Enterprise AI platforms compared: 15 criteria, 12 platforms

Most enterprise AI projects do not fail on the model. They fail on a decision made months earlier, in a procurement meeting, when nobody could explain how three shortlisted products actually differed.

The numbers back this up. MIT’s 2026 study found that 95% of enterprise generative AI pilots delivered no measurable return. IDC put the pilot failure rate at 88%. Gartner expects more than 40% of agentic AI projects to be cancelled by 2027, and Deloitte found only 21% of organisations have a mature governance model for autonomous agents. Fiddler AI’s production data is the most sobering: agents that succeed 60% of the time on a single run drop to 25% across eight consecutive runs under production load.

None of that is a model problem. It is a fit problem — the wrong platform for the data, the permissions, the deployment constraints, and the budget the organisation actually has.

So we did the work we would want a vendor to do for us: a 15-criteria framework, and a feature-by-feature matrix of all twelve platforms — ours included — built from public documentation and vendor announcements, with every claim linked to its source.

Method and caveat. Features below were compiled from vendor documentation, release notes, and public announcements current to August 2026, and every claim links to its source. This market ships monthly, and almost every enterprise contract is custom-quoted, so treat prices as order of magnitude and verify features against current documentation before you sign. “Partial” means the capability exists but is limited to one ecosystem, gated behind a paid tier, or narrower than a competitor’s equivalent.

First, the market is four markets

The word “platform” is doing a lot of hiding. Four distinct product shapes are sold under it, and they are not substitutes.

Suite-native assistants. AI added to a stack you already bought: Microsoft 365 Copilot with Copilot Studio, Google Gemini Enterprise (which absorbed Agentspace), Salesforce Agentforce 360, ServiceNow Otto — the unified layer that now folds in Now Assist and the Moveworks assistant ServiceNow acquired for $2.85B — and IBM watsonx Orchestrate. Cheapest to start, strongest inside their own boundary, weakest outside it.

Work-AI and search-first platforms. Vendor-neutral layers that index everything and answer across it: Glean (roughly $200M ARR at a $7.2B valuation, 100+ connectors), Dust, Writer, ChatGPT Enterprise, and Sana, now part of Workday. They win on cross-system retrieval and vary enormously on whether they can do anything with what they find.

Sovereign and private-deployment platforms. Built for organisations that cannot send data to a shared cloud: Cohere North, which runs on-premise, in a VPC, or air-gapped on as few as two GPUs, and our own platform. Smaller ecosystems, much harder constraints solved.

Build-your-own toolkits. Onyx, Dify, n8n, Flowise, LibreChat. Free or nearly free at the licence line, excellent if you have a platform team. The cost moves from the invoice to the payroll.

Knowing which of the four you are shopping for eliminates most of the confusion before you write a single requirement.

The 15 criteria

Four questions a buyer has to answer: what does it know, what can it do, who controls it, what does it really cost.

What it knows

1. Connector breadth and freshness. How many systems, and does the index re-sync automatically or drift stale? Ask for the re-sync interval, not the connector count.

2. Permission inheritance. Does the assistant respect the source system’s access rules per document, or flatten everything into one index? This is the most common cause of a stalled security review. Some platforms charge extra for it.

3. Citation and verifiability. Does every answer link back to a source a human can open? Without this you cannot use the output in a regulated process.

4. Answer format. Text only, or tables, charts, forms, slides, spreadsheets? A dataset rendered as a paragraph is a worse answer than a dataset rendered as a table.

What it can do

5. Write operations. Can it take action in your tools, or only read from them? A large share of “AI agent” products are read-only retrieval with an agentic name.

6. Approval model. An explicit pending → approved → executed gate before anything is written, or does the agent act and log afterwards?

7. Identity for actions. Does the agent sign in as the individual user, or as one shared service account? This determines whether an audit can attribute an action to a person. Shared service accounts are the quiet governance failure of 2026.

8. Autonomy modes. Chat only, or scheduled runs and event triggers? Real savings come from work that happens before anyone asks.

9. Extensibility. Custom connectors, MCP support, installable skill packages, sandboxed code execution. Ask what happens when your workflow needs the one integration nobody has built yet.

Who controls it

10. Deployment and data residency. SaaS-only, EU region, your VPC, on-premise, air-gapped. Note that if your cloud provider is US-incorporated, an EU region alone does not give you full sovereignty.

11. Model flexibility. One vendor’s models, or your choice per workload — and can you switch without rebuilding the knowledge base?

12. Governance and audit depth. Workspace isolation, RBAC, SSO and SCIM, replayable audit trail, configurable retention. From 2 August 2026 the EU AI Act’s transparency obligations apply, and deployers of high-risk systems must retain generated logs for at least six months. Log retention stopped being a nice-to-have.

13. Compliance posture. SOC 2 Type II, ISO 27001, ISO 42001, GDPR, HIPAA where relevant. Ask for the trust centre link, not the sales claim.

What it costs

14. Cost model and predictability. Per seat, per conversation, per credit, per vCPU-hour, or a mix. Consumption models are where budgets break, because the cost of an answer varies with how much retrieval and tool use it does.

15. Time to value and who does the work. Weeks or quarters, and does the vendor configure it or hand you a console? This is where the 88% pilot failure rate lives.

Feature comparison

Knowledge and retrieval

Platform	Connector reach	Permission-aware retrieval	Citations	Web search	Deep research mode	Rich answer formats
Mitigate AI	Files, SharePoint, Drive, Notion, Confluence, Jira, web crawl; daily re-sync	Yesper-workspace isolation	Yes	Yes	Partial	Yescharts, tables, forms, tabs streamed in chat (OpenUI)
Microsoft 365 Copilot	Deep in Microsoft Graph, thinner outside	Yeswithin Microsoft	Yes	Yes	YesResearcher and Analyst agents	YesOffice artifacts
Gemini Enterprise	Validated connectors plus BYO-MCP	Yeswithin Google	Yes	Yes	Yes	YesWorkspace artifacts
ChatGPT Enterprise	Apps and connectors: Salesforce, ServiceNow, enterprise RDBs, vector DBs	Yes	Yescompany knowledge cites sources	Yes	Yes	Yesdocs, spreadsheets, canvas
Glean	100+ connectors, permissions-aware knowledge graph	Yesstrongest in category	Yes	Yes	YesAgent Sandbox, programmatic tool calling	Yesslides, docs, Canvas co-authoring, image generation
Dust	60–100+ connectors plus MCP	YesRBAC with dual-layer agent permissions	Yes	Yes	Partial	Partial
Cohere North	Built-in connectors plus Compass search index	Yes	Yes	Partial	Partial	Yestables, documents, presentations
Agentforce 360	Deep in Salesforce; Data 360 for the rest	Yeswithin Salesforce	Yes	Partial	Partial	YesIntelligent Context over unstructured data
ServiceNow Otto	Deep in ServiceNow plus Moveworks enterprise search	Yes	Yes	Partial	Partial	YesAI Data Explorer, plain-language querying
watsonx Orchestrate	Integration catalog plus custom tools	Yes	Yes	Partial	Partial	Partial
Onyx (open source)	40–50+ connectors, MCP	Yesper-document mirroring, paid tier	Yes	YesSerper, Brave, SearXNG, Firecrawl, Exa	YesDeep Research mode	Partial
Dify / n8n / Flowise	Build your own	Build your own	Build your own	Yes	Build your own	Build your own

Agents and automation

Platform	No-code agent builder	Multi-agent orchestration	Scheduled runs	Event triggers	Approval gate on writes	Code execution sandbox	Computer use
Mitigate AI	Yesper-workspace configuration	Partial	Yes	Yesincluding email events	Yeson every write	Yessecure sandboxes	No
Microsoft 365 Copilot	YesCopilot Studio	Yesagent-to-agent via Work IQ	Yes	Yes	Configurable	Yes	YesGA, browser and desktop
Gemini Enterprise	YesAgent Designer, visual flowchart	YesAgent Registry and Gateway	Yes	Yes	Semantic governance policies	Yes	Partial
ChatGPT Enterprise	Partialagent mode from a prompt	Partial	Partial	Partial	Partial	Yes	Yesagent mode
Glean	YesAgent Builder with debug mode	YesAgentic Engine 2	Yes	Yes	Varies by agent	YesAgent Sandbox	No
Dust	Yesno-code, anyone on the team	Yeschained multi-step agents	Yes	Yes	Varies by agent	Yesvia MCP code interpreter	No
Cohere North	Yesassistant builder	YesNorth Automations	Yes	Yes	Yes	Yes	Partial
Agentforce 360	YesAgentforce Builder, Agent Script	YesAtlas Reasoning Engine, hybrid reasoning	Yes	Yes	Yes	Partial	Partial
ServiceNow Otto	Yes	YesAction Fabric, "agent of agents"	Yes	Yes	Yesapprovals across departments	Partial	Partial
watsonx Orchestrate	Yesdrag-and-drop and natural language	YesAgentic Control Plane	Yesnative scheduling	Yes	Yes	YesADK, Python	No
Onyx	Yescustom agents with instructions and actions	Partial	Partial	Partial	Build your own	Yes	No
Dify / n8n / Flowise	Yesvisual canvas	Yes	Yes	Yes	Build your own	Yes	Partial

Surfaces and extensibility

Platform	External-facing widget	Slack / Teams	MCP support	Custom connectors	Skill or agent catalog	API and SDK
Mitigate AI	YesSSO, guest, or anonymous modes	YesSlack and Teams	Yes	Yesbuilt to order	Yesinstallable skills	Yes
Microsoft 365 Copilot	Via published Copilot Studio agents	Teams native	Yes	YesPower Platform connectors	Yesagent store	Yes
Gemini Enterprise	Partial	Google Chat native	YesBYO-MCP	Yes	YesAgents gallery, Skills Registry	YesAgent Development Kit
ChatGPT Enterprise	No	Via apps	Yes	Yes	Yesapps and connectors	Separate OpenAI API
Glean	Partialinternal-first	Yes	Yesremote MCP servers	Yes	YesLibrary, agent marketplace	Yes
Dust	Partial	Slack native	Yesadmin-added MCP servers	YesREST, webhooks, OAuth2	Yesshared agent workspace	Yes
Cohere North	Partial	Yes	Partial	Yesflexible APIs	Partial	Yes
Agentforce 360	Yesstrong customer-facing story, voice with SIP	Slack native	Yes	YesMuleSoft	YesAgentExchange	Yes
ServiceNow Otto	Yesportal and voice agents	Yes	YesAction Fabric, runs headless	Yes	Yes	Yes
watsonx Orchestrate	Yes	Yes	Yes	Yes	YesAgent Catalog, 150+ prebuilt agents and tools	YesADK, Python
Onyx	Partial	Slack and Discord bots	Yesincluding its own MCP server	YesOpenAPI actions	Partial	Yes
Dify / n8n / Flowise	Build your own	Build your own	Yes	Yes	Community templates	Yes

Governance and security

Platform	SSO / SCIM	Per-user identity for actions	Workspace isolation	Replayable audit trail	Retention and residency controls	Published certifications
Mitigate AI	Entra ID, Okta, Auth0, Google, Keycloak	Yesagent signs in as each person, never a shared account	Yesend-to-end, knowledge and users separated	Yesevery action, replayable	Yesconfigurable per workspace	Not published — residency solved by self-hosting
Microsoft 365 Copilot	Entra native	Yeswithin Microsoft	Yes	YesPurview	Yes	Extensive
Gemini Enterprise	Google native	Yeswithin Google	Yes	Yes	Yes	Extensive
ChatGPT Enterprise	Yes	Partial	Yesworkspaces	Yesdetailed audit logs	Yesdata residency, no training on your data	SOC 2
Glean	Yes	Partial	Yes	Yes	Yes	SOC 2, enterprise controls
Dust	SAML, OIDC, SCIM	Yes	YesSpaces	Yes365-day retention	YesEU or US residency	SOC 2 Type II, GDPR, HIPAA-ready
Cohere North	Yes	Yes	Yesfull data isolation	Yes	Yesincluding air-gapped	SOC 2 Type II, ISO 27001, ISO 42001, GDPR, HIPAA, CCPA
Agentforce 360	Salesforce native	Yes	Yes	YesAgent Health Monitoring, error and escalation rates	Yes	Extensive
ServiceNow Otto	ServiceNow native	Yes	Yes	YesAI Control Tower	Yes	Extensive
watsonx Orchestrate	Yes	Yes	Yes	YesAgentic Control Plane	Yes	Extensive
Onyx	Paid tiers only	Depends on setup	RBAC in paid tiers; community edition is all-or-nothing	Paid tiers	You control it entirely	SOC 2 Type II in enterprise edition
Dify / n8n / Flowise	Paid tiers	Build your own	Free tiers are all-or-nothing	Build your own	You control it entirely	Varies

Deployment, models, and commercials

Platform	Deployment options	Model choice	Cost model	Reported price	Time to production
Mitigate AI	Our cloud, or self-host via Docker Compose, Kubernetes, AWS EKS	OpenAI, Anthropic, Google, Mistral, DeepSeek, OpenRouter — per workspace, swap anytime	Flat subscription with per-workspace budget caps	From €400/month up to 50 users; e-commerce workspace from €600/month; enterprise custom	Under a week, vendor-run
Microsoft 365 Copilot	Microsoft cloud only	Microsoft-hosted, plus Anthropic Claude in Researcher and Copilot Studio	Per seat plus consumption credits	$18–30/user/month plus M365 base; Copilot Studio credits $200 per 25,000, or $0.01 each	Days to start, months to govern
Gemini Enterprise	Google Cloud only	Model Garden, 200+ models	Per seat plus agent compute, memory, storage	$21–30/user/month typical, up to $60+; vCPU-hour, GiB-hour and GiB-month metering on top	Weeks
ChatGPT Enterprise	OpenAI cloud, residency options	OpenAI models only	Per seat, negotiated	~$60/user/month reported, ~150-seat minimum, ~$108K/year floor; Frontier custom	Days to weeks
Glean	Vendor cloud, private cloud options	Multiple providers	Per seat plus AI add-on plus FlexCredits	~$45–50/user/month plus ~$15 AI add-on, ~100-seat minimum; median contract ~$97.5K/year	Weeks to months
Dust	Vendor cloud, EU or US residency	Claude, GPT, Mistral	Per seat, credits metered	$30/seat/month Pro, $150/seat/month Max; enterprise custom	Days to weeks
Cohere North	On-premise, VPC, hybrid, air-gapped, from 2 GPUs	Cohere models	Custom	Custom	Weeks, with deployment engineering
Agentforce 360	Salesforce cloud, plus AWS	Salesforce-managed	Per conversation, Flex Credits, or per seat	$2/conversation, or $500 per 100K Flex Credits, or from $125/user/month; Data Cloud extra	Weeks to months
ServiceNow Otto	ServiceNow cloud	ServiceNow-managed	Custom	Not published	Months
watsonx Orchestrate	IBM cloud, hybrid	watsonx plus third-party, any framework	Subscription plus usage	Essentials from $500/month; standard custom	Weeks to months
Onyx	Your infrastructure, or Onyx Cloud	Any — including local Ollama, vLLM, LiteLLM	Open source plus paid tiers	Community $0 (MIT); cloud ~$16–20/user/month; enterprise custom	Days to self-host, months to harden
Dify / n8n / Flowise	Self-host or vendor cloud	Any	Open source plus paid cloud	Free self-hosted; paid cloud tiers	Fast to prototype, slow to production

What the tables do not show

Three things matter more than any single row.

The suite trap is real, but not always a trap. Microsoft Copilot has the widest deployment in the world — around 150 million seats, with daily utilisation reported at 28–32%. That gap is not a Microsoft failure. It is what happens when an assistant can see Outlook, Teams and SharePoint but not Salesforce, Confluence, your ERP, or the custom database where the answer actually lives. If 90% of your knowledge genuinely is in one vendor’s stack, buy that vendor’s assistant and stop reading. If it is not, a suite-native assistant will be excellent at a fraction of the job.

Search is not action. The search-first generation is superb at finding things. Glean’s retrieval is the strongest in the category — and its reported 25% first-year churn tells you what happens when finding the answer is where the product stops. The question that separates a demo from a deployment: after the assistant finds the invoice discrepancy, who fixes it? Note how many cells in the approval-gate column read “varies by agent” — that is the difference between a governed action and a hopeful one.

Consumption pricing is a governance question, not a finance one. Copilot Studio bills credits per feature, so a single agent response can burn several depending on how much retrieval it does. Agentforce charges per conversation or per credit. Gemini’s agent platform bills compute by vCPU-hour, memory by GiB-hour, storage by GiB-month, with billing switches flipping through 2026. A mid-market Agentforce deployment can run $6,650–18,800 per month once Data Cloud is included. None of that is unreasonable — but you cannot forecast it, and an unforecastable line item is the one finance kills first.

Where we fit

We are in the sovereign bucket, deliberately.

Our platform. Model-agnostic by design — pick the model per workspace and swap it when something better ships, with no retraining and no re-ingestion. Every write passes a pending → approved → executed gate, and the agent signs into each tool as the individual user rather than a shared service account, so an audit can name a person instead of a robot. Workspaces are isolated end to end: the finance agent cannot see HR data. Deployment runs in our cloud or on yours via Docker Compose, Kubernetes or AWS EKS, with configurable retention and residency. Cost is tracked per workspace, per provider, per user, with budget caps — a flat subscription rather than a credit meter. Answers render as charts, tables and forms inside the conversation rather than walls of text. Pricing starts at €400/month rather than a six-figure floor, and we run the deployment ourselves — ingestion, connectors, onboarding — so production is a week away, not a quarter.

How to actually run the evaluation

Score each of the 15 criteria 0–3 for your context, weight the ones your compliance function cares about at double, and shortlist three. Then, before signing anything:

Pick one real workflow that crosses at least three systems. Not a demo. Something a named person does weekly and hates.
Run it end to end on each shortlisted platform with your own data. Retrieval quality on your documents is the only benchmark that predicts anything.
Force a write. Make the agent change something in a live system, then audit who did it. Many products fail here quietly.
Model the cost at 10x the pilot volume. Ask the vendor to do it in writing.
Ask what happens in month 13. Renewal price, data export, connector ownership.

The organisations that reach production are not the ones with the best model. They are the ones that picked a platform matching their permissions, their infrastructure and their budget — then deployed one workflow properly instead of ten badly.

If you want a second opinion on your shortlist, including an honest read on whether we belong in it, get in touch.

Sources

Market data and adoption

Microsoft

Google

OpenAI

Glean, Dust, Cohere

Salesforce, ServiceNow, IBM

Open source

Regulation and sovereignty