# The cost of AI access for employees: two models

> A four-line budget for the cost of AI access for employees: assistant seats, the access layer, connectors and admin time, worked for 50 and 200 people.

**TL;DR** The cost of AI access for employees has four lines: assistant seats, the access layer, connectors and admin time. Elaichi prices the access layer per seat: Gold lists at $15 per user per month in USD, or $10 per user per month billed annually, and Auditor, Billing Admin and Guest seats are free. Building the access layer yourself is the right call in an air-gapped network or in front of a bespoke legacy system, and expensive almost everywhere else.

## What makes up the cost of AI access for employees

Some companies give employees AI assistants access to real work systems: email, CRM, ticketing, code and internal docs. Every one of them pays for four things, whatever the vendor. A three-person pilot usually only shows the first one on an invoice. A rollout past that size surfaces all four, and finance will ask about each.

1. **Assistant seats.** One license per person who chats with Claude, ChatGPT, Cursor, or similar.
2. **The access layer.** The software sitting between those assistants and your internal apps, enforcing who can call what, under whose credentials, and logging that it happened. MCP (Model Context Protocol) is the standard way an AI assistant calls tools in other apps. You either write MCP servers yourself, run an open-source gateway, or buy [a managed control plane](/blog/what-is-an-mcp-gateway/).
3. **Connectors.** One integration per external app you want an assistant to reach, plus maintenance every time that app changes its API.
4. **Admin time.** Recurring hours, at somebody's loaded hourly cost. They go on mapping identity-provider groups to roles, deciding access requests and running offboarding. Add the hours spent producing evidence when a security questionnaire or audit arrives.

Only assistant seats have a public, comparable price tag. The other three are where most budget estimates break down: either omitted entirely, or guessed without a method. Pull seat prices from the vendor's current pricing page on the day you build the model. List prices and tiers move, and a stale number is worse than a labeled estimate.

Two of the four stay as variables in any version of this formula. One is the assistant seat price, call it **S**. The other is your admin's fully loaded hourly cost, call it **A**. The two worked models below price the access layer and connectors against one vendor, Elaichi. S and A stay as inputs you supply from your own numbers. Treat the Elaichi figures as a worked example, not a universal benchmark. The per-seat comparison across vendors is in [what AI access costs per user](/blog/mcp-gateway-pricing/).

## Worked cost model: a 50-person company

Start from roles, not headcount. Assume 50 employees. Thirty-six run tools from an assistant. One compliance reviewer only reads evidence, one person owns the invoice, and twelve never touch any of it.

In Elaichi, a billable seat is an active member whose role is not Guest, Billing Admin or Auditor. These three role names are Elaichi's own vocabulary, not an industry standard. The pattern is common across access-management products: free read-only or billing-only roles, paid roles for anyone who acts. The reviewer takes a free Auditor seat and the invoice owner a free Billing Admin seat. The twelve are not members at all. That leaves 36 billable seats.

Gold lists at $15 per user per month in USD, or $10 per user per month billed annually. At the monthly figure, 36 seats is $540 a month. Billed yearly it is $4,320 a year, or $360 a month equivalent. [The pricing page](/pricing/) shows the figure for your country, and the final price follows the country your card was issued in.

One trap belongs in the model. A custom role is always a billable seat, even one holding nothing but view permissions. Building a read-only analyst role to save money does not save money. [Auditor is the only free read-only role](/blog/auditor-seat-for-compliance-reviewers/). It opens billing details, the connector catalog, restrictions and API token metadata. It also opens SSO and SCIM settings, logging destinations and the audit log. It does not open other members' connections or toolboxes.

Connectors cost nothing for anything already in the catalog. Elaichi's catalog holds 650+ connectors of two kinds. Most are ones Elaichi authors and runs itself. The rest are native MCP connectors: the vendor builds and runs its own MCP server, and Elaichi handles sign-in, access and audit. Check [the connector catalog](/connectors/) against your app list before you budget a line here. For an app that is missing, an Org Owner or Org Admin has two routes. They can bring in its remote MCP server as a custom connector, or build one from config. Both require Gold.

Admin time at 50 people covers five recurring jobs:

- Connecting the first apps. Each person connects their own account once, with their own sign-in at the app.
- Writing restrictions. A rule targets a role or one member, never a team and never the whole organization.
- Deciding access requests, in the console under Governance, with a confirmation step tied to the approver's identity.
- Offboarding. A preflight returns the leaver's whole estate, connections, shared toolboxes and templates, before anything changes.
- Repairs. A connection whose token refresh failed shows Needs sign-in until its owner reconnects it.

There is no audited industry figure for how many hours that is at your company. Headcount, app count and the number of custom rules you write all move it. The first week is the heaviest one, because that is when apps get connected and the first roles and restrictions get written. 14-day Black trial, no credit card to start. Buying Gold or Black during the trial keeps its remaining days rather than granting a fresh 14, so it is one continuous trial, and the seats are first charged when it ends. A workspace that buys nothing moves to the free Personal plan. Time the five jobs yourself during it, then multiply measured hours by A.

| Layer | 50 people |
|---|---|
| Assistant seats | 36 x S |
| Access layer, Gold monthly | $540 |
| Access layer, Gold billed yearly (monthly equivalent) | $360 |
| Connectors in the catalog | $0 |
| Admin time | measured hours x A |

Across all 50 heads, the access layer at the monthly list price is $10.80 per employee per month.

## Worked cost model: a 200-person company

Same arithmetic, different shape. Assume 150 billable seats out of 200 people. Gold is $2,250 a month at the US-dollar list price, or $1,500 a month equivalent billed yearly. Across 200 heads that is $11.25 per employee per month, close to the 50-person figure. The access layer is priced per seat, not per app or per policy.

Seats scale with headcount. Admin does not, for most of the work. A restriction binds a role or one member, so one block rule on the Member role governs 150 people as cheaply as it governs 36. The number of rules follows the role count, not the headcount. Write allow rules carefully. An allow rule switches on an allowlist for its whole layer, so [an allow rule naming nothing blocks everything](/blog/empty-allow-rule-denies-everything/) for its target.

Two costs do scale with headcount at 200. Access requests arrive per person. How many arrive depends on how tightly the initial access mapping was drawn, so count them during the trial. Offboarding arrives per leaver. A leaver with shared connections, toolboxes, templates or custom connectors needs decisions before removal completes. Budget that per departure, not per head, and read [what offboarding touches](/blog/offboarding-when-the-agent-holds-access/) before the first one.

Gold includes SSO with SAML or OIDC, SCIM provisioning and [group-to-role mappings](/blog/mapping-sso-groups-to-roles/). That is what keeps the admin line close to flat at this size. A joiner lands on the role their group maps to, and a mapping confers at most one role. A SCIM deprovision suspends the member, which revokes their MCP grants and API tokens. In Elaichi, removing or suspending a member revokes every live grant in the same transaction as the membership change. Role, team and restriction changes reach every surface about two minutes.

Audit evidence is the other line that quietly grows with headcount and time. Elaichi writes one entry for each connected-tool call that reaches execution (a call refused earlier writes none). That record is the artifact a SOC 2 or customer security review asks for. Generating it costs nothing beyond the Gold subscription already paid. Elaichi's certifications, SOC 2 included, are published in its [Trust Center](https://trust.elaichi.ai). Elaichi retains 90 days of audit history by default. If your auditor wants seven years, export to your own Datadog comes with the Black plan. The rest of that decision is in [audit log retention for AI actions](/blog/audit-log-retention-ai-actions/).

Two billing mechanics matter before the finance review. A seat added mid-term is charged its prorated price within minutes. A bulk SCIM add, such as onboarding a whole department at once, is collected as one charge rather than many line items.

If the 200-person company also wants workflows, shared tables, dashboards and the Elaichi Agent, that requires Black. Black lists at $24 per seat per month in US dollars, or $18 billed annually. At the monthly list price, 150 seats is $3,600 a month. Black also bills usage past a monthly allowance of 1,000 run minutes and 1 GB-month, pooled across the organization. The rates are $0.02 a run minute and $5 a GB-month. Only active working time counts; time spent waiting on a person or a timer is not billed. Gold and Personal carry no usage-based billing at all. That makes the Gold line the one with zero variance month to month. It matters when the line item has to survive a finance review unchanged.

## Evaluating the access layer: DIY gateway or managed control plane

Building the access layer yourself is the right call in two specific situations and an expensive ongoing habit in the rest. Here is what each path actually requires:

| | DIY build | Managed control plane | Hybrid |
|---|---|---|---|
| Upfront work | OAuth authorization server with dynamic client registration and PKCE (S256), consent/grant storage, credential storage with refresh, pre-call restriction checks, per-call audit logging | Subscription setup, SSO/SCIM mapping | Managed platform + one custom MCP server for a legacy system |
| Recurring work | Token refresh breakage, API changes breaking connectors, a new connector per department request | Vendor maintains connectors and auth flow | Vendor covers catalog connectors; you maintain only the bespoke one |
| Who it fits | Air-gapped networks; bespoke internal systems no vendor will ever publish a connector for | Most companies with ten or more standard SaaS apps in common use | Companies with one or two legacy systems plus a standard SaaS stack |
| Cost shape | Engineer-days upfront, then engineer-days every year for maintenance | Per-seat subscription, usage-based add-ons for workflow platforms | Subscription + amortized engineering time for the one custom piece |

Price a DIY build honestly against what a control plane starts with. You need an OAuth authorization server with dynamic client registration and PKCE using S256, which is what MCP clients expect. You need per-person consent and grant storage, plus credential storage with refresh. You need a restriction layer checked before each call leaves. You need an audit record per call naming the person, the tool and the account. Then add one connector per app, with maintenance each time that app changes its API.

Count it in engineer-days at your own loaded day rate. Then count the recurring half again every year. Token refresh breaks on its own schedule. A vendor changes an endpoint, and the connector stops until someone patches it. Each new app a department requests becomes a ticket: read the docs, write the server, configure scopes, deploy, update routing. While that ticket sits open, people tend to paste data into a chat window instead of waiting.

Two situations keep building the right call. One is an air-gapped network. The other is a bespoke internal system no vendor will ever publish a connector for. If keeping the gateway inside your own network is a hard requirement, other vendors are built specifically for that constraint. [Lunar.dev](https://www.lunar.dev/) presents itself as an enterprise MCP gateway and AI control plane. Its FAQ says MCPX is fully self-hosted inside the customer's infrastructure (checked November 2026). [Kong's docs](https://developer.konghq.com/ai-gateway/) describe AI Gateway as a connectivity and governance layer that puts LLM, MCP and agent-to-agent traffic under one control plane. On premises it runs on self-hosted Kong Gateway with AI plugins (checked October 2026). [Tyk's MCP Gateway docs](https://tyk.io/docs/ai-management/mcp-gateway/overview) say Tyk Gateway sits in front of the customer's own MCP servers and Tyk-managed REST APIs. It proxies that MCP traffic through a centrally managed gateway layer (checked October 2026). Elaichi is hosted, and no current plan offers a self-hosted deployment. The longer version of this trade-off is in [self-hosted MCP servers versus a managed platform](/blog/self-hosted-mcp-servers-vs-control-plane/).

The hybrid approach is usually cheaper than either pure option. A company that already runs an MCP server for one legacy system can add it as a custom remote MCP connector on Gold. Restrictions, OAuth scopes and the audit log then cover that connector the same way they cover any catalog connector. You avoid paying for governance twice, once in engineering time and once in subscription cost.

One cost appears in neither column, and it is usually the largest: not choosing a path at all. People connect personal API keys and run MCP servers on their own laptops. The bill arrives later, as expense-report line items and an access review question you cannot answer. That is the real comparison case, and [replacing personal MCP servers](/blog/replace-personal-mcp-servers/) sets out what it costs to undo once it has spread.

## When this budget line should be zero

Say three people use an assistant, each connecting their own accounts through the vendor's own connectors. There is nothing in this framework to buy yet. Elaichi's Personal plan is free with no end date, but it is scoped to one person. The limits are 1 member, 10 connections across at most 3 connectors, 10 synthetic tools, and 1,000 connected-tool calls a month. Sharing access with a colleague is what starts the Gold requirement.

The line stops being zero when you cannot answer one question. The question is which person, in which app, had an assistant change something, and under whose account. [When you do not need an MCP gateway yet](/blog/when-you-dont-need-an-mcp-gateway/) sets out that threshold in more detail.

## How to build the number for your own company

Work the four layers in order: seats, access layer, connectors, admin time. Keep every figure traceable to a pricing page or a stopwatch reading, not an industry average. The spread between a light user's admin burden and a heavy one's is wide. An average tells you little about your own spreadsheet. Measure your own five jobs during a trial before you commit a number to a budget line.

The per-seat comparison across vendors sits in [what AI access costs per user](/blog/mcp-gateway-pricing/). The governance work the admin line pays for is described under [governance use cases](/use-cases/governance/). The reason budgets get rebuilt six months in is covered in [why AI pilots fail](/blog/why-ai-pilots-fail/).

## FAQ

### How much does a governed MCP control plane cost per user?

Elaichi Gold lists at $15 per user per month in USD, or $10 per user per month billed annually. Black lists at $24 per seat per month, or $18 per seat per month billed annually, plus usage beyond a monthly allowance of 1,000 run minutes and 1 GB of storage, at $0.02 a run minute and $5 a GB-month. Guest, Billing Admin and Auditor seats are free and are not billed. Prices are lower in many countries, and [the pricing page](/pricing/) shows the figure for the visitor's own country.

### Do AI assistants count as billable users?

No. In Elaichi you pay per person who signs in, except Guests, Billing Admins and Auditors, and AI apps do not count as users. Each person connects their own AI client to the same MCP endpoint and approves their own grant, so adding Claude, ChatGPT and Cursor for one employee is still one seat.

### Is it cheaper to build our own MCP servers than to buy a control plane?

Rarely, unless the network is air-gapped or the system is a bespoke internal one nobody will publish a connector for. A do-it-yourself access layer means building an OAuth authorization server with dynamic client registration and PKCE, per-person consent and grant storage, credential storage with refresh, a restriction layer checked before each call, an audit record per call, and one connector per app with ongoing maintenance. Count that in engineer-days at your own loaded day rate, and count it again each year.

### What does the admin time line in an AI access budget actually cover?

Five recurring jobs: connecting apps and helping people sign in to their own accounts, writing restrictions for each role, deciding access requests in the console, running offboarding for each leaver, and reconnecting connections whose token refresh failed. Access requests and offboarding scale with people; restrictions scale with the number of roles, not with headcount, because a rule targets a role or one member.

### How can we get a real number before committing to a plan?

14-day Black trial, no credit card to start. Buying Gold or Black during the trial keeps its remaining days rather than granting a fresh 14, so it is one continuous trial, and the seats are first charged when it ends. A workspace that buys nothing moves to the free Personal plan. Connect the apps your teams actually name, write the restrictions, and time the admin work during those two weeks. When the trial ends, the workspace moves to the free Personal plan and nothing is deleted.

## Read next

- [Elaichi alternatives: an honest list by situation](/blog/elaichi-alternatives/) — An honest list of Elaichi alternatives by situation: native connectors, Zapier MCP, Composio, Merge Agent Handler, Mintive, Glean and self-hosted servers.
- [MCP server vs MCP gateway: the request path](/blog/mcp-server-vs-mcp-gateway/) — MCP server vs MCP gateway: a server runs one system's tools; a gateway sits in front of many, adding sign-in, per-person access and one audit log.
- [Why AI pilots fail, and how to get to daily use](/blog/why-ai-pilots-fail/) — Why AI pilots fail: the AI cannot reach your apps, nobody trusts it, or nobody owns it. The fix for each cause, and how to measure daily use.