One list makes your AI agents compound. Nearly every organisation now runs agents in production, and the research published this month suggests the next gain comes not from building more of them but from putting the ones you already have in a single place. On 22 September 2026 Guild.ai published The AI Agent Management Gap, a Morning Consult survey of 362 IT decision makers at companies with 100 or more employees, fielded between 4 and 9 August. Its headline finding is genuinely good news. Ninety-six per cent of organisations report having AI agents in production, and 47 per cent of those are running dozens or more. Agents have stopped being a pilot. They are now a working population, and populations behave differently from projects.
That shift is where the opportunity sits. The same survey found that 61.2 per cent of engineering and IT teams lose time rebuilding or maintaining agents that already exist somewhere else in the organisation. Read as a problem, that is duplicated effort. Read as an opportunity, it is a stock of finished work that nobody has catalogued yet. Most organisations have already paid to build the lead qualification agent, the case summarisation agent and the quote checking agent at least once. A shared list turns each of those from a one-team asset into a building block the next team starts from, which is the cheapest capability increase available this quarter.
The market has already decided it wants this. Nearly eight in ten respondents, 79.7 per cent, expect to increase investment in agent management infrastructure, 83.9 per cent call real-time visibility of agent activity extremely or very valuable, and 81 per cent rate a centralised management hub as important. Confidence is high alongside it: 93.2 per cent feel prepared for their long-term agent strategy. For any organisation running a CRM at the centre of its revenue operation, that is an unusually clear signal about what to build next, and the first version of it is a list.
What did the 2026 AI agent management survey find?
The AI Agent Management Gap, published by Guild.ai on 22 September 2026 and conducted by Morning Consult among 362 IT decision makers at organisations with at least 100 employees, found that AI agents are now close to universal in production while the tooling that manages them is still arriving. Ninety-six per cent of organisations report agents in production and 47 per cent run dozens or more. On the management side, 42.7 per cent have a centralised dashboard or monitoring tool, 39.8 per cent keep logging or audit trails, 35.9 per cent require an agent to be registered before deployment, and 31 per cent have an automated way to stop a malfunctioning agent immediately, with 76.5 per cent able to respond within minutes by automated or manual means. The margin of error is plus or minus five percentage points.
The number that drew most of the coverage was a confidence gap. Some 96.4 per cent of IT decision makers said they were confident their organisation holds a complete and accurate inventory of its agents, while 66.7 per cent of organisations with agents reported an agent-related operational consequence in the past twelve months. That reads at first as overconfidence, but there is a more useful interpretation available. People are confident because each of them can see their own agents clearly. What nobody can see is the whole set at once. That is a tooling gap rather than a discipline gap, and tooling gaps are the kind that close quickly once somebody owns them.
What is shadow AI when almost every agent is sanctioned?
Shadow AI is any AI tool or agent used for work without the knowledge or approval of the people accountable for data, security and compliance. In 2026 the shape of it has changed. The Guild.ai survey found that 60.6 per cent of respondents think employees are likely deploying agents without approval, while only 35.9 per cent require registration before deployment. The point worth drawing out is that most of those agents are not rogue. They are useful work done by capable people who had no route to register what they built, because the route did not exist yet. Shadow AI today is largely an administrative gap rather than a behavioural one, which makes it considerably easier to close.
The constructive move is to make registration worth doing rather than merely mandatory. Microsoft’s Agent 365 registry, which took over from the Entra agent registry blades on 1 May 2026 and now sits in the Microsoft 365 admin centre, lists agents built in Foundry, Copilot Studio, Agent Builder and SharePoint alongside agents from Amazon, Google and Salesforce in one catalogue. The design lesson there is the inclusiveness. A registry that accepts only agents built on one platform records a fraction of reality. A registry that accepts anything, and gives the builder a visible entry, a named owner and a reusable description in return, gets filled in voluntarily. As we found when looking at who is actually building enterprise agents, the people doing the work are usually delighted to have it recognised.
Reuse is the fastest return on the agents you already run
Reuse is the fastest return because the expensive part of an agent is rarely the model. It is the scoping, the data mapping, the action permissions and the testing against real records, and every one of those is already paid for on an agent that works. When 61.2 per cent of engineering and IT teams report losing time to duplicative agent work, the implied inventory of finished, tested, unlisted components is substantial. A catalogue entry recording what an agent does, which systems it touches, who owns it and what it may change converts that sunk cost into something the next team can adapt in an afternoon rather than specify from scratch over six weeks.
This compounds in a way that building more agents does not. Each registered agent makes the next one cheaper, because the hard decisions it encodes, what a qualified lead is, which fields an agent may write, when a person signs off, are reusable across every agent touching the same objects. A sales team that has already agreed what its quoting agent may change has done most of the work for its renewals agent. That is why a list is a better next investment than another build. It improves the economics of everything already shipped, including the agents nobody has looked at since launch.
How do you measure CRM ROI across a portfolio of agents?
Measure CRM ROI on a portfolio of agents at two levels. At the agent level, track the share of a defined workload each one carries end to end, the time saved per completed task, and the quality of what it hands back, the same measures you would apply to any process improvement. At the portfolio level, track the three figures that only exist once you have a list: how many agents are registered, how many of those were built by starting from an existing one, and how much of the population has a named owner. The second is the number that shows a registry paying for itself, because every reuse is a build you did not have to fund.
The portfolio view is also what makes the business case legible to a finance function. A single agent saving two hours a week is a rounding error in most budgets. Forty agents, each with a measured workload share and a known owner, is an operating capability with a cost line and a return, and it can be reviewed like one. Organisations that reach that point tend to find the conversation changes from whether to fund AI work at all to which part of the portfolio to extend next, which is a far better conversation to be having in a planning cycle.
Why does CRM data quality decide which agents can be reused?
CRM data quality is important because an agent’s usefulness is bounded by what it can reliably read, and its reusability is bounded by whether another team’s records look like yours. An agent that qualifies leads well in one region because that region fills in industry and employee count consistently will underperform in a region that leaves both blank, and the fault sits in the data rather than in the agent. This is why the registry and the data work belong together. Recording which fields and objects each agent depends on produces, as a by-product, a ranked list of the records that matter most to your revenue operation, which is the most practical data quality backlog an organisation can have.
That by-product is worth more than it sounds. Most data quality programmes stall because the scope is the whole CRM and the benefit is abstract. A dependency list inverts both. It tells you that eleven agents rely on the account industry field and two rely on a custom renewal date, which makes the sequencing obvious and the benefit specific. Teams that work this way report the same pleasant surprise: the cleanup is smaller than feared, because the fields agents genuinely use are a narrow slice of the fields a mature CRM contains.
Registering agents stops CRM technical debt before it forms
CRM technical debt is the accumulated cost of customisations, integrations, automations and workarounds that still run but that nobody has documented, reviewed or decided to retire. It is usually discovered years later, during a migration or an upgrade, when understanding what exists costs more than rebuilding it would. Agents are the newest contributor to it and also the easiest to handle well, because the population is young. An organisation registering its agents in 2026 is documenting a two-year-old estate rather than running archaeology on a twenty-year-old one. The ordinary cost of good hygiene is lowest right now, and it never gets lower.
The practical version is modest. Give every agent an owner, a one-line purpose, a list of the systems and fields it touches, a note of the actions it may take without a person, and a review date. Five fields. That is enough to make the estate searchable, enough to answer an auditor, and enough to let the next builder find prior art. The 35.9 per cent of organisations already requiring registration before deployment have, without framing it that way, started the cheapest technical debt programme available to them, and the 79.7 per cent planning to invest in management infrastructure are about to join them.
The Sirocco perspective
We work across Salesforce, HubSpot and Microsoft Dynamics 365, and the registry question is one where independence from any single vendor genuinely changes the advice we give. Each platform is building a good catalogue of the agents built on itself, and Microsoft’s has already reached beyond its own boundary. In practice, though, most of the organisations we work with run agents on at least two platforms plus something built in-house, and the list that helps them is the one spanning all of it. That is the clearest reason to use an independent CRM partner on this particular piece of work: a vendor’s catalogue is naturally shaped by what that vendor can see, while the useful inventory is shaped by what your organisation actually runs.
Our advice to clients this quarter is unglamorous and reliably effective. Start the list before you need it, keep it to five fields, and make the first entry the agent somebody is most proud of. A registry that begins as a showcase gets filled in. One that begins as a compliance exercise does not. The organisations starting now will spend next year reusing what they have already built, while the cost of cataloguing is still measured in hours rather than quarters.
If you are deciding where your next AI investment should go, or working out how to see the agents you already run across more than one platform, we would be glad to talk it through. You can schedule a consultation with our team.
Get in Touch
If you are ready to build one view of every AI agent running across your Salesforce, HubSpot or Dynamics 365 estate, tell us what you are running today and we will help you shape the first version of that list.
