Most companies do not have an AI problem. They have a coordination problem. Their agents work, but each one works alone, and the handoff between them is still a human copying text from one tab into another.
This is now the single most common reason agent projects stall. Not model quality. Not budget. Just wiring.
How many AI agents does a typical company actually run?
More than most people expect. Salesforce's 2026 Connectivity Benchmark Report puts the average at around 12 AI agents per company, with that number expected to reach roughly 20 by 2027.
The problem is what those 12 agents do together, which is usually nothing. About half of them run completely on their own, with no connection to the others.
Gartner's forecast on this is blunt: more than 40% of agentic AI projects are expected to be cancelled or paused by the end of 2027. The reasons given are rising costs, unclear business value, and weak risk controls. All three trace back to the same root — agents that were deployed one at a time, by different teams, with nobody owning the layer between them.
Why do AI agents end up disconnected?
It is almost never a deliberate decision. It happens the same way in most companies.
Each agent was bought by a different team
Support bought a ticket triage agent. Marketing bought a content agent. Finance bought an invoice reader. Each purchase made sense on its own. None of them were evaluated as part of a system, because at the time there was no system.
Nobody owns the layer in between
A single agent has a clear owner. The connection between two agents owned by two different departments has no owner at all. So it does not get built.
There is no shared memory
Agent A learns something about a customer. Agent B has no way to know it. Every agent starts every task from zero, which means the same context gets rebuilt — usually by a person — over and over.
The protocols arrived after the agents did
Standards like MCP (Model Context Protocol), A2A (Agent-to-Agent) and ACP (Agent Communication Protocol) now exist to let agents share context and call each other. But most companies deployed their first agents before adopting any of them. Retrofitting a protocol onto five agents built on five different assumptions is real engineering work, and it keeps getting postponed.
What does the disconnection actually cost?
The cost is invisible on any invoice, which is exactly why it survives. Here is where it shows up.
Work gets done twice. A lead fills a form. The intake agent captures it. The CRM agent never hears about it, so someone re-enters the same details by hand. You paid for automation and still pay for the typing.
Handoffs fail silently. A support agent resolves a ticket that was actually a billing issue. There is no path to the finance agent, so the ticket closes and the billing error stays. Nobody finds out until the customer complains again.
Context dies at the boundary. Your content agent writes a case study. Your outreach agent sends cold emails to the exact industry that case study was written for. Neither knows the other exists, so the email goes out without the one asset that would have made it land.
Nobody can measure anything. When each agent logs separately, there is no end-to-end view of a task. You cannot say what a completed workflow costs or how long it takes. That missing number is why leadership eventually calls the whole programme unproven — and it is the number one predictor of a paused project.
Agents alone vs agents connected
| Agents working alone | Agents working as a system | |
|---|---|---|
| Handoff | A human copies output from one to the next | Routed automatically by a manager layer |
| Context | Rebuilt from zero on every task | Shared state passed between agents |
| Quality control | Each agent trusted on its own output | An evaluator checks output before it moves on |
| Failure | Silent — you find out from a customer | Surfaced the same day, with the step that broke |
| Cost visibility | Per tool, per licence | Per completed workflow |
| Adding an agent | Another disconnected island | Plugs into the existing routing layer |
What does a connected multi-agent system look like?
This is the architecture we run internally at ZACSOL, and the same shape we build for clients. It is deliberately simple — four responsibilities, clearly separated.
1. A manager agent that routes
One agent is the entry point. It receives the request, decides which specialist agent should handle it, and passes the work along with the context already attached. Users talk to one thing, not seven.
This single decision removes most of the copy-paste problem, because the handoff is now a routing step instead of a human step.
2. Specialist agents that do one job well
Below the manager sit narrow agents. A content agent for posts and scripts. An outreach agent that takes a lead list, runs contact and follow-up, and carries a thread through to close. Each one is small enough to test, debug, and replace without touching the others.
Narrow beats general here. A specialist agent with a tight scope is far easier to make reliable than one general agent asked to do everything.
3. An evaluator agent that checks the work
This is the piece most teams skip, and it is the one that decides whether the system survives contact with real users. A separate agent grades what goes into and comes out of the specialist agents. Bad output gets caught inside the system instead of in front of a customer.
Without this layer, every quality problem becomes a human review problem, and the promised time saving disappears.
4. Human approval where the stakes are real
Anything that leaves the company — an email to a prospect, a published post, a payment — stops for a person to approve. Full autonomy sounds impressive in a demo and is the wrong default in production.
Anthropic's 2026 agentic coding research found that while developers use AI for roughly 60% of their work, only a small fraction of tasks can be handed off completely. People still review and steer. The same is true of business agents. Design for it instead of fighting it.
A four-question check on your own setup
Answer these honestly. They take two minutes and they are more useful than any vendor assessment.
1. If agent A finishes a task that agent B should pick up, does a human move it? If yes, that is your first integration. 2. Can you name the cost of one completed workflow end to end? Not the licence cost — the workflow cost. 3. When an agent produces something wrong, who catches it, and how long does that take? 4. If you added a ninth agent tomorrow, would it plug into anything, or would it be another island? Any question you cannot answer cleanly points to a missing layer, not a missing tool.
The short version
Buying more agents does not compound. Connecting the agents you already have does. Most companies are one routing layer and one evaluator away from getting real value out of tools they already pay for — and they keep deploying agent number thirteen instead.
If you want to see what that looks like on your own stack, describe your setup to ZAC Consultant. It maps which of your agents should be connected, what the routing layer needs to do, and what the work involves — in about three minutes, free, no call required.
For the delivery side of this work, see our AI automation and business process automation services, or browse production systems we have shipped.