A client showed us their AI dashboard last month. Six agents in production, three more in pilot, and a slide claiming they had cut ops time by 40 percent. That was the headline. The real number was buried in their IT bill. They were paying for nine separate vector databases, four different model providers, and a finance team that could not get a clean monthly cost line for any of it.
This is what agent sprawl looks like in 2026, and we are seeing it across most of the mid-market clients we onboard.
The shape of the problem
When an enterprise rolled out SaaS in 2018, every department picked its own tool. Marketing got HubSpot. Sales got Outreach. Support got Zendesk. Each one had its own database, its own user model, and its own bill. By 2021, most ops teams had spent two years cleaning that up.
The AI agent rollout is doing the same thing, just compressed into eighteen months. Marketing buys an agent that drafts ad copy. Sales buys one that summarizes calls. Support runs a chatbot. RevOps builds a forecasting agent. Nobody has talked to anyone else, and now you have four agents reading from four copies of the same CRM, each with its own prompt template and its own definition of what counts as a "qualified lead".
Deloitte put a number on this in their 2026 enterprise prediction report. Integration and governance now eat up to 60 percent of agent project budgets at enterprise scale. That is not the model cost. That is not the engineering hours. That is the tax you pay for not having a shared orchestration layer when you started.
What we look for in the first audit
We do not let a client add another agent until we have answered four questions about the ones already running.
Where is the context coming from? If three agents are pulling from three slightly different copies of "the customer record", we collapse them first. The fastest way to do this in 2026 is to put MCP servers in front of your CRM, your data warehouse, and your support inbox, then have every agent talk to those servers instead of holding its own integration. Microsoft adopted MCP as the integration standard for Windows AI Foundry and Microsoft 365 Copilot earlier this year. Most clients we work with have not made that switch yet, and the savings on duplicated infra alone usually pays for the consolidation work. Who owns the prompt? This sounds like a soft question. It is not. If marketing's agent and sales' agent both define "ICP" in their system prompts, and one says "5 to 50 employees" and the other says "10 to 100", every downstream score is wrong. We move definitions out of prompts and into a shared config file the agents read at runtime. It is two days of work and it ends about half the cross-department arguments we see. What does the agent escalate, and to whom? The single most common failure mode we see in 2026 is an agent that silently does the wrong thing because nobody defined the escalation path. Voice agents are especially bad at this. If the prospect says "I'm not interested but my colleague might be", an agent without an escalation rule will end the call. A human SDR would have asked for the colleague's name. We bake explicit escalation triggers into every agent we ship, and we route them to a real person, not another agent. What is the agent allowed to write? Read access is cheap to give. Write access is where the audit work lives. We default new agents to read-only for the first two weeks, and we only grant write scope after we have logs of what they would have written. About one in three reveals at least one decision we would not have approved.The cost math nobody puts on the slide
Vendor decks for agent platforms quote you the model cost. They almost never quote you the integration cost. Here is what we have seen on three recent client engagements (numbers are rounded and anonymized):
A 200-person SaaS company added a sales-call summarizer. Model cost: $1,800 a month. Integration with Salesforce and Gong, plus the policy work to handle PII redaction: $42,000 in our hours over six weeks. The agent paid for itself in week 11. The integration paid for itself in week 38.
A 50-person ecommerce ops team built a support-tier-1 agent. Model cost: $600 a month. Building the MCP server in front of their order system so the agent could check shipment status without four different API keys floating around: $11,000. They needed exactly one agent. The infra they built is now ready for the next two.
A regulated mid-market healthcare client deployed an internal research agent. Model cost: $2,200 a month. Compliance review, audit logging, and the agent-output retention pipeline: $96,000 over four months. They have not added a second agent yet. They told us they probably would not, given the per-agent compliance overhead they discovered.
That last one is the one to pay attention to. The 2026 industry estimate is $60,000 to $300,000 for a regulated agent build, with up to 60 percent on integration and governance. The healthcare client's number sits exactly on that curve.
The right number of agents is smaller than you think
Most clients come to us thinking they need five or six. After the audit, the number we ship is usually two or three. The ones we keep tend to be specialists with sharp scopes (one agent, one job, one set of tools), not generalists who try to do six things and end up doing none of them well.
The org-chart metaphor that everyone is using in 2026 is actually right. You want a coordinator (or a thin orchestration layer) and a few specialists, not nine generalists who all read your CRM. The specialists are easier to test, easier to swap out when a better model ships, and easier to audit when something goes wrong.
We are running our own internal stack on three agents right now. One handles inbound qualification. One handles client reporting. One does ad-creative drafts. The orchestration sits in n8n with MCP servers in front of our shared data. It is boring infrastructure. That is the point.
If you are about to greenlight your fifth agent, the smarter move is probably to shut down two and tighten the three you keep. We can tell you that for free. The hard part comes after.