Something quietly excellent happened inside enterprise software over the first half of this year, and the number that captures it landed on 6 August. Opsin Labs, examining production environments across eight industries between March 2025 and June 2026, found that 67 per cent of the AI agents now running inside enterprises were built by people with no engineering background at all. Go-to-market teams. Customer success. Operations. The people who actually run the process are now the people automating it.
That is the outcome CRM programmes have been chasing for the better part of two decades, and it arrived faster than almost anyone forecast. The same research found that enterprise environments now average roughly one agent per employee, live or in draft, and that workforce interactions with agents grew fourteenfold between January and June 2026. The capability is real, and the teams getting the most from it are pairing that new build freedom with one small discipline at the moment of creation: giving each agent access to exactly the job it does, and nothing more. That single habit is what separates a fleet of useful agents from a pile of abandoned experiments, and it is far easier to establish now than to retrofit later.
Who is actually building enterprise AI agents in 2026?
Enterprise AI agents are increasingly built by business users rather than developers. The Opsin Labs State of Agentic Adoption report, published on 6 August 2026 and drawn from sixteen months of production data across eight verticals, found 67 per cent of agents were created by staff without an engineering background, most commonly in revenue, customer success and operations roles. Platform vendors have made this possible on purpose: HubSpot opened Agent Builder to Professional and Enterprise tiers, Salesforce ships Agentforce builders to admins, and Microsoft puts agent creation inside Copilot Studio alongside the Power Platform tools business teams already use. Building an agent is now closer to configuring a workflow than to writing software.
The scale of it is the part worth sitting with. An average of one agent per employee is not a pilot programme, it is a working population. And unlike the wave of departmental tooling that preceded it, this one is mostly happening inside the sanctioned platforms, where the audit trail already exists. That is a much better starting position than most organisations had with spreadsheets, macros or unmanaged integrations.
Why business-built agents lift CRM adoption
How do you improve CRM adoption rates? The most reliable lever has always been reducing the distance between the people who feel a process is broken and the people allowed to fix it. Adoption falls when a seller logs a request with IT, waits a quarter, and receives something that answers last quarter’s problem. It rises when the person who noticed the friction can resolve it inside the same week. Agent builders inside Salesforce, HubSpot and Dynamics 365 collapse that distance, which is why business-built agents tend to get used: the person who built it already wanted it.
There is a second adoption effect that is easy to miss. An agent that drafts follow-ups, summarises an account before a call or chases a missing close date makes the CRM feel like it gives something back rather than only taking input. Sellers have complained for years that the system is a reporting tax. When the record starts producing work instead of just absorbing it, the incentive to keep the record accurate changes completely. That is the flywheel worth engineering for, and it is now within reach of teams who never had a development budget.
This is not shadow AI, and the difference is the opportunity
What is shadow AI? Shadow AI is the use of AI tools and models outside the organisation’s sanctioned, governed environment: staff pasting customer data into a consumer chatbot, teams buying an unapproved assistant on a corporate card, or an automation running against company systems with credentials nobody has reviewed. It is invisible to IT by definition, which is what makes it hard to manage. Business users building agents inside the CRM platform is the opposite situation. The activity is sanctioned, it is logged, the identity is known and the permission model is already there.
That distinction is genuinely good news, and it deserves more attention than it gets. The hardest problem in AI governance has always been discovery: you cannot govern what you cannot see. Organisations whose agents are being built inside their CRM and productivity platforms have skipped that problem entirely. What remains is configuration, which is a solvable, well-understood piece of administration rather than an intractable visibility gap. Plenty of companies are still hunting for their shadow AI. If yours is being built in the open, you are starting several steps ahead.
Scope each agent to the job, not to the builder
Here is the one adjustment worth making early. The same Opsin research found that among agents provisioned beyond default settings, 60 per cent were given broad allow-all access rather than permissions matched to their task, and 60 per cent ended up with capabilities exceeding their original stated intent. The cause is mundane rather than alarming: an agent inherits the access of whoever built it, because that is the fastest route to a working prototype, and nobody revisits the choice once it works.
The fix is equally mundane, and every major CRM already supports it. In Salesforce, give the agent its own user and a permission set built for the task rather than cloning an administrator profile. In Dynamics 365, define a security role scoped to the tables and columns the agent genuinely touches. In HubSpot, set app scopes and user permissions to the objects in play. This is perhaps a day of work per agent pattern, and it pays for itself beyond security: a scoped agent produces cleaner audit logs, makes usage reporting meaningful, and is far easier to hand over when the person who built it moves on. That is work most teams would benefit from doing anyway.
Why CRM data quality decides what your agents can do
Why is CRM data quality important when agents enter the picture? Because an agent’s usefulness is capped by the data model it reads. An agent asked to flag at-risk renewals can only do so if renewal dates are populated and consistent. An agent that routes enquiries needs an ownership field that reflects who actually covers the territory. When the underlying records are complete, an agent built in an afternoon by a customer success manager will outperform a far more sophisticated system running on partial data. Data quality is the multiplier on every agent an organisation builds.
The encouraging part is that agents make this investment visible in a way that dashboards never did. A stale field used to produce a slightly wrong chart that everyone learned to discount. Now it produces an agent that misses a renewal, and the gap gets fixed within days. Teams that have spent years arguing for data quality budget are finding the argument suddenly makes itself, because the cost of a blank field has become concrete and immediate rather than theoretical.
Permission sprawl is technical debt you can pay down early
What is CRM technical debt? It is the accumulated cost of shortcuts taken inside a CRM platform: redundant custom fields, overlapping automations, unused integrations and permissions granted for a one-off reason and never withdrawn. Each shortcut is rational on the day it is made. The cost arrives later, as slower releases, harder upgrades and a system nobody fully understands. Broad agent permissions are simply the newest entry on that list, and they behave exactly like the older ones.
The difference this time is timing, and it favours you. Most organisations are twelve to eighteen months into agent building rather than a decade in, which means the cleanup is measured in dozens of definitions rather than thousands. Setting a standard now, one scoped role per agent pattern, a named owner, and a review date, costs a fraction of what the same exercise will cost in 2028. Teams that establish the convention while their agent population is still young get to keep building at speed without accumulating the drag. That is a rare chance to get ahead of a debt curve rather than climb out of one, and it is available right now.
The Sirocco perspective
Why use an independent CRM partner for work like this? Because the pattern we are describing is the same on every platform, while the implementation details are entirely different. We work across Salesforce, HubSpot and Microsoft Dynamics 365, and we see the same encouraging shift in all three: the people closest to the customer are building the automation, and it is working. What differs is where the permission model lives, how agent identities are represented, and which controls come as standard versus which need designing. An independent partner brings the pattern from one platform to another without needing to argue that any particular product is the answer.
Our advice to clients this quarter is short and optimistic. Let your revenue, service and operations teams keep building. That energy is the most valuable thing your CRM programme has produced in years, and slowing it down to wait for a governance committee would be the wrong trade. Instead, give them a scoped role template, a named owner per agent and a quarterly review, then get out of the way. Those three habits take a week to establish and they let you keep the pace you have found. We have written before about treating AI agents as managed identities and about measuring agents so they earn budget to scale, and both hold up well against this year’s data.
The organisations that will look strongest eighteen months from now are not the ones that locked agent creation down to a central team. They are the ones that kept the doors open and made scoping a habit at the point of build. That is an achievable position from where most companies stand today, and the teams starting this quarter will feel the benefit long before the end of next year. If you would like a second opinion on how your agent estate is shaping up, schedule a consultation and we will take a look with you.
Get in Touch
If your revenue, service or operations teams have started building agents inside Salesforce, HubSpot or Dynamics 365, we are happy to review how they are scoped and where the quick wins sit. If you would rather start by building one together, a one-day AI hackathon does exactly that.
