The Newest Asset Is the Agent.
and Trust in AI-Driven Service
About the Authors
Dave Rosenlund
Trundl
Global Head of Software and Solutions
Atlassian Community Champion
Dave Rosenlund is Global Head of Software and Solutions at Trundl, a Platinum Atlassian Solution Partner, and an Atlassian Community Champion. He works at the intersection of enterprise service management, AI adoption, and Atlassian platform strategy, and writes the “AI That Works” and “Beyond ITSM” series on the Atlassian Community, among the platform’s most-read practitioner perspectives on AI in service operations.
Gary Blower
Lansweeper
Atlassian Solution Architect
Gary Blower is an Atlassian Solution Architect at Lansweeper, with more than twenty years across service management and the Atlassian ecosystem. He works hands-on with Jira Service Management, Assets, and Rovo, with deep expertise in configuration management and asset and security tracking, and is a frequent speaker in the Atlassian Community.
MaSonya Scott
Atlassian
AI & Service Management Evangelist
MaSonya Scott is an AI & Service Management Evangelist at Atlassian, with more than twenty-eight years across enterprise service management and technology, including roles at AWS and Accenture. She is a widely heard public voice on Service Collection, Assets, and Rovo, helping enterprise and regulated organizations adopt AI with the governance to trust it.
Adoption is ahead of trust
AI is already in production. The binding constraint is trusted context, not more capable models.
Adoption is running ahead of trust
Artificial intelligence is no longer a horizon technology. It is inside enterprise service workflows today, routing incidents, drafting responses, classifying risk, and making calls that used to take a room of people an afternoon.
Yet most leaders will admit the same thing in private. Adoption is running ahead of confidence.
Adoption is running ahead of confidence.
Atlassian’s State of Teams 2026 research shows the gap from two sides: roughly 85 percent of knowledge workers now use AI, while only 6 percent of executives report clear, organization-wide return on their AI investment. Broad adoption on one hand, real difficulty proving enterprise-wide value on the other. Our argument is that for many deployments the binding constraint is no longer capable models. It is whether the AI has trusted context and a workable process to act on. That is the distance between AI that is used and AI that is trusted.
Adoption is ahead of value
85%
of knowledge workers now use AI.
(Atlassian, State of Teams 2026)
6%
of executives report clear, organization-wide return on their AI investment.
(Atlassian, State of Teams 2026)
74%
of security leaders have traced an incident back to an asset nobody was managing.
(Trend Micro, 2025)
Atlassian saw the same pattern inside its own walls, and wrote it up in Leading with Context. Rewiring how 14,500 employees work with AI did not come down to more models or wider access. It came down to context: making sure the information the AI reasons over is structured, current, and connected to how the work flows. The teams that got real value did not just adopt AI. They rebuilt how they work around it.
Which sets up the question this paper is about. An AI answer at two in the morning, or a Rovo-assisted change at any hour, is only ever as good as what sits beneath it. And by the end, the newest thing you will have to govern beneath it is the agent itself.
So what sits beneath it?
Switch on the AI, and you inherit your data’s problems
The common move is to enable the agent first and fix the context underneath it later.
The order is backwards.
An agent only knows the environment your systems expose to it. For most organizations a large part of that environment lives in a configuration management database that is stale, partial, or quietly wrong, and the agent has no way to tell. Give it a record nobody has trusted in years and it will turn that record into fast, confident, wrong output.
One of us has written about this under a blunter heading: the CMDB is lying to AI. The technology is not broken. The AI is faithfully repeating what the data tells it, and the data is out of date. Garbage in, garbage out is the oldest rule in computing. What is new is the speed and the confidence. The agent does not just inherit the bad record. It acts on it, at machine scale, and treats the result as settled.
So the work is not switching the AI on. The work is everything underneath it. That foundation has two halves: the half you can discover, and the half nobody wrote down. The rest of this paper is about both.
The two halves of the foundation
What an agent stands on is the record you can discover
and the context nobody wrote down.
The discoverable half: getting the
record right
Start with the half you can find.
Good asset intelligence begins with continuous discovery across the estate: devices, servers, cloud resources, and installed software, registered or not, found as broadly and continuously as practical rather than once a quarter. Finding what exists, then normalizing and enriching it, hardware, software, vulnerability, and dependency data alike, before any of it reaches Assets. That discovery is the foundation the rest of Atlassian Service Collection (Jira Service Management, Customer Service Management, Assets, and Rovo) stands on, and it is precisely the job Lansweeper is built for.
But discovery only tells you what exists and what it looks like. It does not tell you which services matter, who owns them, or which dependencies are business-critical. Worth remembering, because the instinct that follows discovery is usually the wrong one. Feed everything you find straight into Assets, on the theory that more data means more visibility, and you get the opposite: a record nobody trusts.
The discipline is to discover broadly and keep the record narrow. Discover the whole estate; security, compliance, software asset management, and endpoint teams all depend on broad inventory. Keep narrow only the governed configuration model the agents reason over: the objects, attributes, relationships, and services that carry decision value.
Start with a bounded set of high-value services, one or two to begin with, as Atlassian’s own guidance suggests, and extend only when the next service adds more decision value than it costs to keep current. In our experience the returns from adding more fade earlier than teams expect.
In one customer use case, a US healthcare network preparing to move its service platform to the cloud found that the first real task was not moving data but deciding what deserved to come along. Its configuration database had swollen past the point of usefulness, and the opening work was subtraction: cutting more than 30,000 objects to bring the record back under a workable threshold, and splitting one overloaded schema into a physical-inventory view and a separate configuration view. A record earns its keep by what it leaves out as much as by what it holds.
A record earns its keep by what it leaves out as much as by what it holds.
The same restraint applies to attributes. Copy every field a discovery tool can offer, and someone now has to keep all of it current, or it becomes noise. In one environment, a single unmaintained field turned one asset into the most unreliable device on record, 750 incidents logged against it, because staff gave up finding the right one and defaulted to whatever was in front of them.
There is a second gap, and most organizations live with it quietly. Ask the IT team and the security team how many assets they have and the numbers will not match. In one anonymized client estate, the CMDB counted 11,982 assets and the security tooling counted 11,481: same environment, two teams, two realities. The gap was less error than scope, IT counting for uptime and security for exposure, but until someone reconciles the two, neither can act with confidence, and that pause is where risk compounds. Independent research puts a number on the cost: 74 percent of security leaders say they have traced an incident back to an asset nobody was managing, according to a 2025 global study commissioned by Trend Micro.1
The fix is not one repository but one reconciled reality: a shared asset identity, clear system-of-record rules, and a trusted operational view both teams work from. Give IT and security that, and a vulnerability lands against an asset IT already recognizes, with an owner IT already knows, instead of starting the round of whose-device-is-this that usually follows.
None of this decides itself. Someone chooses which services matter and which attributes earn their place, and if that reasoning lives in one person’s head, the record drifts the day they go on leave.
That decision, and the reasoning behind it, has to be written down where the next person, or the next agent, can use it. We will come back to how.
One test tells you whether any of this is working, and it is not an asset count and not an SLA. Service Collection can hit every service-level target in the book while the person who raised the ticket has a miserable time, because an SLA measures the process, not the experience. A trusted record is necessary for a better experience. It does not produce one on its own. But get the record wrong and a poor experience is guaranteed, whatever the service-level report says.
That is the half you can discover. The harder half is the one no scan reaches.
The half no scanner reaches
Concede the point first. Accurate asset intelligence is necessary.
It is not sufficient.
Beyond the assets a scanner can find sits the context it never will. What really depends on what. Why a particular change is risky. The workaround that keeps a fragile service alive and lives only in one engineer’s memory. Scanners find the assets. They do not find the judgment.
Scanners find the assets. They do not find the judgment.
We have called this the unwritten layer, and named the people who hold it together the human compensation layer, and made the fuller case for both in the companion paper Portable Context. Here it is enough to name it: the tacit half of how an organization runs, undocumented because it never needed documenting, until an agent inherits the gap and answers from it anyway.
This is where data quality stops being IT housekeeping and becomes a governance question. For an agent to be trusted, the knowledge it draws on has to be reliable, relevant, and current. An agent working from documentation nobody has reviewed in a year will route to owners who left, cite policy that has changed, and summarize a stale knowledge base into a confident wrong answer. The governance framework said yes. The data said not yet.
It is worth being honest about the boundary here. Trusted context, the discoverable record and the unwritten layer both, is a foundation for reliable agents. It is not the whole of trust. An agent with perfect context can still be handed permissions it should not have, act outside the policy that governs it, or drift without anyone watching. Trust also depends on what the agent is allowed to see and do, the rules it runs under, how you know it is working and can reconstruct a failure, and who owns the outcome when it is wrong. This paper concentrates on context because it is the layer most often skipped and the one the others stand on: you cannot set sensible permissions, or judge whether an agent behaved well, against a record you do not trust. Two of those other layers, authority and accountability, come back on their own once the agent itself becomes something you have to govern.
So the foundation has two halves that fail in different ways. The discoverable half fails when the record is bloated or contested. The unwritten half fails when the context that matters was never captured at all. Unless the difference is made explicit, an agent treats both as equally true.
The foundation
Two halves that fail in different ways
The
discoverable
half
The record you can scan.
Fails when it is bloated or contested.
The
unwritten
half
The context no scanner reaches, what depends on what, why a change is risky, the workaround in one engineer’s memory.
Fails when it was never captured at all.
Which raises the real question. If half of what an agent needs was never written down, how do you write it down without turning the whole organization into a documentation project?
Building the capability
Making context something the platform can build,
hold, and keep current.
From context to capability
The answer is not to know more. It is to make context something the platform
can hold and keep current.
Two disciplines carry that load.
The first captures knowledge as a by-product of the work rather than as a separate task nobody has time for. Knowledge-Centered Service, a practice the service community has refined for roughly three decades, does this for the knowledge that flows through service work: the resolution, the workaround, the article that saves the next person an hour, written down in the doing rather than after it. The governance decisions behind the record itself, which service to track, which field to retire, which configuration item to stop maintaining, are a slightly different job, best captured as lightweight decision records that keep the reasoning with the decision.
Together they answer the question the last section left open: how the reasoning behind a good record survives the person who made it. Skip both and a well-designed record decays the same quiet way an unmaintained field does.
The second discipline treats an agent less like software you install and more like a new hire you train. You do not hand a new analyst the keys on day one. In practice that means the controls you would give any capable newcomer: scoped permissions, a set of cases you test it against, a supervised rollout, and monitoring that widens its authority only as it earns trust. Level it up on real context and validate it against real cases before it goes live, rather than switching it on and hoping.
None of this is specific to one vendor. Any organization serious about AI-ready service management needs a way to build platforms from its real context rather than a template, and to keep that context governed and current. Trundl builds this way. Contextual Delivery is the operating model, its method for building a platform from an organization’s real context; Rapid Deploy is the delivery platform Trundl’s solution engineers use to execute it, turning codified context, existing configurations, runbooks, and operational knowledge into governed platform configuration the customer’s teams then own and refine. The point is not speed. It is that AI-ready context has to be built and maintained, not written down once and left to rot.
AI-ready context has to be built and maintained, not written down once and left to rot.
Get this right and the foundation stops being a pile of accurate facts and becomes something an agent can reason over.
What it looks like in practice
Put the pieces together, the discoverable record, the unwritten layer, and an operating model to keep both current, and the operating picture takes a concrete shape.
Lansweeper · Service Collection · Trundl
The trust chain, from discovery to a governed agent
Consider one path through it, drawn as an illustration rather than a specific customer. Lansweeper discovers a newly exposed vulnerability and the asset it sits on. Because that asset is already normalized and matched in Assets, it arrives with an owner and a service attached rather than as an orphan. Rovo and Jira Service Management triage it against that context, propose a response, and route it for human approval instead of acting blind.
The action is taken, the record and the reasoning are updated as the work happens, and the result is measured against whether the service, and the person behind it, are better off. Every step leans on the foundation underneath it. Break the record and the vulnerability lands on an orphan asset, and the whose-device-is-this round begins.
The foundation beneath that sequence is not hypothetical. Agropur, a North American dairy cooperative with more than 7,000 employees, connected Lansweeper’s discovery directly into Service Collection’s Assets so that every device, patch, and piece of unauthorized software carried a record before a ticket was raised, and stood the foundation up on a five-month deadline. Agropur’s David Rossignol, director of technology adoption and support, said so in the published customer story: “The combination of Jira and Lansweeper provides a strong ITSM foundation.” The value is not the reference itself. It is that the record-first foundation is one teams run in production, not a whiteboard ideal.
Real environments are messier than one clean path. Discovery sources disagree, records duplicate, syncs lag, ownership is disputed, and a connector fails now and then. Keeping one trusted reality is partly the work of reconciling at those seams, which is why a narrow, governed, shared record matters more once AI is consuming it at speed, not less.
The investment has to cover both sides. Spend on the discoverable half and you get a record the platform can trust. Spend on the unwritten half and you get the context that record can never hold on its own. Skip either and the AI inherits the gap.
Governing the agent
A material production agent is a configuration item.
Govern it like one.
The newest asset is the agent
Here is where the argument turns forward.
For decades a configuration management database existed because complex environments have dependencies that stay invisible until something breaks. A server is a configuration item. An application is a configuration item. A critical dependency is a configuration item. So consider the agent you just switched on.
Not every agent belongs in that record. A tool that drafts a suggested reply carries far less risk than one that can act on production unattended. The threshold is materiality: how much autonomy the agent has, how much business impact it can cause, what data it can reach, whether it can execute changes on its own, and whether it runs in an environment the business considers critical. Below that line, lighter tracking is fine. Above it, the agent has crossed into territory the CMDB was built for.
Concretely: an agent that drafts a reply for a person to send is low materiality. One that auto-triages and routes without touching production sits in the middle. One that can execute a change in production on its own is high, and it is the kind that belongs in the record.
From low to high
The materiality ladder
Low materiality
Low
Drafts a reply for a person to send.
Medium materiality
Medium
Auto-triages and routes without touching production.
High materiality
High
Executes a change in production on its own. This is the kind that belongs in the record.
Levels are illustrative. Materiality also turns on data sensitivity, business process, reversibility, and consequence, not autonomy alone.
A material production agent takes actions with business impact. It depends on upstream data and downstream workflows. It changes behavior when it is retrained or re-prompted. It can fail, degrade, or produce something unexpected under load. It may face regulatory scrutiny that demands proof it was authorized to act.
That is not a pilot. That is a production asset, and the newest configuration item in your environment is the agent itself. The CMDB that was lying to AI at the start of this paper now has to hold the AI as one of its own records.
The newest configuration item in your environment is the agent itself.
Be precise about scope. Enterprises model configuration items differently, some as a logical agent with separate deployments, some by environment. What matters is that at minimum the material production deployment, the instance that can actually act, carries a governed record. And reauthorization is not annual paperwork; it is triggered by material change: a new model or version, changed instructions or tools, a new permission, or a new data or knowledge source. How much scrutiny each earns depends on its risk. Adding production write access is not a prompt tweak, and the review should not treat them the same.
Treating a material agent as a configuration item is not bureaucracy. It is what makes AI in production defensible, to customers, to auditors, and to the board. A workable standard covers what you would track for anything else that can act:
The agent record
A workable standard
| Attribute | Why it matters |
|---|---|
| Traceability and change management | |
| Accountability | |
| The guardrails, documented | |
| What is really doing the work | |
| Data quality and reliability | |
| Resilience and continuity planning | |
| Compliance and audit evidence | |
and on whose approval | Observability and investigation |
| Operational readiness | |
| Governance as an ongoing thing, not a one-time sign-off |
Three questions belong at the center of any deployment review for an agent you now govern like any other asset. First, transparency: can you reconstruct what it did? If a regulator or a board member asked why the agent took a specific action, could your team show what it received, retrieved, invoked, and changed, and on whose approval, in plain language, rather than “the model decided”? This is reconstruction, not a promise to explain the model’s inner reasoning. Second, authority: what is it not allowed to do? Every trusted agent has hard limits, an escalation path, and a way to stop it if it starts failing quietly. Third, accountability: when it is wrong, whose phone rings? The answer cannot be “the algorithm did it.” It has to resolve to a named human who owns the outcome. Call that last one the no-algorithm-shrug rule.
Three questions for any deployment review
The AI Trust Triad
Transparency
Can you reconstruct what it did? Show what it received, retrieved, invoked, and changed, and on whose approval, in plain language.
Authority
What is it not allowed to do? Hard limits, an escalation path, and a way to stop it if it starts failing quietly.
Accountability
When it is wrong, whose phone rings? It resolves to a named human who owns the outcome. The no-algorithm-shrug rule.
MaSonya has a name for the three together, the AI Trust Triad. They are the shortest honest checklist a leader can carry into a deployment review, and they are worth documenting before an agent goes live, in something as light as a one-page trust brief, so they are answered ahead of the review rather than after the incident. The reconstructability line in the table is what lets someone answer the first question after the fact rather than only in the moment.
The horizon
Foundation quality has leverage well beyond the
service desk.
The horizon, and the price of
not knowing
One more turn, because the stakes are getting larger than any
single product.
Assets is no longer only a Jira Service Management feature. Atlassian now treats it as a platform-level source of context that feeds the Teamwork Graph, which grounds Rovo and, increasingly, other agents across the platform rather than inside Service Collection alone.2
The mechanics will keep shifting, but the strategic point does not: the better your asset-and-context foundation, the more the whole platform can do with it, and the further behind you fall without it. Foundation quality has leverage well beyond the service desk.
Which is the note to end on. The service leaders who win with AI will not be the ones who bought the most of it. They will be the ones whose foundation was ready for it, who did the unglamorous work of a trusted record and a captured context before they turned the agents loose.
Where to start is the least glamorous move, and the one with the most leverage. Before you widen an agent’s scope, audit the context it will stand on, because stale context is common and hard to see from the output.
Three things turn that into a foundation you can trust. Continuous discovery, so the record reflects what you actually run, which is the role Lansweeper plays in this architecture. A trusted record and the unwritten layer captured inside Service Collection, so agents reason over context you trust. And an operating model that builds and maintains both rather than writing them down once, which is what Trundl’s Contextual Delivery and Rapid Deploy are for.
Get those right, and the agents finally have something solid to stand on.
AI did not remove the need to know your environment. It raised
the price of not knowing it.
References
[1] Survey of 2,250 IT and security professionals across 21 countries, conducted by Sapio Research for Trend Micro, 2025.
[2] Atlassian reports that grounding responses in the Teamwork Graph produced 44 percent more accurate answers using 48 percent fewer tokens (vendor-reported). On availability: Forge connectors that bring outside data into the graph are generally available, while the interface for querying the graph directly remains in early access with production limits.
About Trundl
Trundl is an enterprise delivery technology company, founded in 2014 and headquartered in San Jose, California, with operations across the United States, Canada, and India. An Atlassian Platinum Solution Partner and a monday.com Silver Partner, Trundl is ISO 27001 certified and trusted by more than 400 customers worldwide. Its Contextual Delivery operating model and Rapid Deploy platform build enterprise platforms from an organization’s real context. Learn more at trundl.com.
About Lansweeper
Lansweeper is the leading AI Cyber Asset Intelligence Platform, giving organizations one trusted, continuously validated view of every IT, OT, cloud, and IoT asset. It turns fragmented asset data into a shared source of truth that IT and Security can align around and act on with confidence, powering safer automation and more reliable AI-driven decisions. Over 30,000 environments worldwide trust Lansweeper. Learn more at lansweeper.com.