar reports what it is doing — “Reading the script steps, wording and routes…”, then “Turning the review findings into recommendations…” — and finishes with Review finished — N new suggestion(s). or Review finished — nothing new to suggest. New suggestions appear straight away, with no need to reload. If the script has not changed since its last review, the toolbar says Already reviewed at this version — save a change to make a new review worthwhile.What suggestions can change
Recommendations are deliberately narrow. They can only:- Correct the wording of a voice step’s script
- Rewrite a voice step’s conditions
Where suggestions come from
Two sources feed the review:- The script itself — read on its own, looking for wording that will be misheard, conditions that overlap, and gaps a caller could fall through.
- Flagged calls — any transcript message you flag with Flag this response in the call detail, with an optional note about what was wrong. Flags are gathered at the start of each review and turned into suggestions for the steps involved.
Next Steps
- Creating a Script — the other creation methods: from a template, from an existing script, or from scratch
- Script Editor Guide — review and refine every part of the generated script
- Knowledge Base — where researched FAQs and document facts land, and how to keep it lean
- AI Receptionist Overview — finish setting up a generated receptionist with a number and tools