SUPERWISE services
Agent build
For teams ready to turn one defined workflow into a tested agent in their environment.
$35,000 to $75,000 pilot · $90,000 to $200,000 production6 to 10 weeks pilot · 3 to 6 months production
Outputs you keep
The working agent, in your environment
Guardrail and scope definition
Test results and user acceptance record
What you get
Working outputs, not a presentation about them
One real workflow in your business runs with an AI agent your team has tested, understands and can extend, inside your own environment, in 6 to 10 weeks.
Use case brief and success measures
One to two pages naming the chosen workflow, who uses it, what data it touches, what the agent must never do and the measures that will decide whether it worked, set with your sponsor at kickoff. The yardstick for the pilot and a template for choosing the next workflow. If someone new joins the project, this is the page that explains what you are building and why.
The working agent, in your environment
A configured, running agent for the one pilot workflow, connected to the sandbox or limited-scope data agreed in scope, built with SUPERWISE Agent Studio components. Your team runs real work through it during acceptance testing and after handover. It is the live example that turns "what is an agent" into something people can see and touch.
Guardrail and scope definition
The written list of what the agent is allowed to read, what it is allowed to produce, which actions need a person's approval and the guardrails that enforce each limit, with the test cases that show each guardrail working. Answering your security, legal or compliance reviewer, and deciding later whether to widen what the agent is allowed to do.
Test results and user acceptance record
The functional and edge-case tests we ran, the guardrail checks, and the notes from your own people who tried the agent on real examples, with what passed, what failed and what was fixed. The evidence behind the production decision, and proof to a board, client or auditor that the agent was tested before it was trusted.
Training session for your team
A hands-on walkthrough for the people who will use and look after the agent: how to use it, how to read what it did, when to override it and how to report a problem. Your team runs the agent without us in the room, and new colleagues can be brought up to speed from the session notes.
Operating and extension guide
Documentation of how the agent works: its inputs and outputs, the data connections, the guardrails, the configuration and how to change or extend it. Handing the agent to a successor, adding a step, or giving it to another provider if you choose to work with someone else.
Next use cases roadmap
A short, ranked list of the next workflows worth building, based on what the pilot taught us about your data, your systems and your people. Planning the next quarter's AI work and your budget request, so the first agent is the start of a sequence and not a one-off.
Outcomes
What will we have at the end?
A workflow your people stop doing by hand
The agent takes over a defined, repeatable piece of work, such as sorting incoming requests or pulling facts out of documents, and your team reviews its output instead of producing it from scratch. We set the baseline at kickoff, so the time saved is measured against how long the task takes today.
A decision about AI you can make with evidence
The pilot ends with test results, user acceptance feedback and a production decision, so the choice to go further, stop or change course rests on how the agent performed on your own work. Nobody has to take a demo on faith.
Something your leadership and your clients can be shown
You get a live agent in your environment, the guardrails it runs under and the record of how it was tested. When a board member, client or auditor asks what the AI does and what stops it from doing more, you can show them the agent and the controls instead of describing a plan.
Examples
What could this look like in our business?
These are examples of suitable work, not claims about past client engagements.
Example: intake triage
Requests arrive by email or form. The agent reads each one, tags its type and urgency, drafts the routing and hands it to the right person. A person approves anything that leaves the building.
Example: document extraction into a system
Invoices, claim forms or order documents arrive as files. The agent reads them, pulls out the fields your team keys in today and writes them to a staging area in your system. A person checks exceptions before anything posts.
Example: claim or ticket summarization
Long claim files or support ticket threads get a short, structured summary with the facts, open questions and the next step, so the reviewer starts from the summary and opens the source when something looks off.
Example: internal knowledge questions on approved data
Staff ask questions about policies, procedures or product documents and get answers drawn only from the documents you approve, with the source shown. Anything outside those documents gets a plain "not in the approved material".
Example: scheduling and follow-up drafts with approval
After a call or a job, the agent drafts the follow-up message and proposes the next appointment. A team member reads it, edits it and sends it. The agent sends nothing on its own.
Example: warranty or returns review
The agent checks a returns or warranty request against your written rules, lists what it found and recommends approve, deny or escalate. A reviewer makes the call and the reasons stay on record.
Success measures
How do we know it worked?
We agree the measures at kickoff and use them in the final decision.
Measures agreed at kickoff
- Task completion without help: the share of real cases the agent finishes correctly with no person stepping in, counted on a test set from your own work.
- Escalation rate and reasons: how many cases the agent hands to a person, and why, so you can see whether it escalates the right things.
- Time per task against the baseline: how long the task takes with the agent compared with the time it takes today, measured at kickoff and again at acceptance.
- Cost per completed task: what each finished task costs to run, compared with the cost of the same task done by hand.
- User acceptance: the people who will use the agent test it on real examples and sign off, or tell us what has to change before they will.
Your team
What does our team need to give?
About 8 to 12 hours a week across your team, most of it from the domain expert, and more in the acceptance weeks. These are planning figures; we confirm them against your workflow when we scope.
Executive sponsor
Chooses the workflow with us, confirms the success measures at kickoff, clears blockers and makes the production decision at the end.about 1 hour
Domain expert
Knows how the work is done today. Walks us through the workflow, supplies real examples and the rules people follow, and judges whether the agent's output is right.about 4 to 6 hours
System and data owner
Grants access to the sandbox or the limited data sources, answers questions about the systems involved and approves any connection to a live system.about 2 hours in weeks 1 to 6
Approver and acceptance testers
Two or three people who will use the agent test it on real cases in the acceptance weeks and report what works and what does not.about 2 to 3 hours each in weeks 7 to 10
Duration
A build path your team can see
6 to 10 weeks pilot · 3 to 6 months production
Weeks 1–2
Confirm the workflow, success measures, data, and guardrails.
Weeks 3–6
Build, connect, and test the agent with your team.
Weeks 7–10
Complete UAT, training, documentation, and the production decision.
Data and security
Is our data safe?
How the engagement handles data
- Where it runs: the agent is built and demonstrated inside your environment, using SUPERWISE Agent Studio, so your team sees it working where it will be used and not in a separate demo.
- What data it touches: the pilot connects only to a sandbox or limited-scope data sources that you approve in scope. Anything wider, such as a live system or regulated data, is a scope change we agree before we connect it.
- Protected health information: if the workflow involves PHI, tell us at the start. The SUPERWISE platform is not covered by a BAA, so PHI stays out of it, and we scope the pilot around data that is not PHI.
- Guardrails and audit trail: guardrails and observability are set up from the first week, so each limit on the agent is written down and tested and what the agent did can be reviewed afterwards.
- What we never do: we do not widen the agent's access or connect it to live data without your approval, and the agent does not take an action outside the approvals defined in its scope.
Ownership
What do we own when it is over?
At the end you own the decisions, the documents and the configuration: the use case brief, guardrail definition, test results, training notes and operating guide are yours, and so are the agent's source and configuration as listed in the scope. The agent itself runs in SUPERWISE Agent Studio, so how long it keeps running after we leave depends on the platform subscription, which is separate from the engagement fee and is agreed in the scope before you sign. If you stop working with us after the pilot, you keep the documents, the configuration and the guide, which are written so another team can carry the work forward.
After the engagement
What happens after?
The pilot ends with a production decision. If you go ahead, the production build is a separate engagement, from $90,000 over 3 to 6 months, for an agent running on live data with integrations, monitoring and handover. If you want us to keep a live agent monitored and maintained, Managed AI operations starts from $3,500 a month on a 12-month term. If the pilot shows you need to map more systems first, Data and systems discovery is the usual step before another build. Support after handover is agreed in the scope.
Draft pricing for review
Know the range before the first call
We confirm a fixed scope and price after a scoping call.
| Tier | Price range | Duration | Included in a line |
|---|---|---|---|
| Pilot | $35,000 to $75,000 | 6 to 10 weeks | One pilot workflow |
| Production | $90,000 to $200,000 | 3 to 6 months | Sandbox or limited data connections |
Not included
- A second workflow. The pilot builds one; a second workflow raises the price and is scoped separately.
- Third-party software subscriptions, including model, data and system licenses your agent depends on.
- Fixing your source systems. If the pilot reveals data that is missing, inconsistent or locked in a legacy system, we report it and you decide on the clean-up as a separate piece of work.
- A production rollout. The pilot runs on sandbox or limited data; production is the next engagement.
- Ongoing monitoring and maintenance after handover. That is Managed AI operations.
- Legal advice or regulatory sign-off for your use of AI.
Fit
Know whether this is the right starting point
Teams of 50 to 500 people with a selected workflow
Good fit if
- For teams ready to turn one defined workflow into a tested agent in their environment.
- Teams of 50 to 500 people with a selected workflow
- You want a defined handover your team keeps.
Start somewhere else if
The workflow or supporting data is not yet clear.
Start with Data and systems discoveryWho delivers and proof
Delivered by the SUPERWISE services team
We have not published a case study of an agent build engagement, so we do not show one. The nearest real stories on this site are Renova Health, where AI assistants give chronic care managers the information they need while every AI decision is logged and traceable to its source, and Bridgestone, where a governed workflow connects warranty claim inspections, approvals and resolutions with end-to-end traceability. Neither is presented here as the result of this service.
Related published work
The nearest approved examples for this work.
Agent build questions
What to expect before you start
A pilot runs $35,000 to $75,000 over 6 to 10 weeks. The low end covers one workflow, sandbox or limited data, guardrails, testing and user acceptance, training and documentation. A second workflow, more integrations such as an ERP or line-of-business system, and regulated or sensitive data all raise it. A production agent starts from $90,000. We confirm a fixed price after a scoping call, and the pricing on this page is draft pricing for review.
Choose one that is repeated every day or week, follows rules a person can write down, has data you can reach, and is visible enough that people will notice when it works, but where an error is easy to catch. We help you pick in the first two weeks, using the Quick Scan or an assessment if you want a structured way to compare candidates.
You need people who know the workflow and someone who can grant access to the data and systems involved. You do not need AI engineers: we build and configure the agent and train your team to use and extend it. If you have no IT staff, tell us in the first call so we plan the access and security questions with whoever looks after your systems.
We agree the success measures at kickoff and test against them with your own cases, so you find out during the build, not after. If the results fall short, the readout says so, with the reasons and what would have to change. The decision to stop, adjust or go to production is yours.
You keep the documents, test results, configuration and operating guide, and the source and configuration listed in the scope. The agent runs in SUPERWISE Agent Studio, so keeping it live depends on the platform subscription, which is agreed separately before you sign. We tell you exactly what continues and what stops before the engagement starts.
A chatbot answers questions. This engagement builds an agent for one defined piece of work, with guardrails, tests, user acceptance and documentation, inside your environment and with your team involved throughout. A freelance developer can build a prototype. This includes the testing, the written limits on what the agent is allowed to do, the training and the guide your team needs to run it.
The service path
Where this fits
Learn about the decisions
Does the AI advise a person, or act on its own?
Set the agent's authority before its workflow is built.
Is it safe to let an AI agent act on company systems?
Choose the permissions, checks, and stop conditions the build needs.
What is prompt injection, and how do you defend against it?
Understand one of the threats the build and tests must cover.
Products this service puts to work
Accelerate this engagement with Council
Use Council to turn the workflow requirements into a reviewed system design before and during the build.
The SUPERWISE platform
The working agent runs on the platform with its policies and activity record.
Talk to us
Tell us about Agent build
A few details help the services team route your request.