Most enterprise AI agent shortlists look identical on paper. A chat window, a workflow builder, a connector library plastered with 200 logos, and a pricing page that hides the only number you care about. Then you point the thing at your actual ERP, and three of the connectors need custom OAuth work that nobody scoped.
The gap between the demo and the deployment is where the buying decision actually gets made. So rather than rank vendors by how impressive their launch keynote was, here is what separates the agents that survive contact with a real organisation from the ones that quietly get switched off in month five.
Judge an agent by what it finishes, not what it shows
The single most useful metric in this category is completion rate on a defined task. Not accuracy on a benchmark, not tokens per second. Completion rate: of the 1,000 invoices that arrived this month, how many did the agent take from intake to posted, without a human finishing the job?
That number is brutal and honest. A support agent that resolves 61% of tier-one tickets end to end is worth far more than one that drafts a reply for a human to approve 100% of the time, because drafting tools add review work rather than removing it. The teams getting real returns treat autonomous AI agents as systems judged by what they finish, and they write the completion threshold into the contract before signing anything.
The platforms worth a serious look
There is no single winner. Enterprise agent platforms tend to be very good inside the ecosystem that birthed them and mediocre outside it, which is why most large companies end up running two or three side by side.
Microsoft Copilot Studio and Azure AI Foundry
The default choice for Microsoft-heavy estates. Identity, permissions, and data governance already exist in Entra ID, and the agent inherits them instead of bolting on a parallel permission model. Strong on internal knowledge and Power Platform workflows. Weaker when the work lives in a legacy system with no API.
Salesforce Agentforce
Consumption pricing, tied closely to CRM objects. Excellent for service triage, case summarisation, and sales follow-up where the data is already in Salesforce. If your customer records are split across four systems, half the value evaporates before you start.
ServiceNow AI Agents
The sleeper pick. ITSM and HR service delivery data already sits in a well-structured workflow model, which is exactly the substrate an agent needs. Approval chains, SLAs, and escalation paths are pre-built, so the agent plugs into process rather than inventing it.
UiPath and the RPA lineage
Less fashionable, quietly effective. If your bottleneck is a 2009 thick-client application that will never have an API, screen-level automation plus an LLM layer for judgement calls beats a pure-API agent that simply cannot reach the system.
Build-your-own with LangGraph, CrewAI, or the OpenAI Agents SDK
Reasonable once you have more than about five agent use cases and a platform team to own them. Before that, you are paying engineers to rebuild connector management and observability that a vendor already ships.
The practical test is not feature comparison. It is whether the platform survives your actual systems — your auth patterns, your rate limits, your 14-year-old database with the weird schema. Run a connector against your three ugliest systems in week one and watch what breaks.
Where enterprise agents actually pay for themselves
Across the deployments that hold up, value clusters in a small number of places. These are the ones that consistently show a return inside two quarters:
- Accounts payable exception handling. Matching a purchase order to an invoice to a receipt, then routing the 15% that genuinely need a human. Teams report cutting handling time from double-digit minutes per document to low single digits.
- Tier-one support resolution. Password resets, order status, returns, address changes. High volume, narrow scope, clear success criteria.
- Procurement intake. Turning a rambling email from a department head into a structured request with the right approval path attached.
- Engineering toil. Migration work, dependency upgrades, on-call triage that summarises an incident before the human joins the bridge.
- CRM hygiene. Post-call note capture, deal-stage updates, and contact enrichment that sales reps otherwise skip.
There is a pattern in that list. Every item is high-volume, bounded, and has an unambiguous definition of done. Real AI agent examples doing actual work almost always look mundane from the outside, and that is the point.
The line items nobody puts in the business case
Seat licensing is the visible cost and usually the smallest. The rest hides in four places: consumption fees from retries and long context windows, a second model call for verification on anything customer-facing, maintenance of connectors as upstream APIs change, and human review time for the residual cases the agent hands back.
Budget the review time honestly. If an agent resolves 70% of tickets and escalates 30%, the economics look great only if you actually reduce headcount or redeploy those hours. Plenty of pilots produce a beautiful dashboard and no change in staffing, which is why finance starts asking pointed questions in quarter three.
Permissions decide the shortlist faster than features do
Ask one question early: does the agent act as the user, or as a service account with broad access? Acting as the user preserves your existing permission model and keeps the blast radius small. A service account with read access to everything is faster to build and much harder to get past security review.
The related issue is prompt injection. Any agent that reads inbound email or web content can be manipulated by that content. Guardrails, allow-listed tool calls, and human approval on irreversible actions are not optional extras for a system that can issue refunds or move money. Insist on a full audit trail of every tool call, attributed to a real identity, exportable to your SIEM.
Pilots rarely fail at launch. They fail in month three.
The pattern is consistent. Weeks one to four go brilliantly because the team is hand-holding the agent and picking clean examples. Weeks six to twelve is when volume arrives, edge cases multiply, and the agent starts quietly degrading on the 8% of inputs that look different from the training set.
Mapping where an AI agent workflow breaks before you scale is worth more than any additional feature. Watch the handoff points in particular — the moment the agent passes a task to a human, or to another agent, is where context gets dropped and work gets duplicated.
Voice is a different bet entirely
Text agents inside internal workflows are relatively mature. Voice agents on customer calls are not, and voice AI has not had its ChatGPT moment by most operators’ own admission. Latency, interruption handling, and accented speech still produce failure rates that are fine for a restaurant booking line and unacceptable for a regulated financial conversation. Pilot it, do not bet the contact centre on it yet.
A 30-day evaluation that predicts production
Skip the six-month proof of concept. Pick three bounded tasks with real users attached, and measure four things weekly:
- Completion rate, meaning tasks finished without human intervention, tracked per task type.
- Cost per completed task, including retries, verification calls, and review labour.
- Escalation quality, meaning whether the handoff to a human carries enough context to be useful.
- Time to connect a new system, measured in engineer-days, because that number tells you how expensive use case four will be.
If completion rate is climbing, cost per task is falling, and your engineers can wire a new system in under a week by day 30, you have something worth scaling. If completion is flat and every new connector needs a bespoke integration, the platform is the problem, not the model. Knowing that in a month rather than a year is the whole game.