How to choose an AI agent platform

Evaluate against your workflows, not demos

The market is noisy: every tool with a chat box now calls itself an agent platform. Demos will not help you tell them apart, because a demo is a rehearsed conversation about the vendor's data. The only test that matters is whether the platform can run one of your workflows — on your channels, under your approval rules, inside your budget.

So before you take a single call, write down two or three workflows you actually want to hand over. For most SMEs they look like answering WhatsApp enquiries about prices and stock, triaging a shared inbox and drafting replies, or producing a weekly research digest. Then score every vendor against those, using the checklist below.

A note on bias, since this page is on a vendor's website: we build one of these platforms, and the checklist mirrors the things we chose to build. It is not neutral. But every item is a real failure mode we have seen businesses hit, and every question works on any vendor — including us.

1. Channels: can agents live where your customers are?

An agent that exists only in a web widget is a chatbot with better grammar. Your enquiries arrive on WhatsApp, email, Telegram, and Slack, and events arrive from your systems as webhooks — if the platform cannot meet them there, your team becomes the copy-paste layer.

• Which channels are supported today — WhatsApp, Telegram, email, Slack, SMS — and which are roadmap?
• Can incoming events (a new email, a webhook from your store) start work, or must a human open a chat first?
• Can one agent serve customers on one channel and report to staff on another?

2. Human control: who approves what, and where?

The difference between a useful agent and a liability is what it may do unsupervised. Look for graduated autonomy rather than an on/off switch: research and drafting can run freely, while outbound messages, money, and system changes wait for a human until trust is earned. Our approval-first playbook covers this design in depth.

• Is autonomy graduated — levels or scopes per agent — or all-or-nothing?
• What waits for approval by default, and can gates be set per category: financial actions, external sends, file writes?
• Where do approval requests arrive — in tools your team already uses, or a dashboard someone must remember to check?

3. Audit: is every action logged and reviewable?

When a customer disputes what they were told, or you are deciding whether to widen an agent's autonomy, you need the record — not a reconstruction. A log is table stakes; an immutable trail covering everything the agent read, drafted, and sent is what actually settles arguments.

• Is the trail immutable, or can entries be edited or deleted?
• Does it cover every action, or only outbound messages?
• Can you see who approved each action, and when?

4. Cost control: caps, metering, and BYOK

Model usage is a variable cost, and variable costs surprise people. Alerts are not control — a hard cap is. Watch for per-seat pricing too: it taxes exactly the adoption you are hoping for.

• Is there a hard spend cap agreed before launch, or only usage alerts after the fact?
• Is AI usage on a transparent meter you can inspect, or bundled into an opaque fee?
• Is pricing per-seat? Can you bring your own model keys (BYOK) — and does that cost extra?

5. Isolation and privacy: whose infrastructure, whose data?

Agents hold conversation history, your documents, and credentials for your connected systems, so evaluate the platform like any other processor of customer data. Be suspicious of absolute promises — sweeping guarantees are marketing, not architecture — and get specific about the verifiable basics instead.

• Is each customer's workspace private and isolated, or shared?
• How are connector tokens and API keys stored, and who can access them? Are there role-based access controls?
• Is data encrypted at rest and in transit, and are specific cloud regions available for data-residency needs?

6. Memory and grounding: does it know your business?

A platform that forgets everything between sessions makes your team repeat themselves forever; one that answers from the model's general training will guess at your prices. You need both persistent memory and grounding in your own documents.

• Does agent memory persist across sessions, or reset with every conversation?
• Can answers be grounded in a knowledge base built from your documents — price lists, policies, SOPs?
• When a document changes, how quickly do answers follow?

7. Operations: who hosts, monitors, and maintains it?

This question separates platforms from toolkits. Someone must host the runtime, watch it overnight, apply updates, and fix the connector that breaks — if that someone is your team, price their hours into the comparison. The trade-offs are laid out in our hosted agents vs DIY automation comparison.

• Who hosts and monitors the agents — the vendor or you?
• What is included: setup, maintenance, monitoring? What is the setup timeline, in business days?
• When something breaks at 2am, whose problem is it?

8. Exit: what leaves with you?

Assume you will want to leave someday, and check the door before you walk in. Contracts, keys, and knowledge are the three things to trace.

• Can you cancel anytime? Is there a money-back window?
• If you brought your own model keys, do those accounts and their history stay yours?
• What do you keep if you go — your documents, your procedures, your conversation records?

How Olano answers these

Briefly, and in the spirit of the checklist — put these questions to us on a call rather than taking a webpage's word for it:

• Channels: 15+ messaging channels — WhatsApp, Telegram, Slack, Discord, email, SMS, Signal, and more — plus inbound webhooks, with 75+ built-in integrations and 450+ connector tools behind them.
• Human control: trust levels 0–4 (Observer → Assistant → Collaborator → Autonomous → Developer) with per-category approval gates; anything outbound waits for approval by default.
• Audit: an immutable audit trail on every action.
• Cost: a transparent meter with a hard spend cap agreed before launch; no per-seat pricing; BYOK at no extra cost, routing across 18+ LLM providers.
• Isolation: a private, isolated workspace per customer; encrypted connector tokens and API keys; role-based access controls; encryption at rest and in transit; custom AWS/GCP regions for data-residency needs.
• Memory: persistent per-agent memory plus knowledge-base grounding over your uploaded documents.
• Operations: hosted and managed — setup, hosting, maintenance, and monitoring included; one workflow is typically live within 1–2 business days of onboarding, bigger rollouts in 1–3 weeks, and everything is quoted and approved before we build.
• Exit: cancel anytime, a 30-day money-back guarantee, and BYOK means your model accounts stay yours.

And where we are honestly not the answer: if your need is individual work — one-off drafting, general Q&A, a small team without customer channels — a general-purpose AI workspace like ChatGPT or Claude may be all you need. An agent platform earns its keep when agents live in your channels, respond to business events, and operate under approvals, audit history, and spending controls.