GaaS Examples: Agentic as a Service in Practice
8 min readConcrete GaaS examples: what an agent does in ticket triage, lead qualification, document processing and code review, what people approve and what is measured.
Zunkiree Labs Team
· Updated
Agentic as a Service (GaaS) means a provider runs AI agents that complete a defined piece of work for you, and you pay for completed outcomes rather than software seats. In practice that looks like an agent working through a queue, a person approving the risky steps, and a measured result for each unit of work.
Key takeaways
- A GaaS example always has three parts: what the agent does, what a person approves, and what outcome is counted.
- The best first workflows are repetitive, high-volume and easy to check, such as ticket triage, lead qualification, document processing and code review.
- Human approval is part of the design. It sits at the steps where a mistake is costly or hard to reverse.
- The outcome must be defined before the work starts, because it is what you are billed on and what you judge the service by.
- Start with one workflow, measure it against the manual baseline, then expand.
What does a GaaS arrangement look like in general?
The idea is explained in What is GaaS?. The short version: with software as a service you buy a tool and your people operate it. With GaaS you hand over a goal and a provider operates the agents, the integrations and the monitoring, and you pay for the results. Zunkiree Labs describes this on its Agentic as a Service page as agents that receive goals, make decisions, take actions and deliver measurable outcomes.
Anthropic's engineering guidance on agents draws a useful line here. It separates workflows, where models and tools follow predefined code paths, from agents, where the model directs its own steps and tool use. It also advises building in human checkpoints, where an agent pauses for feedback at set points or when it hits a blocker. Every example below follows that pattern: the agent does the routine work, and a person holds the decisions that matter.
The four examples that follow are generic illustrations of how such an arrangement could be set up. They are not descriptions of a specific client, and the numbers you would track depend on your own baseline.
Example 1: How does support-ticket triage work as GaaS?
- What the agent does: reads each new ticket, classifies the issue, checks it against the knowledge base and past tickets, drafts a reply for the common cases, and routes the rest to the right queue with a summary attached.
- What a person approves: any reply that involves a refund, a policy exception, a legal or safety topic, or an unhappy customer. Support staff also review a sample of the agent's routine replies each week.
- What is measured: tickets resolved without a human rewrite, time to first response, share of tickets routed to the correct queue the first time, and customer satisfaction on agent-handled tickets.
The billable outcome is usually a resolved ticket that was not reopened within an agreed window. Counting only closed tickets would reward an agent that closes things badly, so the definition needs a quality condition.
Zunkiree Labs has one client case study in this area. Admizz Education, an education consultancy, used Zunkiree Search to power an inquiry system where students ask questions in natural language and complex queries are escalated to human counselors with context preserved. The Admizz case study reports 45% faster response times and a 60% reduction in manual inquiry handling. Those figures come from the client case study, not from an independent audit.
Example 2: How does lead qualification work as GaaS?
- What the agent does: takes inbound enquiries and form fills, enriches them with public company information, scores them against criteria you define, asks the lead a few clarifying questions, and books a meeting or passes the record to sales with a short brief.
- What a person approves: the scoring criteria before launch, any outbound message to a high-value account, and every disqualification of a lead that matches a strategic target list.
- What is measured: qualified leads accepted by sales (not merely leads generated), meetings that took place, and the rate at which sales rejects the agent's qualification.
This is where outcome definitions matter most. A "lead generated" is easy to inflate. A "lead accepted by sales" ties the provider's revenue to something your team has actually judged. The rejection rate gives you a running quality check on the agent's judgement.
Example 3: How does document processing work as GaaS?
- What the agent does: reads incoming documents such as invoices, contracts or forms, classifies each one, extracts the key fields, validates them against rules and existing records, and posts the data to the downstream system or routes the document to the right team.
- What a person approves: exceptions. Documents the agent flags as unusual, missing information, or low confidence go to a reviewer. Large-value or irreversible postings can require sign-off every time.
- What is measured: documents processed correctly without human touch, exception rate, extraction accuracy on a sampled audit, and turnaround time from receipt to routing.
The CMS Group case study on Zunkiree Labs' site follows this shape. CMS Group processes contracts, invoices, compliance forms and internal records, and the case study states that 85% of its document processing workflows are now automated, with staff reviewing exceptions instead of processing every document by hand. The page labels its results as illustrative, based on a representative engagement, and says individual outcomes vary and are not guaranteed. Treat the 85% as that client's reported figure, not a promise.
Example 4: How does code review work as GaaS?
- What the agent does: reviews each pull request for bugs, style issues, missing tests and risky patterns, runs the checks it is allowed to run, and posts comments with suggested fixes.
- What a person approves: the merge. The agent comments; a developer decides. Changes to authentication, payments or infrastructure need a named human reviewer regardless of what the agent says.
- What is measured: review comments that developers accepted, time from pull request to first review, defects caught before merge, and defects that escaped to production despite review.
Anthropic's guidance recommends that agents check ground truth from their environment at each step, for example test results or code execution output, rather than trusting their own reasoning. Code review is a good example because that ground truth is available: a test either passes or it does not.
What do these examples have in common?
- A narrow, defined scope. Each workflow has clear inputs and a clear definition of done.
- An escalation path. The agent knows what it should not decide and hands it over with context. UiPath's order-to-cash example shows the same pattern: when a discrepancy such as a price mismatch appears, an agent routes the case to a person and automation resumes after it is resolved.
- A checkable outcome. The result can be verified by someone other than the agent.
- A baseline. You know how long the work took and how often it went wrong before the agent arrived.
What controls should surround the agent?
Autonomy adds risk. Anthropic notes that agentic systems can compound errors, since a mistake early on can carry through later steps, and recommends testing in sandboxed environments with guardrails and setting stopping conditions. Tungsten Automation's comparison lists controls for agentic systems: audit logs for every action, defined authority boundaries, least-privilege access and human review for high-risk decisions.
Zunkiree Labs' own GaaS page describes its approach in similar terms: ephemeral sandboxed execution environments, Human-in-the-Loop gates for high-risk actions, audit logging, and no access to production systems without explicit approval gates. When you evaluate any provider, ask to see these controls for your specific workflow, not in general.
How do you choose a first GaaS workflow?
- Pick work that is repetitive, high-volume and already documented.
- Choose a workflow where a wrong answer is cheap to catch and fix.
- Write the outcome definition and the quality condition in plain language before any build starts.
- Record the manual baseline for two to four weeks, so the improvement is measurable.
- Agree which steps need approval and review the approval rate over time. If people approve everything without changes, the gate may be moved or removed. If they change a lot, the agent needs work.
- Run it as a pilot on one workflow before expanding, as Zunkiree Labs suggests for teams moving from SaaS to GaaS.
Be cautious about the market noise. Most published numbers about AI agents come from vendors, so treat them as a reason to start small and measure on your own workflow, not as a statistic to rely on.
Where Zunkiree Labs fits
Zunkiree Labs is a GaaS company: it designs, deploys and operates agents itself, so customers pay for completed outcomes rather than managing infrastructure or seats. Its Agentic as a Service page lists workflows such as ticket triage, lead qualification and code review as typical starting points, and its pricing model is cost per outcome. If you want to test the approach on one workflow, the page describes a pilot on a single process before wider rollout.
Frequently asked questions
What is an example of Agentic as a Service?
A provider runs an agent that triages your support tickets: it classifies them, drafts replies for routine cases and routes the rest to staff. A person approves sensitive replies, and you are billed on tickets resolved to an agreed definition.
Which workflows suit GaaS best?
Repetitive, high-volume work with clear inputs and a checkable result, such as ticket triage, lead qualification, document processing and code review. Work where mistakes are costly or irreversible needs stronger human approval.
Do people still approve things in GaaS?
Yes. Good GaaS designs place human approval at high-risk steps and for exceptions. The agent handles the routine work and a person decides where the stakes are high.
How is the outcome measured in GaaS?
By a definition agreed before work starts, for example a ticket resolved and not reopened, or a lead accepted by sales. A good definition includes a quality condition so volume alone is not rewarded.
Are the case study figures guaranteed?
No. They are results reported in Zunkiree Labs' client case studies. The CMS Group page says its results are illustrative and individual outcomes vary and are not guaranteed.
Sources
- Anthropic: Building effective agents
- UiPath: What is robotic process automation (RPA)?
- Tungsten Automation: Agentic AI vs Traditional RPA
- Automation Anywhere: RPA vs agentic AI (vendor article, cites Gartner)
- Zunkiree Labs: Admizz case study
- Zunkiree Labs: CMS Group case study
- Zunkiree Labs: Agentic as a Service