Skip to content

/articlesGuides

How to choose a platform for company assistants

An enterprise AI agent platform buyer’s guide: 10 criteria to test live, from integrations and approvals to data isolation, Arabic and channels, and a pilot plan.

Updated
Published
min read
07
On this page
09
On this page09

Start with the task, not the vendor

Most failed purchases fail before the demo. The team compares features in general, signs, and then finds the platform cannot reach the one system that matters.

Write two or three tasks first, each as a sentence someone would actually say. For example: “Answer WhatsApp price questions from our price list and prepare an Odoo quotation”, or “Check this Dynatrace alert and draft a ServiceNow incident for my approval”.

For each task, list what it reads, what it changes, who approves, and what done looks like. That list becomes your test script.

  • Systems read: for example Odoo products, Splunk logs, a Gmail inbox.
  • Systems changed: quotations created, tickets updated, emails sent.
  • Approvers: which role approves which action, and who covers when they are away.
  • Language and channel: Arabic, English or both; website chat, WhatsApp, Telegram, Slack, or inside Odoo.
  • Success measure: time to answer, steps removed, errors avoided.

Ten criteria and the questions to ask

Use this table as a score sheet. Score only what the vendor shows on your task. A roadmap slide scores zero.

CriterionWhat good looks likeAsk the vendorRed flag
IntegrationsNamed read and write actions for your systems, plus a way to add your ownWhich actions exist for our systems, and which change data?A wall of logos with no list of actions
Per-tool rulesEach action runs directly, follows your rules, or always needs approvalCan reading email run freely while sending always needs a person?One on/off switch for all actions
ApprovalsThe task pauses, a person approves the exact action, the task continuesShow a paused task approved by someone other than the requester.The same user clicks “confirm” in the chat
AuditWho asked, what ran and who approved, recorded per taskCan we see who approved what?Only chat transcripts are kept
Data isolationEach company’s data kept apart by a check every query passesWhat stops a query from returning another company’s data, and what does that check not cover?“Every customer has its own prompt”
CredentialsEncrypted, tied to their owner, keys rotatedHow are credentials encrypted, and can any screen or API show one?Credentials visible in an admin screen or sent to the model
ArabicUnderstands Gulf and Levant dialects, answers in clear Arabic, right-to-left documentsRun our task in Arabic end to end.Arabic only through a translation layer
ChannelsEvery channel goes through the same rules and approvalsDo WhatsApp and website chat follow the same approvals?Each channel is a separate product with separate rules
Model choiceSeveral model providers; tools and rules live outside the modelWhat changes if we switch models next quarter?Assistants tied to one provider
PilotFixed scope, real systems, agreed measures, clean exitWhat do we keep if we stop after the pilot?Pilot only in the vendor’s sandbox with sample data

How deep do the integrations go?

Count actions, not logos. An integration that only searches tickets is very different from one that can create an incident and close it. Ask for the list of actions per system, and the risk class of each: read, write, destructive (deletes data) or external send (reaches people outside the company).

Ask how you add a system the vendor does not cover. MCP (Model Context Protocol) is a common standard for plugging tools into assistants. Check that tools you add follow the same rules as built-in ones.

In Botify, for example, Gmail send, reply and forward always need approval, Odoo product search runs directly, and tools from an MCP server you connect need approval until someone reviews them.

Rules, approvals and audit: what to verify

Rules only count if they are checked before the action runs. Ask where the permission check happens. A clever prompt must not be able to talk the model into something the user is not allowed to do.

Approvals must survive a restart. If a task waits for a manager overnight, it should continue from the same step when the decision arrives.

Then open the audit record for a task you just watched. It should show who asked, what ran, and who approved or rejected.

Security: data isolation and credentials

A shared platform keeps many companies’ data in one database. Ask where the company filter is enforced and what it does not cover. In Botify it is enforced in the application on every top-level query; nested writes and raw SQL carry their own filter. Botify does not give each company a separate database.

For credentials, ask which cipher is used, whether each value is tied to its company, and how key rotation works. Botify uses AES-256-GCM, records the key version in each encrypted value, and re-encrypts old values to the current key every hour.

Answers like “we use the cloud provider’s encryption” are a reason to ask more.

Arabic and regional requirements

For companies in Saudi Arabia and Jordan, Arabic is a requirement, not an extra. Test with real messages from your customers and staff, including dialect and English terms mixed in. Check that reports read correctly right to left.

Data protection law shapes the purchase too. Saudi Arabia’s Personal Data Protection Law has been fully enforceable since 14 September 2024 and applies to organisations outside the Kingdom that process the personal data of its residents. Jordan’s Personal Data Protection Law No. 24 of 2023 took effect on 17 March 2024.

Ask each vendor where prompts, tool results and logs are processed and stored, and bring in your data protection officer before the pilot. Botify does not state a hosting region or certifications on this site; ask for what your review needs.

Channels and model choice

Check that every channel your customers and staff use goes through the same rules. In Botify today that is website chat, WhatsApp, Telegram, Slack, inside Odoo, the HTTP API and MCP. WhatsApp and Telegram also take voice notes. A phone agent is on its way.

Email and Microsoft Teams are integrations in Botify, not channels: the assistant can read and draft in Gmail or Outlook and post in Teams, but customers do not chat with it there.

Models change faster than contracts. Botify works with Claude, OpenAI, Gemini and xAI models, and keeps tools and rules outside the model.

How to run a pilot that proves something

A good pilot is small, real and measurable. It uses your systems and your people, runs the task you chose, and ends with a decision.

  • Pick one task with clear value and moderate risk, such as WhatsApp price questions or alert triage.
  • Connect real systems with credentials created for the pilot, with the fewest permissions that work.
  • Set rules before the first run: reading open, anything that changes data or contacts customers behind approval.
  • Agree success measures up front, and record how the work is done today.
  • Test failures on purpose: a user without permission, a rejected approval, an Arabic request, a removed credential.
  • Review the audit log with security and your data protection officer at the end.
  • Confirm the exit: how you export your setup and data, and how credentials are deleted.

Common questions

01

Is an AI assistant platform the same as a chatbot builder?

No. A chatbot builder answers questions. An assistant platform also takes actions in your systems, so it needs per-action rules, approvals and an audit log. Test those as carefully as the answers.

02

Should we build our own platform instead?

Building makes sense if assistants are your core product and you can staff the platform for years. Most companies underestimate the work around the model: data isolation, credential storage, approvals, audit and keeping integrations working.

03

How many criteria should our request for proposal include?

Few enough that every vendor must show each one. The ten in this guide cover what usually decides success; add your sector’s requirements from security and compliance.

04

Why does Arabic support matter?

Customers and staff in the Gulf and Jordan write in Arabic, often in dialect and mixed with English terms. A platform that only translates loses meaning and layout, so test end to end in Arabic.

05

Which actions should need approval in a pilot?

Anything that changes data outside the platform or contacts people: sending email, sending quotations, creating or updating tickets, deleting anything. Read-only lookups can run directly.

Sources

  1. [1]Saudi Arabia’s Personal Data Protection Law becomes enforceable (Clyde & Co, 2024)
  2. [2]Jordan issues first personal data protection law (Clyde & Co, 2023)

Where you’ll find this in Botify

Start with one assistant on WhatsApp or your website.

Book a demo and we’ll show you an assistant answering questions like your customers’. We set up the first one with you and review the results together before you add more.