ChatGPT Enterprise, Microsoft 365 Copilot and Claude Enterprise give each employee an assistant. Odokai gives your team agents that run its processes. See how the four compare on cost, fifteen capabilities and thirteen enterprise controls, with every figure sourced.
Most UK SMEs choosing an AI platform end up with the same shortlist. Copilot is already on the Microsoft bill, ChatGPT is the name everyone knows, and someone is usually trying Claude. The question is which of them will take work off your team, and which is just another subscription.
If you want an assistant for every employee, one of the incumbents is likely the right buy, and the tables show which. If you want agents and apps built around how your business runs, Odokai gives you your choice of hundreds of models, your data where you put it, knowledge that belongs to your business, and a fixed monthly bill.
What each one is, in a paragraph.
A sales-negotiated workspace putting the ChatGPT assistant in front of the whole organisation. SSO and SCIM, role-based access, a compliance API covering eDiscovery, DLP and SIEM export, data residency across ten storage regions, customer-managed encryption keys, custom retention and SLAs. The default choice for a large organisation that wants one standalone assistant, independent of any office suite, with mature controls.
Not a standalone product: a $30 per user per month add-on to a qualifying Microsoft 365 plan, or $21 as Copilot Business for up to 300 users. In exchange, AI sits inside Word, Excel, PowerPoint, Outlook and Teams, grounded in the tenant's own data through Graph, inheriting the entire Microsoft compliance stack. The path of least resistance inside a Microsoft estate, and weak outside one.
One seat covers the full suite: chat, Claude Code, delegated multi-step work, design, connectors to Google Workspace, Microsoft 365, GitHub and Slack, plus enterprise search, memory, skills and projects. The $20 seat fee covers access only; all usage is billed separately at API rates, with optional spend limits for the organisation and each user. Strong for heavy, variable users.
A multi-agent operating system: private, persistent agents set up around your team and the processes it runs. Deployed as a shared managed service in AWS London, or single-tenant in an AWS account we manage or in your own AWS account. Built for document-heavy and regulated work, with a consultancy service that gets your first workflow live.
Compiled in September 2026 from vendor documentation and third-party procurement reporting. No vendor commissioned it. It was produced by an assistant running on the Odokai platform, so Odokai's entries are vendor-sourced.
Seat prices only tell part of the story. Check what each price leaves out, and who pays when usage runs higher than planned.
Swipe or scroll horizontally to compare all columns.
| ChatGPT Enterprise | M365 Copilot | Claude Enterprise | Odokai | |
|---|---|---|---|---|
| Published price | Quote only. Reported at $45–75 per user per month, around $60 typical. | $30 per user per month, annual, as an add-on to a qualifying base licence. Copilot Business is $21 for up to 300 users. | $20 per seat per month, annual, plus all usage at API rates. | Fixed monthly fee with included credits. No figure published; the pricing model is set out separately. |
| Realistic all-in | ~$60 per user per month, reported. | $69–$90 per user per month once the E3 or E5 licence underneath is counted. | ~$37 light use, ~$87 heavy use. Variable, with optional spend limits. | The fixed fee. Infrastructure cost applies only in single-tenant mode. |
| Seat minimum | Around 150, reported. An implied floor near $108k a year. | None, but the base licence is compulsory. | 20 self-serve, 50 sales-assisted. | A seat limit per tenancy rather than a floor. |
| Usage billing | Credit pools, or token-based billing on some contracts. | Mostly included. Cowork and some agent usage are metered in Copilot Credits. | Everything billed separately at API rates. No included allowance. | Monthly credit cap with a warning threshold and a soft stop when requests are admitted. |
| Who carries the overrun | Buyer, moderate. Higher on token billing. | Buyer, low. | Buyer, high, unless spend limits are set. | Buyer, low. The cap is buyer-set and enforced. |
| Procurement route | Sales-negotiated, annual, twelve-month term. | Self-serve or Enterprise Agreement, annual. | Self-serve or sales-assisted, multi-currency on assisted. | Sales-assisted: a call to scope the first workflow, then a fixed fee in writing. |
All figures in USD as published or reported in September 2026, excluding tax. Microsoft's SMB bundle promotions ended on 30 June 2026 and base prices rose on 1 July 2026 (E3 $36 to $39, E5 $57 to $60), with the add-on itself held at $30. Partners report enterprise Copilot promotions running to 30 September 2026. Vendor pricing changes often; re-verify at contract time.
An assistant writes text. Odokai agents read your exports, spreadsheets and documents and answer with the figures, charts and tables you would build yourself, with every step on the record.
Fifteen capabilities, assessed the same way for all four platforms.
Swipe or scroll horizontally to compare all columns.
| ChatGPT Enterprise | M365 Copilot | Claude Enterprise | Odokai | |
|---|---|---|---|---|
| General assistant chat | ||||
| Embedded in the office suite | Word, Excel and PowerPoint add-ins | Its core strength | Excel, Word and PowerPoint add-ins; Outlook in beta | Via connectors, Microsoft 365 included |
| Grounding in company data | Through Graph | With citations and per-file grants | ||
| Custom agents or assistants | GPTs and workspace agents | Copilot Studio | Skills | The central concept |
| Model choice and bring-your-own model | Own models only | OpenAI and Anthropic models; bring your own through Foundry in Copilot Studio | Own models only | Any provider with an OpenAI-compatible API, including Anthropic, OpenAI, Gemini, OpenRouter or a local model |
| Agentic multi-step workflows | Cowork, Researcher and Agent Mode | As durable tasks | ||
| Developer tooling | Codex | Agents Toolkit and SDK; GitHub Copilot sold separately | Claude Code | Shell, Git, SDK, CLI, browser control |
| Persistent memory across sessions | At platform level rather than per user | |||
| Background and scheduled work | Tasks | Scheduled prompts and Copilot Studio triggers | Cowork scheduled tasks and Claude Code routines | Durable across deploys, and scheduled |
| Multi-agent orchestration | Subagents in ChatGPT Work and Codex | Child and connected agents in Copilot Studio | Subagents, workflows and agent teams | Panel, pipeline, triage, debate, map-reduce, delegation trees |
| Human approval gates on tool calls | Admins set per-app action limits and ask-before-write permissions | Per-tool approval in Copilot Studio agents | Admins set "Needs approval" per connector and tool, by role | Per-tool policy, dual control, SLAs |
| PII redaction before the model sees it | DLP blocks sensitive content; no redaction | Redaction; DLP blocking in beta | On by default | |
| Internal app building | Canvas | Frontier | Artifacts with storage and connectors | App Studio |
| Broad third-party integrations | 100+ connectors, Power Platform and MCP | 146 reviewed connectors plus 8 first-party anchors | ||
| Versioned shared workspace | Shared projects | SharePoint | Shared projects, not versioned | Versioned |
Odokai entries come from Odokai's own documentation and have not been independently audited. "Partial" covers preview features, connector-based support, and anything not offered as a built-in capability.
Thirteen controls. The incumbents lead on certification. All four offer approval gates on tool calls. Odokai is the only one that redacts PII before the model sees it.
Swipe or scroll horizontally to compare all columns.
| ChatGPT Enterprise | M365 Copilot | Claude Enterprise | Odokai | |
|---|---|---|---|---|
| SSO and SAML | Entra | Provided as part of your engagement, on the shared platform or your own instance. | ||
| SCIM provisioning | Implemented as part of your engagement. | |||
| SOC 2 or ISO 27001 | SOC 2, ISO 27001 and ISO 42001 | Via Microsoft 365 | SOC 2 Type II, ISO 27001 and ISO 42001 | ISO 27001 in progress. In your own AWS account, the infrastructure sits under your certifications and AWS's. |
| Compliance API for eDiscovery, DLP, SIEM | Auditor exports as CSV, signed JSON and PDF, but no API. | |||
| Audit logs | Purview | Classified, encrypted, roughly seven-year retention, admin viewer. | ||
| Customer-managed encryption keys | In single-tenant. Customer-owned KMS keys, rotation and CloudTrail. | |||
| Data residency options | Ten regions | Microsoft 365 geo and EU Data Boundary, excluding Anthropic models | US-only inference option; no EU or UK | London for the managed service, the customer's own region single-tenant, or local models. |
| No training on customer data | Default | Default | Enforced through an approved-provider allowlist. | |
| Multi-factor authentication | TOTP, recovery codes, organisation-wide enforcement. | |||
| Role-based access control | Typed permission catalogue, default deny. | |||
| Approval gates on tool calls | Per connector | Per tool in Copilot Studio | Per connector and tool | Dual control, SLAs, fail-closed. |
| PII redaction before the model | DLP blocks sensitive content; no redaction | Redaction; DLP blocking in beta | On by default. | |
| Data classification policy | Purview sensitivity labels and DLP | Via Purview and DLP hooks | Per-class rules, including private-model-only. |
Odokai's entries are vendor-sourced. Incumbent entries come from public vendor documentation, with certifications listed only where publicly claimed.
Six ways Odokai works differently from the other three platforms.
A shared managed service in AWS London, single-tenant in an AWS account we manage, or single-tenant inside your own AWS account with its own VPC, database and KMS keys. The other three are SaaS services. If your data can't leave your environment, Odokai can run inside it.
Any provider with an OpenAI-compatible API, including Anthropic, OpenAI and Microsoft models, plus hundreds more through OpenRouter and Gemini, and thousands of open models through Ollama, routed per workload and per data class. ChatGPT Enterprise and Claude Enterprise use their own models; Copilot offers OpenAI and Anthropic models. Only Odokai runs local models, so sensitive work never has to leave your environment.
Agents hold domain knowledge, memory and skills, and the platform keeps what they learn. All four have persistent memory, but in the other three it belongs to a user or a project. In Odokai it builds up as shared organisational knowledge, so it stays when people leave.
Approval gates with dual control, PII redaction on by default, per-class data rules including private-model-only, and permissions that deny by default. The incumbents offer approval gates, and Copilot classifies data through Purview. None of them redacts PII before the model sees it.
A fixed monthly fee with included credits, a warning threshold, and a cap you set. Claude Enterprise bills all usage at API rates with no included allowance; spend limits can stop it, but not reduce the rate.
Panels, pipelines, triage, debate, map-reduce and delegation trees, plus background work that survives a deployment and scheduled work that runs without anyone present. The incumbents run subagents and scheduled tasks; Odokai gives you these patterns ready to use.
Microsoft 365 Copilot, comfortably. The work already happens in Word, Excel, Outlook and Teams, the identity and compliance stack is already in place, and deployment friction is close to zero. ChatGPT Enterprise alongside it where a meaningful part of the organisation works outside Microsoft. If what you want is an assistant for everyone, this is the assistant to buy.
Claude Enterprise. The strongest developer tooling of the four, one seat covering chat, coding and delegated work. The trade is usage-based billing; ChatGPT Enterprise's standard seats, which include Codex, are the alternative where fixed seat pricing matters more.
One of the three incumbents, if your board needs the vendor scale and certification those platforms hold today. Odokai is a small UK company, live for over a year, with no SOC 2 or ISO 27001 certification yet; an ISO 27001 plan is in progress. Source code escrow and other contractual protections are available for an Odokai engagement.
Copilot's fixed add-on, a fixed per-seat Claude Team plan, or Odokai's fixed fee with a buyer-set cap. Each avoids Claude Enterprise's usage billing, where every token is billed at API rates with no included allowance.
Odokai. Agents hold the context of your business, apps are built around your processes, and work runs on a schedule whether or not anyone is watching. You choose from hundreds of models, your data stays where you put it, and there is an auditable record of everything the agents know and produce. The incumbents give every employee an assistant. Odokai gives your business a workforce.
OpenAI's help centre, documentation and release notes; Microsoft Learn and Microsoft's Copilot pricing and licensing pages, plus partner reports for current promotions; Anthropic's help centre, pricing page and trust centre; and for Odokai, its own published platform documentation checked against its repository on 13 September 2026. Incumbent entries were last checked on 27 September 2026. The full assessment, including figures and method, is available as a PDF.
If you want the best assistant, buy Copilot where the work runs on Microsoft, Claude where engineers want coding agents, and ChatGPT Enterprise where familiarity matters most.
If you want AI that becomes part of how your team works, Odokai is an agentic operating system with a consultancy service to get you started. You build your own agents, apps and workflows on it, using the models you choose, in AWS London or your own AWS account, with a fixed monthly bill.
Notes on governed models, operational automation, and how teams move from AI-assisted pilots to broader adoption.
18 October 2026
The ICO has opened a call for evidence on agentic AI, closing 20 November. What it covers, six checks before an agent touches personal data, and what does not change yet.
Read more16 September 2026
Five stages, three approaches, and a warning about the traps. How SMEs cut invoice handling time by 70 to 90 percent, and what the first working version looks like.
Read more