Start with one repeatable workflow. Give the agent the same permissions as the person who’d normally do the work, and keep a human approving anything that can’t be undone. That’s our recommendation, not a rule any vendor states. Pieces of it do show up in three recent announcements: Trovata holds its treasury agents to the same permissions as human users, Google mentions human oversight for Gemini for healthcare, and DevRev says its service agents’ actions are reversible. None of them are built for a 30-person company, but the decisions behind them carry over.
What’s different about an agent compared with a chatbot?
A chatbot answers a question. An agent does a piece of the work. In its AWS write-up, Trovata describes agents that run on schedules, triggers or rules. They monitor activity, explain variances, flag anomalies and send updates without someone prompting them each time. Trovata also stresses that its AI is built into the treasury system and tied to real treasury records, not bolted on as a generic chatbot.
The other two announcements point the same way. Google says Gemini for healthcare goes beyond information retrieval and can run workflows with many steps, like synthesizing medical records and reviewing prior authorizations. DevRev says it built its Computer and Operator agents to carry hard customer and employee requests through to completion.
So the practical question isn’t “which chatbot should we buy?” It’s “which job could software finish on its own, inside the systems we already use?”
How do you pick the first workflow for an agent?
Pick a job that’s repeated often, pulls from systems you can already connect, and produces output a person can check in a minute or two. Trovata’s pitch is that most treasury teams have 5 to 15 workflows ready for agents today. That’s a claim about treasury teams specifically, not a rule for every business, but it’s a useful reminder that you’re looking for a short list, not a company-wide overhaul.
Google frames the cost of paperwork in its own terms, saying clinicians lose nearly 28 hours a week to paperwork and administrative tasks. You don’t need a figure that dramatic. Ask each team lead which weekly task eats the most hours of skilled people’s time, and start the list there.
Here’s an imagined example, not a real project. A distributor’s finance lead spends Monday mornings pulling overdue invoices from the accounting system, matching them to customer notes and drafting reminder emails. That job is repeated, the data is in known places, and a person can scan the drafts in minutes. It’s a good first candidate.
- Repeated weekly or daily, with a similar shape each time.
- Data lives in two or three systems you can connect.
- Output is easy to check against a source.
- A mistake is annoying, not expensive, and can be reversed.
What controls should an agent have before it touches real systems?
Set the limits before you write a single prompt. Trovata says its agents run under role-based access control, so an agent is held to the same access limits as the people using the system. That’s a simple rule worth copying: if your bookkeeper can’t see payroll, the bookkeeping agent can’t either.
DevRev’s description adds two more. Its chief technology officer, Ahmed Bashir, says the agents’ actions can be undone and every answer links back to where it came from. DevRev’s Agent Studio also lets people describe an agent’s behavior in conversation while fixing some parts of it with set rules. Google says its healthcare workflows are bounded by guardrails it defines and include a suitable level of human review.
Put together, that gives you a checklist you can apply to any vendor or custom build:
- Permissions: the agent inherits a specific role, not a master login.
- Reversibility: every action can be undone, or needs a person’s approval first.
- Sources: each answer shows where it came from so a reviewer can verify it.
- Fixed rules: anything that must never vary, like refund limits, is a hard rule rather than a suggestion to the model.
- Human checkpoints: you decide which steps wait for approval, and you can change that later.
How do you test an agent and know if it’s working?
Test it on your own messy data, not a clean demo. DevRev says it doesn’t test its models alone. It tests them alongside the layer of software that feeds in company context, applies permissions and limits what the agent can do. Its Enterprise-Bench benchmark covers situations where a record exists in several versions, the facts are spread across separate systems, and what someone may see depends on their role. Those are the same conditions a small business has, usually with less tidy records.
Scheduled agent
- A schedule, trigger or rule fires
- Agent pulls data through a verified API layer
- Agent analyzes variances and anomalies
- Update is delivered within user permissions
Source: AWS Machine Learning
Build a set of real past cases, including the awkward ones: duplicate customer records, a missing field, an invoice that was corrected twice. Run the agent against them, compare its output with what your staff actually did, and note every miss. Then repeat with a person approving each live result for a few weeks before you loosen anything.
Track four things: time saved per run, errors caught by reviewers, cost per run, and how fast it responds. Speed matters more than people expect. Bashir says these agents “have to be able to run and complete themselves in seconds, not minutes.” An agent that takes ten minutes to answer a customer won’t get used.
Treat headline numbers with care. Trovata reports customer results such as $1M+ in annualized cost savings, and says treasury reports that used to take hours or days now take less than a minute. Those are the company’s own figures from its own customers, and they don’t predict what your business will see.
- $1M+Annualized cost savings Trovata reports for its customersAWS Machine Learning
- 5+ hoursPer day saved per accounting team, as Trovata reportsAWS Machine Learning
- 77 percentReduction in annual TMS costs, as Trovata reportsAWS Machine Learning
- 28 hoursPer week clinicians lose to paperwork, according to GoogleGoogle AI
Sources: AWS Machine Learning, Google AI, OpenAI
What are the main risks, and what should you decide first?
The biggest risk is scope. An agent with broad access and a vague job will eventually do something you didn’t expect. Keep the first deployment narrow, and widen it only after the test results justify it.
Maturity is the second. Google’s Gemini for healthcare is initially available in preview, so it isn’t yet a finished product. Whatever you evaluate, ask what’s in production today and what’s still being tested.
Third, decide who owns the agent. Someone in the business needs to review its output, update its rules and decide when to switch it off. Without a named owner, agents drift.
Finally, decide whether to buy or build. A packaged product fits when your workflow matches a common pattern and your systems are supported. A custom agent makes more sense when the work spans your own tools, such as an ERP, a customer database and a shared inbox. For that, an AI agent development partner would typically scope one workflow, connect the data sources, set permissions and approval steps, and build the test set with you before anything goes live.
What should you do next?
This week, list the five tasks that eat the most skilled-staff hours and score each on three points: repeated, checkable, reversible. Pick the top scorer, write down who’d normally do it and what they can access, and collect twenty or so real past examples. That page is the brief for any vendor or developer you talk to, and it’ll show you quickly whether an agent is the right tool. Want help with this? See our work on AI agent development and LLM integration, or book a meeting with our team.
Frequently asked questions
Do we need a large budget to try an AI agent?
Pricing varies by product and setup, and these vendor announcements didn’t publish any. What you can control is scope: one workflow, one data connection and a human reviewing every result keeps the first trial small. Decide your success measures, like hours saved and errors caught, before you start.
Should an agent be allowed to act without approval?
Only for steps that are easy to reverse and have passed testing. DevRev says its actions are reversible, and Google describes human oversight for its healthcare workflows. Start with approval on everything, then remove it step by step where the record justifies it.
Can we trust the results vendors publish?
Treat them as the vendor’s own claims. Trovata’s savings figures come from its own customers, and they don’t tell you what your workflow would produce. Run a pilot on your data and measure it yourself.
Related reading
- How to Choose an AI Voice Agent for Your Business
- Managed AI agents are maturing: what businesses should do now
Sources
- Autonomous treasury agents: How Trovata built an AI-native solution on Amazon Bedrock and AgentCore, AWS Machine Learning, 2026-11-10
- Google Cloud launches Gemini for healthcare., Google AI, 2026-11-10
- DevRev delivers autonomous self-service with GPT-6.1 Sol, OpenAI, 2026-11-04
Researched from the sources above with AI assistance, then checked against them before publishing.
