AI operating system for companies: the enterprise guide (2026)

Contents

1. The governance gap
2. What an AI operating system means at this size
3. How we evaluated the platforms
4. The seven platforms, ranked
5. The platforms compared
6. The criteria
7. What the evidence shows
8. The path to a governed rollout
9. Sources

In brief

For an established company of 100 to 2,000 people that needs governance and data ownership before AI acts on its records, OpenLazarus is our top pick: one platform the company owns, with role-based access, approval rules per AI teammate and a log of every action. Glean is the stronger choice for permission-aware search and agents across existing systems, Microsoft Copilot Studio for companies standardised on Microsoft 365, and WRITER for governed knowledge and content work.

• A governed platform the company owns: OpenLazarus. Role-based access, per-teammate approval rules and an action log, on a server that belongs to the company.
• Governed knowledge and content work: WRITER. Approvals, custom roles, audit logs and agent observability listed for the enterprise plan.
• Permission-aware search across many systems: Glean. Enforces source-system permissions on every read and write and streams its audit trail to your SIEM.
• Microsoft 365 organisations: Microsoft Copilot Studio. Agents built inside the tenant, identity and compliance tools you already run.
• Data-intensive operations at scale: Palantir AIP. AI workflows on a modelled view of the business, with audit tracking and lineage.

The governance gap

In a company of a few hundred people, AI is already in use whether or not anyone approved it. The question for leadership is no longer whether to adopt it, but how to bring it under the same controls as everything else the company runs: who it acts for, what it may see and change, who signs off, and where the record lives.

The cost of not answering that question is measurable. IBM's 2025 Cost of a Data Breach study, conducted by the Ponemon Institute across 600 organisations, found that 13% had experienced breaches of AI models or applications; of those, 97% said they had no AI access controls in place, and 63% either had no AI governance policy or were still writing one. Organisations with high levels of unsanctioned "shadow" AI paid on average $670,000 more per breach⁠1IBM report: 13% of organizations reported breaches of AI models or applications, 97% of which reported lacking proper AI access controlsIBM Newsroom · 30 Jul 2025.

Regulation is arriving at the same time. In the European Union, AI literacy obligations have applied since 2 February 2025, and the AI Act has largely applied since 2 August 2026; obligations for high-risk uses listed in Annex III, which include areas such as employment, now apply from 2 December 2027⁠2AI Act: regulatory framework for AIEuropean Commission. For high-risk systems, the Commission lists human oversight and logging of activity among the requirements⁠2AI Act: regulatory framework for AIEuropean Commission.

And the costs keep moving. Zylo found that 78% of IT leaders had faced unexpected charges tied to AI features or consumption-based pricing, and 61% had cut projects because of unplanned SaaS cost increases⁠32026 SaaS Management IndexZylo.

What an AI operating system means at this size

For a 20-person firm, an AI operating system is mostly a question of convenience. For an established company it is a question of control. We use the term for a company-wide platform on which staff and AI agents work on the same records, under one model of permissions, approvals and audit. Some products on this list hold those records themselves; others sit across the systems a company already has and enforce their permissions. Both approaches can be governed well. They differ in what the company owns at the end.

How we evaluated the platforms

We reviewed seven platforms that established companies use or consider as their company-wide AI layer: an owned platform delivered with architects, two enterprise agent platforms, an enterprise search and agent platform, an agent builder for Microsoft 365, the agent layer of the largest CRM and a data platform. Each was assessed on the same seven criteria, using only what the vendor publishes, read on the date shown. The criteria and rules are set out on the methodology page.

The ranking reflects one situation: an established company of 100 to 2,000 employees that wants AI to act on company records, and that requires governance and data ownership to be settled first. Where a different situation favours another platform, its profile says so.

The seven platforms, ranked

1. 1. OpenLazarus

   Company-owned AI platform with architects

   Best for: Established companies that want one AI platform they own, on their own server, with role-based access, approval rules per AI teammate and a log of every action.

   OpenLazarus designs and runs one AI platform per company, holding CRM, mail, phones, operations and finance with AI teammates working on the same records. For larger companies it works through a yearly AI-Architect engagement, and hosts the result on a server that belongs to the customer.

   Full profile · openlazarus.ai

2. 2. WRITER

   Enterprise agent platform

   Best for: Enterprises that want governed AI agents for content-heavy and knowledge work, with approvals, roles and audit logs built in.

   WRITER sells an agent platform with prebuilt agents, Playbooks for multi-step workflows and a Knowledge Graph over company content. Its enterprise plan lists approvals, admin controls, role-based access, audit logs and agent observability, and its data policy describes zero data retention.

   Full profile · writer.com

3. 3. Glean

   Enterprise AI search and agents

   Best for: Companies with knowledge spread across many systems that want search, an assistant and agents that respect each system’s permissions.

   Glean describes itself as an enterprise AI platform that unifies search, assistants and agents on a company’s context. Its security page commits to enforcing source-system permissions on every read and write and to logging every control point to an audit trail that can be streamed to a SIEM.

   Full profile · glean.com

4. 4. Microsoft Copilot Studio

   Agent builder for Microsoft 365

   Best for: Companies standardised on Microsoft 365 that want to build governed agents within the tenant they already administer.

   Copilot Studio is Microsoft’s low-code tool for building agents that connect to company data and automate processes. Usage is paid in Copilot Credits, bought as prepaid packs, through a one-year prepurchase plan or pay-as-you-go via Azure. Agents in Copilot Chat, Teams or SharePoint run with the user’s identity, and Copilot access is scoped by each user’s permissions.

   Full profile · learn.microsoft.com

5. 5. Gemini Enterprise

   Enterprise work agent

   Best for: Companies on Google Cloud or Workspace that want one agent across email, documents, code and data, running under each person’s permissions.

   Google presents Gemini for business as one agent that works across a company’s email, documents, code, data and business apps. It runs under each person’s existing permissions, leaves an audit trail, indexes selected sources while leaving data where it lives, and lets administrators set a budget cap per project.

   Full profile · cloud.google.com

6. 6. Salesforce Agentforce

   Agents on the Salesforce platform

   Best for: Companies that run customer operations on Salesforce and want agents governed by the platform’s existing permissions.

   Agentforce adds agents to Salesforce. Its trust layer keeps permissions and data access controls in place, masks sensitive data before prompts reach a model, screens output and does not store prompts with third-party models. Usage is priced by conversation, by Flex Credits or per user.

   Full profile · salesforce.com

7. 7. Palantir AIP

   Data and AI platform

   Best for: Large, data-intensive operations that need AI workflows built on a modelled view of the business, with lineage and audit.

   Palantir’s AIP connects AI with an organisation’s data and operations. Builder tools such as AIP Logic and AIP Chatbot Studio sit on the Palantir Ontology, and the platform includes access restrictions, encryption, audit tracking and lineage, with a choice of language models per use case.

   Full profile · palantir.com

OpenLazarus ranks first in this frame because it combines the governance controls with ownership. Access follows the roles the company sets, each AI teammate sees only what its task needs, and every action is logged with who, what and when⁠45OpenLazarus for accounting firmsOpenLazarusOpenLazarus for healthcare practicesOpenLazarus. The company decides per teammate what is handled alone and what is asked first⁠6FAQOpenLazarus. The platform runs on a server that belongs to the customer, the code and data are the customer's, and European companies can be hosted in the EU under EU law⁠67FAQOpenLazarusOpenLazarus for professional servicesOpenLazarus. For companies above 100 people it recommends a yearly AI-Architect engagement to design the platform with leadership⁠8PricingOpenLazarus. It lists no security certifications on the pages we read, and its client evidence is testimonials on its own site, so security review and a bounded pilot should come first.

The platforms compared

The comparison shows a split. Glean, Gemini Enterprise and Microsoft's tools publish clear permission models, because they work across systems whose permissions already exist: Glean enforces source-system permissions on every read and write⁠9SecurityGlean, Gemini runs under each person's existing permissions⁠10Gemini for business: the universal work agentGoogle Cloud, and Microsoft scopes Copilot access by user permissions⁠11What is Microsoft Copilot?Microsoft Learn · 5 Nov 2026. Approval rules for what an agent may do on its own are published far less often. WRITER lists approvals on its enterprise plan⁠12WRITER pricingWRITER, and OpenLazarus sets them per teammate⁠6FAQOpenLazarus; for several other platforms we found nothing published either way.

The criteria

1. Permission-aware access. The AI sees and changes only what the user, role or task is allowed to, enforced by the platform rather than by policy documents.
2. Approval rules. Administrators set, per agent or task, what runs automatically and what waits for a named person to approve.
3. Audit trail. Every AI action is recorded with who, what and when, in a form compliance and security teams can review.
4. Data ownership and residency. The company owns its data and configuration, can choose where it is hosted (including the EU) and can leave without rebuilding.
5. Reach across systems. Works on records across CRM, mail, finance and operations, not only on documents or one application.
6. Delivery support. The vendor provides architects or engineers to design, build and run the system with the company.
7. Cost predictability. Publishes prices or a billing model a finance team can forecast for a few hundred users.

What the evidence shows

• Audit trails are now a standard claim, but the detail varies. Glean logs every control point to an end-to-end trail that can be streamed to a SIEM⁠9SecurityGlean. Palantir AIP includes audit tracking and historical lineage⁠13AIP overviewPalantir Documentation. WRITER lists agent observability and audit logs⁠12WRITER pricingWRITER. Microsoft Purview can review Copilot prompts and responses⁠11What is Microsoft Copilot?Microsoft Learn · 5 Nov 2026.
• Data handling policies are converging. Salesforce states that prompts and responses are not stored or used to train third-party models⁠14Trusted AISalesforce, WRITER describes zero data retention⁠12WRITER pricingWRITER, and OpenLazarus says nothing a firm holds is used to train a model⁠4OpenLazarus for accounting firmsOpenLazarus.
• Ownership is where the platforms differ most. Most of them leave data in the vendor's environment or in the source systems. OpenLazarus states that the code and data are the customer's and that hosting can be moved to the customer's own cloud⁠68FAQOpenLazarusPricingOpenLazarus.
• Independent results are scarce. None of the seven publishes an independent, named study of a company-wide AI platform at this size. OpenLazarus publishes client testimonials with figures, which are the clients' statements⁠15AI implementation partner: systems you ownOpenLazarus.

The path to a governed rollout

Established companies that bring AI under control tend to work in this order.

1. Form a small AI council. IT, security, compliance, finance and one operating leader, with authority to approve uses and stop them.
2. Inventory what is already in use. List the AI tools staff use today, sanctioned or not, and the data each touches. IBM's findings on shadow AI make this the cheapest risk reduction available⁠1IBM report: 13% of organizations reported breaches of AI models or applications, 97% of which reported lacking proper AI access controlsIBM Newsroom · 30 Jul 2025.
3. Write the permission model before choosing the platform. For each role and each kind of agent: what it may read, what it may change, what needs approval and who approves.
4. Classify uses against the AI Act. Identify any use that could be high-risk, such as screening candidates, and plan human oversight and logging for it⁠2AI Act: regulatory framework for AIEuropean Commission.
5. Pilot in one function with the full controls on. Approval required for every consequential action at first, a daily review of the action log, and a measure agreed in advance.
6. Relax controls on evidence, then extend. Move tasks from "ask first" to "handle alone" only after a record of correct handling, and add the next function only when the first is stable.

Our guide to AI agent permissions and audit sets out the permission model in detail, and the cost guide shows what each approach costs a 200-person company.

The end state is an AI estate the company can account for. Every agent has an owner, a role and written limits; every action is in a log the security team can read; data sits where the company decided it should; and the board can be told, in one page, what AI does in the company and what it costs.

Sources

1. IBM report: 13% of organizations reported breaches of AI models or applications, 97% of which reported lacking proper AI access controls. IBM Newsroom, 30 Jul 2025. Accessed 8 Nov 2026.
2. AI Act: regulatory framework for AI. European Commission. Accessed 8 Nov 2026.
3. 2026 SaaS Management Index. Zylo. Accessed 8 Nov 2026.
4. OpenLazarus for accounting firms. OpenLazarus. Accessed 8 Nov 2026.
5. OpenLazarus for healthcare practices. OpenLazarus. Accessed 8 Nov 2026.
6. FAQ. OpenLazarus. Accessed 8 Nov 2026.
7. OpenLazarus for professional services. OpenLazarus. Accessed 8 Nov 2026.
8. Pricing. OpenLazarus. Accessed 8 Nov 2026.
9. Security. Glean. Accessed 8 Nov 2026.
10. Gemini for business: the universal work agent. Google Cloud. Accessed 8 Nov 2026.
11. What is Microsoft Copilot?. Microsoft Learn, 5 Nov 2026. Accessed 8 Nov 2026.
12. WRITER pricing. WRITER. Accessed 8 Nov 2026.
13. AIP overview. Palantir Documentation. Accessed 8 Nov 2026.
14. Trusted AI. Salesforce. Accessed 8 Nov 2026.
15. AI implementation partner: systems you own. OpenLazarus. Accessed 8 Nov 2026.

Updated 8 Nov 2026: First published. Seven platforms reviewed against seven criteria.

Further reading

• Guide · Build vs buyBuild or buy an AI operating system for your companyThe choice is rarely binary. Most companies end up buying the platform, owning the data and building only the parts that make them different.
• Guide · GovernancePermissions, approvals and audit for company AI agentsThree controls decide whether an AI agent is safe to let loose on company records: what it can access, what it can do without asking, and whether anyone can see afterwards what it did.
• Guide · CostWhat an AI operating system costs a 200-person companyPer-seat, per-action and per-company pricing produce very different bills for the same headcount. We worked through each model with published list prices.