Everyone is selling a layer. Ask which one you are buying
Model. Wrapper. Tool harness. Enterprise data. Operations. Five rungs, and most of the market stops at four. Below is the question set we would want you to take to every vendor on your shortlist, including us.
Every AI platform sits on a rung. Work out which one
The word "platform" now covers five very different products. Each rung gives you something real and leaves something with you. The rung decides how much of the work you still own after you sign.
Others sell you a smarter tool. We run your operations.
You get intelligence. You still build everything.
You get answers. You still do the work.
You get actions. You still design the workflow.
You get context. You still run the operation.
The operation runs. Your people stay in control.
They all call themselves an Enterprise AI platform. Here is where each one stops
Contact-center agents
Conversational agents for support queues, on voice and chat.
They answer the conversation. The process behind it stays manual.
We run the process the conversation was about, end to end.
Enterprise assistants
Company-wide copilots that search and summarise across your tools.
Useful in every department, shallow in the one that matters.
We go deep in one operation and act in the systems running it.
Sovereign model providers
Frontier models you can host privately or inside your own region.
They sell the model. The orchestration above it is yours to build.
We run on top of them. A partner layer, not a competitor.
Agent frameworks
Open toolkits for developers to assemble their own agents.
A construction kit. You still staff, run and maintain the result.
We hand over a running operation, not the parts to build one.
Vertical point agents
Single-task agents tuned for one document type or one workflow.
Deep on one task, with no layer underneath to run the rest.
We are that layer, with the specialist work sitting on top of it.
Four things every AI platform will pitch you as a differentiator. None of them are
Each of these comes up in almost every vendor conversation as a reason to choose one platform over another. The whole field ships them now. Treat them as the entry requirement, ask what sits above them, and you will get through a shortlist considerably faster. A SOC 2 report belongs on this list too.
Private deployment
Single-tenant and private VPC are standard. On-premise and air-gapped are where the field still genuinely splits.
Access control
Role-based permissions and enterprise SSO ship as standard. Ask instead what an agent is permitted to do.
Audit logging
Action-level logs are expected everywhere. What differs is whether anyone can act on what the log says.
Multi-model routing
Running more than one model provider is now common rather than clever. Ask what happens when you switch.
Three problems no AI platform has solved, ours included
A comparison that only lists our own strengths is marketing, not information, and you would be right to discount it. So here is where the whole category, us included, is still short.
Compounding on your operation
Most platforms perform about the same on day 90 as on day 1. This is the one we are betting the product on, and it is unfinished work everywhere, us included.
Proving the return
Nobody in this category can hand you a clean before-and-after on cost per resolved task. Ask every vendor on your list, ourselves included, and expect a soft answer.
Exceptions without a person
Every platform routes the genuinely hard case to a human. Reducing how often that has to happen is still an open problem across the whole field.
Six questions worth asking every vendor on your list
Including us. If a vendor cannot answer these in a working demo on your own data, the answer is usually no.
Where does the data sit when the model runs?
Ask for the deployment diagram, not the certificate. Compliance logos describe a company. A diagram describes where your data actually goes.
What happens when the agent is wrong?
Ask to see the audit trail and the approval step live, on real data. Every platform claims a human in the loop. Few can show you the gate.
Who runs it after go-live?
Ask who is on the hook at 2am, and whether that turns out to be your team. This is where most platform pricing quietly becomes a hiring plan.
What does it know on day 90 that it did not know on day 1?
This is the question that separates a tool from an operating layer. Ask for the mechanism, not the roadmap.
How many of your systems does it act in, versus read from?
Reading is integration. Acting is operations. The gap between the two is where most pilots stall before they ever reach production.
What happens when you change model provider?
If the answer is a rebuild, you bought a wrapper. If the answer is a config change, you bought a platform.
The question nobody in this market answers well
Enterprise focus. Private deployment. Working on your data. Orchestration. Governance. Five things every serious vendor now claims, and most of them can back. That race is table stakes, not a differentiator. The sixth one compounds: does the system get better at your operation every week it runs?
Every correction is training data
When a supervisor overrides an agent, that decision is captured and folded back into the workflow, instead of disappearing into a ticket.
Every exception becomes a rule
The edge cases your team handles by hand in month one are the automation running unattended in month four.
Every run narrows the next one
Month 12 is not month 1 with more usage. It is a different system, running the same operation with far less of your attention.
This is the slowest moat to build and the hardest to copy. It is also the only one still worth anything in three years.
What fifty AI platforms will put in writing
Every cell is the vendor's own published wording, linked, because a claim in writing is one you can hold them to in a meeting and a good answer on a call is not. Blue Mesh is deliberately not a row: scoring ourselves in a table we built and publish would tell you nothing.
| Platform | Deployment | Multi-model | Voice | Human approval before an action | Published compliance |
|---|---|---|---|---|---|
| Sovereign orchestration | |||||
| Kamiwaza AI | On-premise, edge, cloud and air-gapped, including classified networks. No multi-tenant SaaS published. source | Yes source | Not published | Not published | Not published (no certification claimed anywhere on their site) |
| Operational platform | |||||
| Palantir AIP | AWS, Azure, Google Cloud, Oracle, on-premises, and disconnected/air-gapped. Multi-tenant vs single-tenant wording not published. source | Yes source | Yes source | Yes source | FedRAMP High, DoD DISA IL-5/IL-6, CMMC; HIPAA, GDPR and ITAR referenced. SOC and ISO referenced only generically. source |
| Suite platform | |||||
| IBM watsonx Orchestrate | SaaS on IBM Cloud, SaaS on AWS, on-premises via Cloud Pak for Data, local Developer Edition, and air-gapped. SaaS tenancy model not published. source | Yes source | Yes source | Not published (overview claims it; the node reference defines no approve/reject semantics) | Unverified (ibm.com returns HTTP 403 to automated fetch; the reachable developer docs name no certifications) |
| Microsoft Copilot Studio | Multi-tenant public cloud; GCC and GCC High government clouds; FedRAMP High; EU Data Boundary. True on-premise and air-gapped not published. source | Yes source | Yes source | Yes source | SOC, ISO 27001/27017/27018/27701, HIPAA BAA, FedRAMP, PCI DSS, HITRUST CSF, CSA STAR, UK G-Cloud, GDPR source |
| Salesforce Agentforce | Multi-tenant public cloud on Hyperforce. Single-tenant, private VPC, on-premise and air-gapped not published. source | Yes source | Yes source | Not published | SOC 2, SOC 3, C5 (ISAE 3000), PCI DSS AoC, ISO/IEC 27001:2022 source |
| ServiceNow AI Agents | Unverified | Yes (bring your own LLM, scoped to Now Assist rather than AI Agents specifically) source | Unverified | Unverified | Unverified (trust and compliance pages return HTTP 403) |
| Hyperscaler | |||||
| AWS Bedrock AgentCore | Multi-tenant AWS across 15 named regions, with VPC connectivity across all services. On-premise, GovCloud, sovereign and air-gapped not published. source | Yes source | Not published | Yes (approver not stated to be human) source | BIO, C5, CISPE, CPSTIC, ENS High, FINMA, GNS, GSMA, HITRUST, IRAP, ISMAP, ISO 27001/27017/27018/27701/22301/20000/9001, CSA STAR, MTCS, OSPAR, PCI, Pinakes, PiTuKri, SOC. HIPAA eligible; FedRAMP being pursued. source |
| Databricks Mosaic AI | Vendor-managed multi-tenant control plane on AWS, Azure and GCP, with customer-managed VPC and PrivateLink. On-premise not published. source | Yes source | Not published | Yes source | SOC 2 Type II, C5, CCCS Medium, DoD IL5, FedRAMP High and Moderate, HIPAA, HITRUST, IRAP, ISMAP, K-FSI, PCI-DSS, TISAX, UK Cyber Essentials Plus source |
| Google Vertex AI Agent Buildernow Gemini Enterprise Agent Platform | Not published / unverified | Yes source | Not published | Yes (via ADK component) source | ISO 27001, ISO 27017, ISO 27018, SOC 1/2/3, PCI DSS, Penetration Testing, scoped explicitly to the named agent service. HIPAA and FedRAMP not on that page. source |
| Snowflake Cortex Agents | Fully managed inside the Snowflake account, with cross-region inference routing. On-premise and private connectivity not published on the product page. source | Yes source | Not published | Not published | Platform-wide list (CSA STAR L1, ISO 9001/27001/27017/27018, SOC 1 and 2 Type II, CJIS, DoD IL5, FedRAMP Moderate and High, GovRAMP, IRS 1075, ITAR, NIST 800-171, HITRUST, PCI DSS, IRAP, C5, TISAX AL3, K-FSI, CE+) with no per-feature scope stated. Only FedRAMP Moderate is explicitly scoped to Cortex Agents. source |
| Model provider | |||||
| Cohere (North) | Customer VPC, on-premise, or Cohere-managed Model Vault. Air-gapped not published. source | Yes (vendor blog only, not corroborated in docs) source | Not published | Yes (vendor blog only, no docs page defines the mechanism) source | SOC 2 Type II, scoped explicitly to 'our API platform' not to North. ISO 27001, ISO 42001 and HIPAA appear as badge images with no scope text. source |
| Mistral AI | Varies by product. Studio: hybrid, dedicated, self-hosted, cloud and on-prem. Le Chat Enterprise: self-hosted, customer public or private cloud, or Mistral cloud. API: regional endpoints EU or US. Air-gapped not published. source | Yes (limited: hosts one third-party open model so far) source | speech models yes; voice-agent and telephony not published | Yes (Mistral Work only, not Studio or Le Chat Enterprise) source | SOC 2 Type II, ISO 27001/27701 source |
| Enterprise assistant | |||||
| Aiseraacquired by Automation Anywhere | Unverified (security and platform pages now redirect off-domain to the acquirer) | Not published | Yes source | Yes source | Unverified on any Aisera-owned page |
| Ema | Runs entirely in the customer environment, fully isolated when required. source | Yes source | Yes source | Yes source | SOC 2 Type II, ISO 27001, ISO 27017, ISO 27701, ISO 42001, CSA STAR, GDPR, EU AI Act. HIPAA not published. source |
| Glean | Single-tenant, either Glean-hosted or in the customer's own AWS, Azure or GCP. Air-gapped not published. source | Yes source | Yes (in-app real-time voice; telephony not published) source | Yes source | SOC 2 Type II, ISO/IEC 27001, ISO/IEC 42001:2023, HIPAA, GDPR source |
| Moveworksacquired by ServiceNow | Regional AWS cloud including US GovCloud, EU, Canada, Australia, UK and Japan. Tenancy model not stated. source | Yes source | Not published | Yes source | ISO/IEC 42001, 27001:2013, 27017:2015, 27018:2019, 27701:2019; SOC 2 Type 2; CSA STAR Level 2; GDPR; CCPA; FedRAMP. HIPAA not published. source |
| Writer | Agent deployment via Docker to cloud, on-premises or local; AWS PrivateLink and GCP Private Service Connect for connectors. Core platform tenancy not published. source | Yes source | Not published source | Not published | Unverified (writer.com returns HTTP 403 to fetch) |
| Contact-center agents | |||||
| Cognigy | Managed Kubernetes on AWS EKS, Azure AKS or Google GKE (recommended). On-premise Kubernetes possible but explicitly discouraged and unsupported. source | Yes source | Yes source | Not published | Marketing-level list only: GDPR, ISO 27001, CCPA, HIPAA, SOC 2, PCI DSS. Trust centre unverified (JS-rendered, empty body). source |
| Cresta | Not published | Not published | Yes source | Yes source | SOC 2 Type II, SOC 3, ISO/IEC 27001, ISO/IEC 27701, ISO/IEC 42001:2023, HIPAA, PCI DSS, GDPR, CCPA, CPRA, TISAX source |
| Decagon | Standard SaaS, single-tenant SaaS in a dedicated VPC, cloud-prem in the customer VPC, and on-premise datacentre by exception. Air-gapped described as bespoke, not a packaged tier. source | Yes source | Yes source | Not published | GDPR, CCPA, EU AI Act, HIPAA, SOC 2, ISO (no standard number published), PCI source |
| Forethoughtacquired by Zendesk | AWS infrastructure. Tenancy not published. source | Not published | Yes source | Not published | ISO 27001, SOC 2, HIPAA, GDPR, CCPA, alignment with NIST 800-53 and 800-171 source |
| Kore.ai | Public multi-tenant SaaS, private cloud dedicated VPC (single tenant), and on-premises. Air-gapped not published. source | Yes source | Yes source | Yes source | SOC 2 Type II, PCI DSS, ISO/IEC 27001:2022, GDPR, CCPA, EU AI Act, DESC CSP (trust centre). FedRAMP appears in docs but NOT on the trust centre. source |
| Parloa | Microsoft Azure with regional hosting and data-residency controls. Single-tenant vs multi-tenant not published. No on-prem published. source | Yes (including bring your own STT, TTS and LLM) source | Yes source | Yes source | ISO/IEC 27001:2022, ISO 17442:2020, SOC 2 Type 1, SOC 2 Type 2, PCI DSS, HIPAA, DORA, GDPR, EU AI Act source |
| Sierra | Not published | Yes (Sierra selects the blend; customer model choice not published) source | Yes source | Not published | SOC 2, HIPAA, GDPR, PCI Level 1, FedRAMP High, CCPA, CSA STAR, ISO 27001, ISO 42001 source |
| Teneo.ai | SaaS, private cloud, on-premise (on-prem corroborated by docs; air-gapped not published) source | Yes source | Yes source | Not published | ISO 27001, SOC 2 Type I and II, Cyber Essentials, GDPR / EU AI Act source |
| Yellow.ai | Not published source | Yes source | Yes source | Not published | SOC 2 Type II, ISO/IEC 27001:2022, ISO/IEC 27701:2019, HIPAA, PCI-DSS v4.0.1 source |
| Voice agents | |||||
| PolyAI | Not published source | Yes source | Yes source | Not published | ISO/IEC 27001, SOC 2 Type II, Cyber Essentials and Cyber Essentials Plus, HIPAA, PCI-DSS, GDPR source |
| Voice infrastructure | |||||
| Bland AI | Bland-managed cloud, your VPC, on-premise, air-gapped (all four named) source | No source | Yes source | Not published | SOC 2 Type I and II, HIPAA, GDPR, PCI DSS v4.0 source |
| Retell AI | Shared multi-tenant cloud; Dedicated Stable Server on Enterprise; deployment in the customer's own infrastructure. Private VPC and air-gapped not published. source | Yes source | Yes source | Not published | SOC 2 Type 1 and Type 2, HIPAA (signed BAA required), GDPR source |
| Vapi | Vendor-hosted cloud; on-premise for large enterprises. Private VPC and air-gapped not published. source | Yes source | Yes source | Not published | SOC 2 Type II, HIPAA (signed BAA required), GDPR, PCI source |
| Agent framework | |||||
| Botpress | Managed cloud; tenancy, VPC and on-prem unverified. Self-hosted docs page returns only a JS redirect stub. | Yes source | Not published | Yes (on outbound MESSAGES, not on tool actions) source | Unverified (security, enterprise and privacy pages all HTTP 403; trust portal Drata-hosted and also 403) |
| CrewAI | AMP is managed multi-tenant cloud; Factory is self-hosted on Helm/Kubernetes; on-premise stated on a vendor blog. Air-gapped not published. source | Yes source | Not published | Yes (on task output, not per tool call) source | Unverified (trust centre JS-rendered; no certification on any fetchable page) |
| Dust | Multi-tenant cloud SaaS. Self-hosted NOT published as a supported edition despite the MIT repo. Regions unverified. | Yes source | Not published | Yes (on state-modifying tool calls, with accept/reject links) source | SOC 2 (type not specified), GDPR. ISO 27001 and HIPAA not verified. source |
| LangChain / LangGraph | Platform: multi-tenant cloud, BYOC in the customer VPC, and self-hosted (both Enterprise). Regions GCP US/EU/APAC and AWS US. Air-gapped not published. Framework: self-hosted library. source | Yes source | Not published | Yes (per tool call, before execution) source | SOC 2 Type 2, HIPAA, GDPR. ISO 27001 unverified (trust portal JS-rendered). source |
| LlamaIndex | Framework is a self-hosted library. LlamaCloud offers SaaS, private VPC across all cloud providers, and self-hosting/BYOC on Kubernetes (terms password-gated). Air-gapped not published. source | Yes source | Not published | Yes (developer-built pause primitive, not a packaged product feature) source | SOC 2 Type II, GDPR, HIPAA, stated for LlamaParse specifically rather than the whole platform source |
| Lyzr AI | Lyzr Cloud SaaS, on-premise inside the customer VPC, and hybrid. Air-gapped not published. source | Yes source | Yes source | Yes source | SOC 2 Type II and ISO 27001:2022 have downloadable audit documents. HIPAA and ISO 42001:2023 are listed as programme standards with no report evidenced. source |
| Rasa | Self-hosted on own infrastructure, private cloud, and fully offline / air-gapped. Multi-tenant cloud not named. source | Yes source | Yes source | Not published | Vendor wording is hedged: 'supports SOC 2 Type II compliance'. Their security page claims controls 'aligned with ISO 27002' and names no achieved certification. source |
| Relevance AI | Multi-tenant cloud across US, EU and AU regions; region fixed at signup. Single-tenant stated to be in development. Self-hosted, on-prem and air-gapped not published. source | Yes source | Yes (outbound phone agent via Twilio; inbound not published) source | Yes (per tool, Approval Required mode with a drafted action) source | SOC 2 Type II, GDPR. HIPAA and ISO 27001 not published. source |
| Stack AI | Four deployment models spanning multi-tenant cloud through air-gapped on-premise; on-prem runs entirely in customer infrastructure on any major cloud or own servers. source | Yes source | partial (audio in and out; telephony not published) | Yes (workflow pause for human input) source | SOC 2 Type II, HIPAA, GDPR, ISO 27001 source |
| Vellum | Vellum Cloud managed, VPC on AWS/Azure/GCP, and self-hosted including environments without internet connectivity. source | Yes source | Not published | Not published | SOC 2 Type 2, HIPAA. GDPR and ISO 27001 not published. source |
| Voiceflow | Vendor-hosted cloud. Tenancy, VPC and on-prem not published. source | Yes source | Yes (native telephony) source | Not published | SOC 2 Type II, ISO/IEC 27001:2022, GDPR, HIPAA source |
| Automation | |||||
| Automation Anywhere | Customer VPC / private cloud, plus vendor cloud across 16 datacentres. On-premise and air-gapped not published. source | Yes source | Not published | Not published (referenced but never defined; approvals doc is gated) | SOC 1 Type 2, SOC 2 Type 2, SOC 3, ISO 27001, ISO 22301, HITRUST, GDPR, FIPS-140 encryption. HIPAA and FedRAMP not claimed. source |
| n8n | n8n Cloud is managed multi-tenant hosted in the EU on Azure; self-hosted is offered. Air-gapped and private VPC not verifiable on a fetched page. source | Yes source | Not published | Yes source | Cloud: SOC 2 alignment plus a public SOC 3 report, GDPR. Self-hosted: security is the operator's responsibility. ISO 27001 and HIPAA not published. source |
| UiPath Agentic Automation | Automation Cloud SaaS, Automation Cloud Dedicated (single-tenant), Automation Suite self-hosted, Automation Cloud Public Sector. Air-gapped not published. source | Yes source | Not published | Yes source | ISO 27001, ISO 27017, ISO 27018, ISO 9001, HITRUST; SOC 1, SOC 2, HIPAA, C5 attestations; FedRAMP for Public Sector; ISO/IEC 42001 and EU AI Act alignment source |
| Zapier Agents | Multi-tenant cloud only, hosted on AWS in the United States. VPC Peering is network connectivity to the customer VPC, NOT a customer-hosted deployment. No single-tenant, on-prem or air-gapped option published. source | Yes (verified for AI by Zapier; model choice within Agents specifically not published) source | Not published | Yes source | SOC 2 Type II, SOC 3, GDPR, CCPA. ISO 27001 and HIPAA not claimed. source |
| Vertical point agent | |||||
| Abridge | Not published | Not published | Yes (ambient capture, not telephony) source | Yes source | HIPAA, SOC 2 Type 1, SOC 2 Type 2, CCPA, TX-RAMP, WCAG. HITRUST and ISO 27001 absent. source |
| Eleos Health | Browser extension over any web-based EHR. Tenancy not published. source | Not published | Yes (ambient audio; telephony not published) source | Yes source | HIPAA, HITRUST, SOC 2 Type II, ISO 27001, ISO 27799, ISO 42001 source |
| Hippocratic AI | Not published | Not published | Yes source | Not published | HIPAA, SOC 2, HITRUST e1 (on-site badges only) source |
| Coding agent | |||||
| Cognition (Devin) | Cloud, with dedicated or on-prem for Enterprise; customer data stored in the customer tenant. source | Not published | Not published | Yes source | SOC 2 Type II, ISO/IEC 27001:2022, CCPA. No HIPAA, no GDPR listed. source |
| ITOps | |||||
| BigPanda | Not published | Not published | Not published | Not published | SOC 2 Type II; security framework described as ISO 27002:2013-based, which is alignment not certification. source |
How this table is built
Public sources only. We read each vendor's own site, documentation and trust pages, and we link the exact page a claim came from. We do not use review sites, analyst notes or press coverage.
We publish no scores, no rankings and no judgement of anyone's quality. Only what each vendor states about itself.
Last verified 18 August 2026. We recheck quarterly, and a row whose sources we have not rechecked in ninety days loses its detail until we have.
If something here is wrong or out of date about your product, write to compare@bluemesh.ai and we will correct it and note the change.
All product names and trademarks are the property of their respective owners and are used here for identification only. No affiliation or endorsement is implied.
Three times you should buy something else
A comparison page that never says this is an advertisement. Here is where we lose, honestly.
You want a chatbot
If the job is answering questions on a website, a dedicated support product will do it faster and cheaper than we will.
You have no process to run
We compound on a real operation with real exceptions. Without one, there is nothing for the system to learn from.
You need it live next week
Deep integration into your systems takes weeks, not days. One workflow as a pilot is the honest starting point.
Send us your evaluation criteria
Bring the scorecard you are running every vendor through. We will fill it in honestly, including the rows where we are not the right answer for you.