AI automation FAQ
Short answers to the questions buyers ask most about AI automation: what it costs, how long it takes, how your data is protected, who owns the result and what happens after launch. Each answer reflects how Scenox Studios works today and links to the page with the full detail.
Getting started
What an AI automation partner does, and what you need before the first call.
What does an AI automation agency do?
An AI automation agency designs and builds software that takes repetitive work off your team: reading documents, answering routine requests, updating records and moving work between systems. Agencies differ most in what happens after launch. We run and monitor what we build, report against agreed metrics every month and fix it when something breaks.
How we workWhat does Scenox Studios do?
We design, build and run automation for business operations: end-to-end workflows, AI agents for tasks that need judgment, knowledge assistants that answer from your documents, private AI models and voice agents. They work inside the tools your team already uses, such as your ERP, CRM, help desk and inbox, with a person approving the actions you choose.
See all solutionsWhich workflow should we automate first?
Pick one repetitive, high-volume workflow that crosses several tools and has a person who owns it. Supplier invoices, document intake, support tickets, inbound calls and lead follow-up are common first choices. Start with one workflow and a fixed scope, then add more once the first has paid for itself. The readiness assessment scores a candidate in 12 questions.
Take the readiness assessmentDo we need an in-house AI team?
No. We build and run the system. Your team brings knowledge of the workflow and approves the actions you choose to keep with a person. What you need in-house is someone who owns the workflow and can answer questions quickly. You don't need engineers or a data science function. Building in-house is the alternative, and it needs a dedicated AI team.
What we need from youWhat do you need from us to start?
An owner for the workflow, a few hours with the people who run it today, access to the tools involved (usually a sandbox first), and a sample of past cases to test against. For invoices, that means a few hundred recent invoices and how each one was handled. That sample is what the prototype is scored on.
See a sample auditWhat if automation isn't worth it for our workflow?
We say so in the audit report. If the volume is too low, the inputs too inconsistent or the saving smaller than the running cost, the recommendation will be not to build, or to fix a process step first. You still have a measured baseline and a map of the workflow, which is useful whatever you decide.
What the audit covers
Cost and ROI
How the price is formed, what runs monthly, and how to check the business case.
How much does AI automation cost?
It depends on the workflow, the integrations and the volume. Cost has two parts: a one-time build, driven by how many systems the workflow touches, how varied the inputs are and how many exceptions need rules, and a monthly run cost for model usage, hosting and monitoring. Most clients start with a fixed-fee workflow audit, which ends with a fixed price for the build.
How pricing worksHow do AI automation agencies charge?
Agencies mostly use four models: a fixed fee per workflow, time and materials at an hourly or day rate, a monthly retainer for running the system, or usage fees per document, call or run. We use a fixed fee for the audit, a fixed price for an agreed build scope and a monthly fee for managed operations. Compare proposals on what they include.
AI automation cost guideWhy don't you publish prices?
We scope and quote each build after looking at the workflow, so a number on a web page would be a guess. Two workflows with the same name can differ a lot in systems, input variety and exceptions. The audit fee is quoted and agreed in writing before any work begins, and the audit report ends with a fixed price for the build.
How pricing worksHow do we calculate the ROI of AI automation?
Take what the workflow costs today, subtract what it costs after automation, including the review time that remains and the cost of running the system, and compare the result with the build cost. Measure the baseline on a real sample of items first. The ROI calculator does the arithmetic from your own volumes and quote and shows the payback month.
Open the ROI calculatorWhat running costs do people forget?
Monitoring and maintenance, and the human time spent on review and exceptions. Someone has to watch failures, re-test when a supplier changes its invoice layout or a model version changes, and handle small requests. Review time sits on your payroll, so vendor quotes tend to leave it out. Model usage is usually the smaller line.
Build and run costs explained
Process and timeline
The four stages, how long each takes, and what you receive at the end.
How long does an AI automation project take?
A pilot running on your own cases usually takes three to five weeks. Shadow running and a staged rollout to production take another two to five, depending on integrations and your security review. Most first workflows are in production within six to ten weeks. The workflow audit before the build takes one to two weeks.
The four project stagesWhat happens in a workflow audit?
We interview the people who run the workflow, time a sample and pull your system data, then run a prototype on your past cases. The report holds a workflow map, a measured baseline, which steps run on rules, models or people, the risks and controls, pilot thresholds and a week-by-week build plan with a fixed price.
Read the sample auditWhat is shadow running?
The system processes live work alongside your team without acting on it. We compare its output with what your team did, and it only starts acting on its own once it meets the thresholds agreed in the build plan. Shadow running sits between the pilot on your past cases and the staged rollout to production.
How a pilot runsWhat do we receive at the end of a project?
A system in production, running on live work and rolled out in stages. You also get an operator console with every task, approval and escalation in one place, an audit trail of each action, documentation and runbooks, training for the people who supervise it, and monthly reporting against the metrics agreed before launch.
What you receiveHow do you measure success?
We agree the success metrics before we build, then report against them every month. The main one is the share of tasks the system completes correctly on your cases, measured per field or decision rather than as one average. Silent errors are tracked separately from escalations, and the pilot has to hit agreed thresholds before anything acts on live work.
How we evaluate systems
Security and data
Training, deployment, what model providers see, and security reviews.
Is our data used to train AI models?
No. Your data is never used to train models for other clients, and we use model provider endpoints that do not train on API traffic, with zero data retention where the provider offers it. Custom models trained on your data belong to your deployment. On-premises deployments make no external model calls at all.
Data handling in the trust centerCan it run in our own cloud or on-premises?
Yes. Systems and custom models can be deployed in your AWS, Google Cloud or Azure account, or on your own hardware, when your security policies require it. There are four deployment models: managed by Scenox in a dedicated environment, in your cloud, on-premises, or hybrid, where models, the retrieval index and the audit log stay in your environment.
Enterprise deploymentsWhat do model providers receive?
Only the input a step needs, such as the text of one invoice or a question with its retrieved passages. Personal data is redacted first where possible, and content a user may not open is filtered out at retrieval time. System credentials never go to the model, because tools are called by the workflow engine with scoped service accounts.
Model providersHow do you protect against prompt injection?
Inbound email, documents and web content are treated as untrusted input. Agents run with scoped credentials and per-action limits, rules are enforced in code outside the model, and sensitive actions wait for approval. Production access to your systems starts read-only and widens in steps, so a bad instruction hidden in a document has little it can reach.
Access controlCan you complete our security questionnaire?
Yes. Send your own questionnaire or a standard one such as SIG Lite or CAIQ, and the engineers who will build the system complete it. We include the architecture and data-flow diagram and the model provider terms. A data processing agreement and the subprocessor list for your system are reviewed during procurement.
The security review packWill our system meet UK GDPR, EU AI Act or PDPL requirements?
We design systems to support your obligations under these laws, with data minimization, retention per workflow, in-region deployment, human oversight and full logging. Whether a specific use case meets them is assessed by you and your counsel, and we supply the architecture and data-flow documents they need.
Regional regulations
Ownership and IP
Who owns what we build, and how you leave if you want to.
Who owns the system you build?
Ownership, hosting and IP terms are agreed per engagement. We can operate the system for you, or hand it over and deploy it into your environment. Your data, prompts and evaluation sets can be exported or deployed into your own cloud, so you are not locked in. Custom models trained on your data belong to your deployment.
Ownership terms to agreeCan we take the system in-house later?
Yes. We can hand the system over and deploy it into your environment instead of running it for you. Every engagement includes documentation and runbooks that say how it works, how to change it and what to do when something breaks, plus training for the people who supervise it. Your data, prompts and evaluation sets can be exported.
What you receive
Models and accuracy
Which models we use, when people approve, and what happens on a mistake.
Which AI models do you use?
It depends on the task and your policies: OpenAI, Anthropic and Google through their enterprise APIs or your own cloud (Azure OpenAI, AWS Bedrock, Vertex AI), or open models such as Llama, Mistral and Qwen hosted in your environment. Each choice is listed in the architecture document for your review, and a model is swapped only after the evaluation suite passes.
How our systems are builtWhat happens when the AI gets something wrong?
Systems check their own work, ask for approval on sensitive actions and send uncertain cases to a person with full context. Uncertain cases stop for review before anything is sent, and every case is logged so the system can be improved. Accuracy is measured in production, with alerts when it drifts.
Human-in-the-loop designWhich actions need human approval?
You decide during design. Payments, refunds and outbound messages usually wait for a named person, along with anything above a threshold you set and any field the system is unsure about. The approver, the decision and the time go into the audit log. Limits and approval points are enforced in workflow code that the model cannot change.
Approvals and audit logsWhat is the difference between a knowledge assistant and a private model?
A knowledge assistant uses retrieval-augmented generation (RAG): it looks up your documents when a question is asked, so answers stay current and cite sources. A private model is a small language model trained on your examples to do one task very well, such as classifying tickets. It suits high-volume work that needs to be private, fast and cheap. Many systems use both.
RAG vs fine-tuning vs small models
Support and running costs
What happens after launch, who pays for model usage, and how incidents are handled.
What does managed operations include?
We monitor, maintain and extend your systems after launch, for a monthly fee that covers hosting, monitoring, model costs and improvements. Accuracy is measured in production with alerts when it drifts, a person reviews a weekly sample of cases, and you get a monthly report against the metrics agreed before launch.
Engagement modelsWho pays for model and API usage?
Managed operations is a monthly fee that covers hosting, monitoring, model costs and improvements. When model use has to be billed and governed inside your own cloud account, models are called through your own subscription, such as Azure OpenAI, AWS Bedrock or Vertex AI. Either way, cost per task is tracked for every workflow and estimated before you commit.
Who pays which costsWhat happens if something breaks in production?
Monitoring alerts us and your named contact. We can pause a workflow so new cases go to your team while the system is checked. The fix is re-tested against the evaluation set of your past cases before the workflow acts on its own again, and you receive a written incident summary of what happened and what changed.
Incidents and support
Agencies and white-label
How agencies offer AI systems to their clients under their own brand.
Do you work with agencies?
Yes. Agencies use us to deliver AI systems to their clients under their own brand, through referral, co-delivery or white-label. In white-label work your brand goes on the operator console, emails and voice agents. We work behind you as the delivery and operations team, and join client calls as part of your team when useful.
The partner programHow do pricing and margins work for agency partners?
We quote you a fixed price for each build and a monthly price for operations, at partner rates. You add your margin, present it as your proposal, decide what the client pays and invoice them yourself. We invoice you. Partner rates are set out in the partner agreement, so you can quote clients with confidence.
White-label pricingWho owns the client relationship?
You do. In white-label and co-delivery work we contact your clients only through you or with your agreement, and the partner agreement includes a non-solicitation clause covering the clients you bring. You handle first-line questions in your usual channels, and we handle anything technical behind your support line.
Who does what
Ask the question this page didn't answer.
A 30-minute call with an engineer. Bring the workflow you have in mind and we will tell you plainly what it would take, or whether it is worth doing at all.
What happens next
- 1
Strategy call
We learn how the workflow runs today and what a good outcome would be.
- 2
Workflow audit
We map the process and measure where a system would save the most time.
- 3
Proposal
A written scope, timeline and fixed price for your first system. No obligation to build.