Choosing an enterprise AI agent platform
Choosing an enterprise AI agent platform comes down to five checks that a feature list does not answer: which tools the agent may call and under whose credentials, which actions need human confirmation, what is logged, what ends a run, and whether any internal data leaves the organisation to reach the model.
Why does a feature comparison not decide this?
Agentic is applied today to products that differ by more than any feature column shows. One is a fixed workflow with a language model in a single step; another plans an open-ended sequence and calls whatever tools it decides it needs. Both list planning, tool use and memory on the datasheet.
The difference that matters in an enterprise is not capability but control: how much of the sequence the system decides, and what stands between a decision and an irreversible action. That is a property of the deployment, not of the model, and it is what the five checks below measure.
The five checks
| Check | The question | A weak answer |
|---|---|---|
| Tool scope | Which tools can it call, and under whose credentials does each call run? | It integrates with everything |
| Confirmation gates | Which actions require human confirmation, and is that list configurable? | The model is very reliable |
| Audit | What is recorded per step, and for how long is it kept? | There is a conversation history |
| Termination | What ends a run: a step limit, a time limit, a cost limit, or nothing? | It stops when the task is done |
| Model placement | Where does inference happen, and does internal data leave to reach it? | The data is encrypted in transit |
Each weak answer is true and irrelevant. Encryption in transit describes how data travels, not whether it should be travelling; a reliable model is still a model, and the control being asked about is what happens on the occasion it is wrong.
How much autonomy should an enterprise deployment have?
| Level | What the system decides | What it requires |
|---|---|---|
| Assisted | Nothing; it drafts, a person acts | Review before anything happens |
| Supervised | The sequence, confirming each action | A person in the loop |
| Bounded autonomy | Everything within a defined tool and budget scope | Tight permissions, complete logging |
| Unsupervised | Everything | Rarely appropriate outside a sandbox |
Most production deployments sit at supervised or bounded autonomy, and the same software can be configured at any level. A vendor describing their product as autonomous has therefore not answered the question: the level is set by the deploying organisation, and the useful question is which levels the platform makes it possible to enforce.
When does model placement decide the shortlist?
Sending internal documents to a hosted model is a data transfer, whatever the interface looks like. Where policy or law forbids that transfer, the shortlist is decided before capability is compared, exactly as it is with face recognition: the platforms that run the model on the customer's own hardware, and the platforms that do not.
- Hosted model, vendor infrastructure: fastest to start, and a transfer to justify for every document the agent reads.
- Hosted model, customer tenancy: a contractual boundary rather than a physical one, which some regulators accept and some do not.
- Self-hosted model, customer hardware: no transfer, and the constraint becomes fitting a capable enough model to the hardware available.
- Air-gapped: self-hosted with no outbound connection at all, which additionally rules out runtime downloads, telemetry and online licence checks.
A requirement written as on-premise is routinely satisfied by software that calls home for licensing. If the intent is an air gap, the requirement has to say so.
What should a pilot actually measure?
- Pick one workflow that a person currently does end to end, and that has a checkable output.
- Give the agent the narrowest tool set that could complete it, using a service identity rather than someone's credentials.
- Run it alongside the person rather than instead of them, and keep both outputs.
- Count the interventions: how often a human corrected, stopped or redid the work. That number, not a success rate, is what tells you the autonomy level to deploy at.
- Read the logs afterwards as an auditor would, months later, and see whether they explain what happened.
A pilot that reports only completed tasks has measured the easy half. The intervention count is the figure that predicts what the deployment will cost to supervise, and it is the one most rarely reported.
Frequently asked questions
- How do I compare enterprise AI agent platforms?
- By five checks a feature list does not answer: which tools the agent can call and under whose credentials, which actions require human confirmation, what is logged per step and for how long, what terminates a run, and whether internal data leaves the organisation to reach the model.
- What is the difference between an AI agent and an agentic platform?
- An AI agent is the program; agentic describes the property of deciding its own sequence of steps. In vendor material the label covers everything from a scripted workflow with a model in one step to open-ended planning, which is why the five checks separate products that the datasheets do not.
- How much autonomy should an enterprise agent have?
- Most production deployments run supervised or with bounded autonomy: the agent decides the sequence, irreversible actions need confirmation, tools are scoped narrowly and every step is logged. Autonomy is a configuration rather than a product tier, so the question is which levels a platform can enforce.
- Can an AI agent platform run on-premise?
- Yes, when the model, the orchestration and the tools all run on local hardware. The constraint is fitting a capable enough model to the available hardware, not the agent architecture. Air-gapped operation is stricter again: no runtime downloads, no telemetry, no online licence checks.
- What should an agent pilot measure?
- The intervention count — how often a person corrected, stopped or redid the work — rather than a completion rate. That figure predicts what the deployment costs to supervise and sets the autonomy level it can safely run at. A pilot reporting only completed tasks has measured the easy half.
- Is sending company documents to a hosted model a problem?
- It is a data transfer, whatever the interface looks like, and whether it is a problem depends on the policy and jurisdiction that apply to those documents. Where the transfer is forbidden, the shortlist is decided before capability is compared: platforms that run the model on your hardware, and platforms that do not.