This guide is for operations and technology leaders who plan an AI agent of their own, and for buyers who want to read a vendor's self-run products as evidence. It sits under AI automation and agents, the cluster pillar and service for agents in operations.
In this guide
- What does a software company learn by running its own products with AI agents?
- Which loop do both products share?
- How do you turn the loop into your own supervised agent?
- Which lessons transfer to a client operation, and which do not?
- When should you not copy a vendor's self-run setup?
- Where AI fits
- What delivery record exists, and what does not?
- Alternatives and selection criteria
- Limits of this guide
- Frequently asked questions
- Next step
What does a software company learn by running its own products with AI agents?
It learns where an agent is useful and where it is not, at its own cost, before a client pays for the lesson. When the product and the operation belong to the same company, nobody can hide a failure behind a contract boundary: if the agent drafts a wrong fix, the team that ships the product finds it.
Two Netbase products show the pattern in public. Leadee, a product Netbase JSC builds, measures how AI answer engines such as ChatGPT, Gemini, Perplexity, Claude and Google AI Mode mention a brand, publishes confidence intervals with the results, and its Fix Engine turns those results into fix tasks. SoftWSP, also built by Netbase JSC, covers firmware and industrial software security for device makers, machine builders and plants, from firmware scan to AI-assisted fix, re-scan proof and monitoring.
Neither product is an agent that runs unattended end to end. Both turn a measurement into a bounded task, and both end in a check that proves whether the task worked. That shape, not the model inside, is the lesson.
Which loop do both products share?
The shared loop has five steps, and each step owns one control. A client operation that copies the steps, not the tools, gets most of the value.
-
Measure
Measures how answer engines mention a brand, with confidence intervals
- What SoftWSP does
- Scans firmware
- Control a client should copy
- A repeatable measurement with a stated margin of error
-
Fix task
Fix Engine turns results into fix tasks
- What SoftWSP does
- Findings feed the fix step
- Control a client should copy
- One finding becomes one bounded task with an owner
-
Fix
Not part of its registered description
- What SoftWSP does
- AI-assisted fix
- Control a client should copy
- AI drafts, a named person approves before release
-
Prove
A repeat measurement, as a method
- What SoftWSP does
- Re-scan proof
- Control a client should copy
- The same measurement repeated, not a self-report
-
Monitor
Not part of its registered description
- What SoftWSP does
- Monitoring
- Control a client should copy
- A schedule and an alert, owned in code
The middle column of each product row comes only from what Netbase has registered about the product; anything not listed there, such as an unattended autopilot mode, is not claimed here.
How do you turn the loop into your own supervised agent?
-
Pick one measurable outcome
Choose a process where a result can be measured the same way twice, such as tickets resolved, documents accepted or findings closed, and write the acceptance rule before any agent is built.
-
Make the agent produce tasks, not changes
The first version should turn a measurement into a proposed task with evidence attached; this keeps the blast radius small while the team learns what the agent gets wrong.
-
Put the approval in a named place
Every AI-drafted change goes to a named reviewer with the authority to reject it, and the rejection reason is logged so the next prompt or rule can improve.
-
Prove with the same measurement
Re-run the original measurement after the change instead of asking the agent whether it succeeded; a re-scan or re-measure is the only proof that counts.
-
Log the repairs, not only the completions
Track edits, take-overs and abandoned runs as first-class events, then use them with the operating cost per accepted task to decide whether to keep, narrow or retire the agent.
Which lessons transfer to a client operation, and which do not?
The loop transfers; the conditions around it often do not. A vendor running its own product controls the data, the release schedule and the risk appetite. A client operation answers to customers, auditors and contracts the vendor never sees.
Lessons that usually transfer:
- Measurement before automation, so the agent has a baseline to beat.
- Tasks as the unit of work, so every AI action has an owner and an audit trail.
- Human approval before release, kept even after the agent looks reliable.
- Re-check by the same measurement, so success is proven, not asserted.
Lessons that need care:
- Internal users fix agent errors quietly. Tian Pan's analysis of why dogfooding an agent is not QA describes staff correcting a wrong draft in seconds rather than filing a bug, so the task logs as complete; instrument repair events instead.
- Ownership of delegated work must stay with a person. In the Pragmatic Engineer's report on Microsoft dogfooding its AI dev tools, a Copilot task launched by a developer stays that developer's responsibility.
- Volumes and data differ. A loop tuned on the vendor's own product will need new acceptance rules, test cases and thresholds for the client's records.
When should you not copy a vendor's self-run setup?
Do not copy it when the process runs a handful of times a month, when no stable acceptance rule exists yet, or when a wrong action cannot be reversed. In those cases a checklist, a form or a plain workflow rule is cheaper to run and easier to audit than an agent. Also do not treat a vendor's own product as proof of a result in your industry: it proves the operating habit, not your outcome.
Where AI fits
Netbase works with the major commercial and open-source AI tools and models, chosen per project, and treats measurement, task routing, human approval and re-check as part of making an agent fit to operate.
- In delivered work
WhatsApp AI chatbot with CRM integration
An AI chatbot for WhatsApp Business synchronised with the client's CRM for lead capture, for a client that is not named, delivered in milestones through testing and optimisation; CRM sync makes each conversation a traceable unit.
- In delivered work
RAG knowledge assistant
A retrieval-augmented knowledge assistant delivered for a client that is not named; answers grounded in the client's own documents are the measurable unit.
- In delivered work
4over4's recommendation engine
It suggests products from browsing history and previous purchases.
What delivery record exists, and what does not?
- What exists. Netbase designs, builds, integrates and runs software across technologies and industries, and builds its own products, including Leadee and SoftWSP described above; the directory of tools Netbase JSC builds lists them. Beyond the 4over4 recommendation engine, Netbase has delivered AI projects for clients that are not named, including retrieval-augmented knowledge assistants, document AI and MLOps pipelines. Netbase runs, maintains and upgrades what it builds, remotely and alongside the client's team, under an agreed support scope.
- What does not. No published Netbase record states a saving, a resolution rate, a fix rate or a time figure for Leadee, SoftWSP or any client agent. The loop above is a method, not a reported result. Most Netbase projects are delivered on fixed-price contracts, with milestone-based, monthly team retainer and KPI-linked models also offered.
Alternatives and selection criteria
There are three common ways to get a supervised agent into operations, and the right one depends on who should own the loop after launch.
- Build it in-house. Choose it when you have engineers who can own measurement, prompts, evaluation and on-call; the risk is that the team learns the loop slowly on live work.
- Buy a platform and configure it. Choose it when your process matches the platform's model closely; check who owns the traces, the acceptance rules and the exit path.
- Have a partner build and run it. Choose it when you want the loop working before your team takes it over. Co-build and build-operate arrangements are agreed deal by deal with the CEO and never published; see build-operate-transfer for software products and the engagement models.
Selection criteria for any of the three: a written acceptance rule per task type, a named approver, a re-check that repeats the original measurement, logged repair events, and a documented way to switch the agent off.
Limits of this guide
- It describes an operating pattern from two live products, not a benchmark; your process, data and regulators set the real thresholds.
- It names no saving, rate or payback, because none is registered for these products.
- Third-party observations come from the cited sources and describe those companies, not Netbase.
Plan the next step with a Netbase consultant
Frequently asked questions
Not automatically. Running its own products shows the vendor has lived with an agent's failures, but you should still ask for the acceptance rule, the approval step and the re-check it would build for your process.
No claim of unattended operation is made. Leadee turns measurements into fix tasks, and SoftWSP moves from scan to an AI-assisted fix with re-scan proof and monitoring; the pattern keeps a measurement and a check around every AI step.
Ask what the product measures, how a result becomes a task, who approves an AI-drafted change, how success is re-checked and which repair events are logged. Vague answers to these questions are a warning sign.
Yes, if one process has a stable acceptance rule. The loop needs a measurement, a task queue, a named reviewer and a repeat check, which many teams can build on the tools they already run.
Next step
Tell us which process you want an agent to run and how you measure it today, and we will book a solution review to set out the measure, task, approval and re-check loop for it. You can also see the enterprise AI agent platform.
Related services and solutions
AI automation and agents that keep people in charge Netbase provides AI automation and agent development for operations teams that want repetitive, multi-step work done by software while people keep approval over decisions that matter. We combine rule-based automation with AI where it adds value, build in human approval, and measure return against a baseline.
Learn More
Enterprise AI agent platform: run AI agents with approvals, audit logs and cost limits An enterprise AI agent platform is a governed runtime where AI agents plan and carry out multi-step work through approved tools, while approvals, audit logs, access rules and cost limits keep every action accountable. Netbase offers it as a forward-looking solution, built inside your own environment rather than licensed as a product.
Learn More
Work with Netbase remotely
Every Netbase engagement is delivered remotely, so clients in any country and any industry can work with us; Netbase has worked with clients in 80 countries. Contracts and invoices are issued by Netbase JSC, and payment is made by bank transfer to Netbase JSC's company account or PayPal Business. Talk to the CEO, Huy Nguyen (David), on WhatsApp or by email.
Discuss a project
Netbase JSC helps organizations design, build, modernize, and operate digital products and AI-enabled business systems.project@netbasejsc.com
+84 937 869 689
91 Nguyen Chi Thanh, Dong Da, Hanoi, Vietnam
Get in touch
Tell us what you want to build, modernize, or operate.