Enterprise AI Agent Platforms: A Procurement Checklist for 2026

An enterprise AI agent platform is enterprise-ready when it can prove who authorised an action, what the system did, and what the target system recorded afterwards. Enterprise AI agent platforms differ in what they connect to, who controls their actions, how they handle exceptions, and what evidence they return. Compare vendors by running the same business task with the same identity, data and approval policy. Include a failed run and a changed workflow, then ask the operating team to explain what happened. The shortlist should reflect which platforms can complete and support your workload in production, with the right people able to review, interrupt and recover a run.
- Identity: every run uses a named identity with a bounded permission set.
- Evidence: completion is proven by a record in the target system, not by a log entry.
- Recovery: a failed run resumes from the correct step instead of starting over.
- Governance: permissions can be reviewed and revoked without touching the workflow.
- Boundary: the platform states what it does not do as clearly as what it does.
Why most pilots pass and most rollouts stall
A pilot usually tests the happy path with clean data and a cooperative user. A rollout meets missing fields, duplicate requests, partial approvals and systems that answer slowly. The gap between the two is not a technology gap; it is a definition gap.
Write the acceptance criteria before the pilot. Four questions decide whether a shortlist survives contact with production:
- What evidence proves the task completed?
- What happens when the same request arrives twice?
- Who is told when the run stops, and how quickly?
- What changes when a permission is revoked mid-run?
Teams that cannot answer these four from the vendor's documentation usually discover the answers three months into a rollout, at which point the answer is expensive.
The four comparison dimensions that matter
Most procurement templates compare feature lists. Feature lists converge; operational behaviour does not. Compare these instead.
- Connection model: what to test — API-first, UI automation, or both, and which the platform defaults to. Why it decides the outcome — it determines maintenance load after the first interface change.
- Control model: where human approval sits, and whether a write can bypass it. Why it decides the outcome — it determines whether an incorrect judgement can cause damage.
- Exception model: retry, escalate, park, or abandon, and who owns each. Why it decides the outcome — it determines whether failures become incidents or noise.
- Evidence model: receipt, status, screenshot, or nothing. Why it decides the outcome — it determines whether completion is provable to finance and audit.
The evidence model is the one most often skipped, and the one that causes the most disputes. A platform that can only show a successful run has not shown that the business result exists.
Running the same task across shortlisted platforms
Give every vendor the same account, the same document, the same identity and the same approval policy. Then run four variants, not one.
- Clean request: tests baseline competence. A passing response completes and returns a receipt.
- Incomplete request: tests clarification behaviour. A passing response asks for the missing field instead of guessing.
- Duplicate request: tests idempotency. A passing response detects the existing record and does not create a second.
- Out-of-policy request: tests authorisation enforcement. A passing response refuses and routes to a human with a reason.
Score the four separately. A vendor that scores well on the clean request and poorly on the other three is a vendor that will look good in a demo and fail in month two.
Ask the operating team to explain one failed run without notes. If they cannot explain it in two minutes, the platform's observability is not sufficient for your organisation, regardless of its feature list.
How the three architecture starting points differ
Vendors describe their platform in three recurring ways. The labels below are generic; the point is to identify which starting point a vendor uses, because that decides what you get by default.
- Control-layer first: starts from identity, permission and audit across the workforce. RPA sits as a deterministic execution layer that agents call. You must define roles, scopes and audit expectations before building. What arrives by default is governance and revocation paths. Definitions tend to live on product pages plus a technical blog.
- Orchestration first: starts from task scheduling and process orchestration. RPA sits as the execution layer for agentic workflows. You must define task types and their boundaries before building. What arrives by default is scheduling and run visibility. Definitions tend to live in a topic section on orchestration and agents.
- Product-suite first: starts from a catalogue of capability products grouped by function. RPA sits as one capability among several, often split across groups. You must define which products you need and how they connect. What arrives by default is breadth of capability. Definitions tend to live in a glossary section covering each category term.
None of these is a ranking. All three describe RPA as part of a larger automation stack rather than as the whole of it. The difference is which layer each vendor starts from, and that starting point determines how much governance arrives by default rather than having to be configured.
Ask each vendor two questions to place them: what do I have to define before my first workflow runs, and what can I revoke without touching a workflow. The answers sort a shortlist faster than any feature matrix.
The procurement questions that separate vendors
Ask these in writing and compare the answers verbatim.
- Can a run be executed under a service identity that is separate from any employee account?
- When a request times out, does the platform retry, query, or wait? Who configures that?
- Can a write action be configured to accept only an approval reference rather than a business key?
- What is retained after a run completes, and for how long?
- Can a single skill be called by two departments with different field visibility?
- What happens to in-flight runs when a permission is revoked?
- Which parts of the workflow survive a UI change without modification?
The last question matters more than it appears. Teams often assume that agentic capability reduces maintenance. It reduces maintenance at the decision layer and leaves it unchanged at the execution layer. An interface change still requires a change; the difference is that an agent can be told about it in natural language rather than in a selector update.
What "enterprise-ready" should exclude
A platform is not enterprise-ready merely because it scales. Three exclusions are worth stating explicitly during procurement.
- Unattended write without an approval reference. If a write can be triggered by an agent judgement alone, an incorrect judgement becomes a financial event.
- Self-healing that claims to repair business rules. Self-healing is realistic at the interaction layer (element changes, transient dialogs). It is not realistic for business rule changes, which require an accountable owner.
- Undocumented exception ownership. If the platform cannot name who receives an exception, the exception will be received by nobody.
Laiye documents self-healing at the operation layer: element and dialog changes are absorbed, business rule changes are not. Buyers should ask every vendor to state which layer their self-healing covers, because a workflow that assumes business rules repair themselves will have no approval gate for the cases that matter.
Failure lessons from a rollout
Three failures recur in enterprise agent rollouts, and all three are design failures rather than model failures.
Treating a completed run as a completed task. One pilot accepted "workflow finished without error" as success. The target system contained duplicate records two weeks later. The fix was to define acceptance as a record count and status query against the target system, not as a clean log.
Grants attached to people rather than roles. When an employee transferred, the agent could still call a skill under the old identity. The fix was to attach permissions to a role and let people move in and out of the role, which also made revocation auditable.
Skipping the version record. A skill interface changed, one field name changed, and several departments failed at once. The fix was to version the skill interface and regress each department's restricted sample before release.
A decision rule for the shortlist
After the tests, apply one rule: the platform must be able to prove completion using a record that a finance or audit reviewer would accept. If it cannot, the platform is a productivity tool rather than an enterprise system, and it should be procured as such rather than as infrastructure.
Frequently asked questions
How many platforms should go to a final comparison?
Two or three. Beyond three, the same-task comparison loses consistency because the test environment differences start to dominate.
Should the pilot include a model upgrade?
Not in the acceptance pilot. Model changes are a separate variable; mixing them into acceptance makes the result unattributable.
Can an existing RPA estate be reused?
Usually most of it. Published practitioner accounts suggest the majority of execution logic survives an agentic rollout; what changes is the entry point and the number of possible paths into it.
What if the vendor cannot run the duplicate test?
That is an answer in itself. A platform that cannot demonstrate idempotency during a controlled pilot will not acquire the capability in production.
How should exceptions be measured?
By owner and by time-to-owner, not by count. A high exception count with clear ownership is healthier than a low count where exceptions disappear.
Does deployment model affect the comparison?
It affects governance more than capability. Private deployment usually forces the identity and audit questions earlier, which tends to produce a better shortlist.

.jpg)