NUVAI · One working layer
From scattered AI tools to one working layer
Renout van Hove for NUVAI
A practical paper on AI in a small or mid-sized business: from scattered tools to one workspace that works directly with your CRM.
By NUVAI. Figures checked on 24 October 2026.
Download the PDF PDF, 27 pages
00
0. Summary
Most companies with ten to two hundred staff already use AI. They just don’t use it in one place. There are a handful of separate subscriptions, some free accounts, and a few people who have each worked out their own way of doing things. Nobody knows exactly where customer data ends up.
This paper is about moving from those scattered tools to one working layer: a single AI workspace per employee, on Claude or Grok, connected through MCP straight to your CRM and your other systems. With fixed permissions, with a person who approves anything the AI wants to change, and with a cost you can work out in advance.
What we found
- In 2025, 20.0% of EU businesses with ten or more employees used at least one AI technology. Belgium stood at 34.5%, the Netherlands at 33.2% and France at 18.2% (Eurostat). The UK is not covered by this dataset.
- 78% of people who use AI at work bring their own tools. At small and medium-sized businesses the figure is 80% (Microsoft Work Trend Index 2024).
- A business AI seat costs $20 a month on annual billing with Claude Team and ChatGPT Business. Grok Business costs $30. Microsoft 365 Copilot Business costs $21 on top of a Microsoft 365 licence.
- The main CRMs now have an official MCP server: HubSpot, Salesforce, Pipedrive, Zoho CRM and Microsoft Dataverse. Teamleader has one for Orbit.
- If you operate in the EU, you have had to support your staff’s AI literacy since 2 February 2025. Supervision and enforcement rules have applied since 2 August 2026.
What you can do this month
- List every AI tool in use across the business, free ones included.
- Choose one workspace and one plan, and cancel the rest.
- Connect your CRM with read-only access first.
- Pick two or three processes from the catalogue in chapter 6.
- Write a one-page AI policy and keep a record of who has had which training.
- of Belgian businesses with ten or more staff used AI in 2025
- 34.5%
- of Belgian businesses with ten or more staff used AI in 2025
- Source S1
- a business AI seat per month, annual billing
- $20
- a business AI seat per month, annual billing
- Source S3, S4
- systems with an official MCP server in this paper
- 6
- systems with an official MCP server in this paper
- Source S11, S12, S13, S14, S15, S16, S17, S18
- AI Act Article 4: supervision and enforcement apply
- 2 August 2026
- AI Act Article 4: supervision and enforcement apply
- Source S19
View as table
| Source | |
|---|---|
| 34.5% of Belgian businesses with ten or more staff used AI in 2025 | S1 |
| 78% of AI users bring their own tools to work | S2 |
| $20 a business AI seat per month, annual billing | S3,S4 |
| 6 systems with an official MCP server in this paper | S11,S12,S13,S14,S15,S16,S17,S18 |
| 2 August 2026 AI Act Article 4: supervision and enforcement apply | S19 |
One working layer, in five levels
01
1. AI is already running in your business
1.1
What the figures say
Every year Eurostat measures how many businesses with at least ten employees use an AI technology. Between 2024 and 2025 the EU figure rose from 13.5% to 20.0%. Belgium is among the leaders, up from 24.7% to 34.5%. The Netherlands went from 23.1% to 33.2% and France from 9.9% to 18.2%. Eurostat flags a break in the Dutch time series in 2025, so compare those two years with some care.
Company size matters a great deal. Across the EU in 2025, 17.0% of businesses with 10 to 49 employees used AI, against 30.4% of those with 50 to 249 and 55.0% of large firms. In Belgium the figures were 28.8%, 54.5% and 76.4%.
These figures count businesses that use AI officially. They say little about what staff do on their own initiative.
Businesses using at least one AI technology
View as table
| 2024 | 2025 | |
|---|---|---|
| Belgium | 24.7% | 34.5% |
| Netherlands | 23.1% | 33.2% (b) |
| France | 9.9% | 18.2% |
| EU27 | 13.5% | 20.0% |
| 10 to 49 staff 2024 | 10 to 49 staff 2025 | 50 to 249 staff 2024 | 50 to 249 staff 2025 | 250 staff or more 2024 | 250 staff or more 2025 | |
|---|---|---|---|---|---|---|
| Belgium | 20.7% | 28.8% | 35.7% | 54.5% | 66.3% | 76.4% |
| Netherlands | 20.0% | 28.8% | 31.2% | 46.6% | 54.1% | 67.6% |
| France | 8.5% | 15.0% | 14.5% | 30.8% | 32.7% | 58.0% |
| EU27 | 11.2% | 17.0% | 21.0% | 30.4% | 41.2% | 55.0% |
1.2
What the figures don’t show
For its 2024 Work Trend Index, Microsoft asked 31,000 people in 31 countries how they use AI. 78% of AI users brought their own tools to work. At small and medium-sized businesses it was 80%. The research dates from May 2024, but little of what we see in practice suggests the share has fallen since.
That is understandable. People want to work faster and reach for the tools they know. The problem is not the use itself but three of its consequences:
- Data. Quotes, client names and figures go into accounts the business cannot see. With a free or personal account you often don’t know whether that data is used to train models.
- Know-how. Each person builds their own prompts and routines. When they leave, that know-how leaves with them.
- Money. Separate subscriptions are paid through expense claims. Nobody sees the total.
Where a provider processes your data matters too. A tool running outside the EU or UK without clear terms can be a legal risk for your business, however well it works.
Bring your own AI
78%
of AI users at work
80%
at small and medium-sized businesses
View as table
| % | |
|---|---|
| of AI users at work | 78% |
| at small and medium-sized businesses | 80% |
1.3
Self-test: how big is the shadow?
Answer these ten questions with yes or no.
Self-test: how big is the shadow?
0 of 10 answered
- 01Do staff use AI tools the business did not choose?
- 02Are AI subscriptions paid through expense claims or personal cards?
- 03Does anyone use a free account for work involving client data?
- 04Are you unsure which tools record or summarise meetings?
- 05Do people paste CRM data or quotes into a chat window?
- 06Is there no written AI policy?
- 07Has nobody kept a record of who has had which AI training?
- 08Do two teams use different AI tools for the same kind of work?
- 09Would you lose someone’s prompts and routines if they left?
- 10Is nobody able to say how much the business spends on AI each month?
Zero to two yeses: you have little shadow AI. Three to five: there is work to do, but it is manageable. Six or more: start with the inventory in chapter 8 before you buy anything new. These thresholds are our own judgement, not a statistic.
02
2. What “one working layer” means
One working layer is not a new system. It is one AI workspace that sits between your people and the software you already have. Your CRM stays your CRM. Your accounts package stays your accounts package. The layer lets the AI read that data, and change it under agreed conditions, without anyone copying and pasting.
We describe the layer in five levels, from the bottom up.
The five levels of one working layer
- What it is
- Who may do what, who approves a change, and who owns which process.
- Examples
- Rights, approval steps, process owners
- Who owns it
- Management and the process owners
Proposed changes, which wait until a person approves them.- What it is
- Projects, instructions and templates the whole team uses, so a quote always looks the same.
- Examples
- Projects, instructions, templates
- Who owns it
- The process owner
Shared instructions and templates the assistant applies to the answer.- What it is
- Claude Team or Enterprise, or Grok. One account per person, managed by the business, with no training on your data.
- Examples
- Claude Team or Enterprise, Grok Business
- Who owns it
- The business, one account per person
Requests and results, signed in with your own OAuth login. No password passes.- What it is
- Each link is an official MCP server from the vendor, or a connector offered by the workspace itself. Each has its own permissions.
- Examples
- Official vendor MCP servers, workspace connectors
- Who owns it
- An administrator, per connection
Records you may already see in the system, read through the vendor MCP server.- What it is
- CRM, email, calendar, documents, accounting, ticketing. This is where the truth lives, and it stays there.
- Examples
- CRM, email, calendar, documents, accounting, ticketing
- Who owns it
- The owner of each system, as today
The difference from scattered tools lies mostly in the bottom two levels and the top one. With scattered tools, each person copies data out of the CRM by hand. With one layer, the AI reads straight from the CRM, with the permissions of whoever asked, and changes nothing until someone clicks “approve”.
Scattered tools versus one layer
View as table
| Before | Four tools, each with its own access to mail, CRM and documents: twelve separate connections. |
|---|---|
| After | One layer with one set of connections: three connections, each with its own permissions. |
2.1
Claude or Grok?
Both work. Claude Team and Enterprise have a broad connector system, shared projects and admin controls for connectors. Grok Business offers connectors (Salesforce and HubSpot among them), the option to bring your own MCP server, role-based access and no training on your data. Choose on the basis of what your team already uses, which connectors you need and what your administrator finds easiest to run. Chapter 5 compares prices, including ChatGPT Business and Microsoft 365 Copilot.
03
3. MCP in plain English
MCP stands for Model Context Protocol. It is an open standard for how an AI assistant talks to other software. Think of it as a plug socket: if your CRM has an MCP server and your AI workspace supports MCP, they fit together.
3.1
What happens when you ask a question
What happens when you ask a question
Step 1 of 6
- 01You ask something like: “Which deals over £10,000 haven’t been updated for two weeks?”
- 02The assistant sees there is a CRM connection and picks the right tool, for example “search deals”.
- 03The connection uses your own login through OAuth, so the AI only sees what you are allowed to see in the CRM.
- 04The CRM’s MCP server runs the search and returns the result.
- 05The assistant answers and points to the records it used.
- 06If it wants to change something, such as a stage or a note, it asks first. You approve or decline.
Two points matter here. The AI never gets more rights than the user. And the MCP specification forbids a server from passing on a token meant for another service (“token passthrough”). That limits the damage if one connection is compromised.
3.2
Which systems have an official MCP server?
| System | Official server | How you connect | Who switches it on | Source |
|---|---|---|---|---|
| HubSpot | Yes, remote (mcp.hubspot.com) | OAuth with PKCE, connector in Claude | A Super Admin approves | Source S11, S12 |
| Salesforce | Yes, Headless 360 MCP server | Your own External Client App, OAuth | Salesforce admin | Source S13 |
| Pipedrive | Yes, maintained by Pipedrive | OAuth, on every plan | Each user, within their own rights | Source S14 |
| Zoho CRM | Yes, four separate servers (read, records, structure, automation) | OAuth, through the Zoho Server Registry | Admin chooses which servers | Source S15 |
| Microsoft Dataverse / Dynamics 365 | Yes, per environment | Remote endpoint or local proxy, Entra app | Admin enables per environment | Source S16, S17 |
| Teamleader Orbit | Yes (api.orbit.teamleader.eu/mcp) | Each user logs in with their own account | Admin grants “MCP read” and “MCP write” | Source S18 |
| Teamleader Focus | Not found, community servers only | To be checked | To be checked | To be checked |
Dynamics 365 Sales has its own MCP server, and Microsoft notes that Claude Desktop is not supported there. If you work on Dataverse, use the Dataverse server.
If your helpdesk or accounts package isn’t listed, check your own tool’s documentation for an official MCP server. A community server may work, but then maintenance and security rest with someone you don’t know.
04
4. Example: Claude and HubSpot in an afternoon
HubSpot makes a good example because its link with Claude is well documented. The steps below take about two and a half hours in total. The timings are our own estimate.
Claude and HubSpot, step by step
0 of 145 minutes done
Some limits of the HubSpot connector worth knowing: bulk changes are capped at ten records at a time, HubSpot’s normal API limits still apply, and pipeline stage validation rules are enforced. Association label validation is not applied automatically.
4.1
Which rights do you grant?
Which rights do you grant?
| Grant | Grant with approval | Never grant |
|---|---|---|
4.2
Where it goes wrong, and how to prevent it
Where it goes wrong
01Broad rights from day one
Start with read access. Add write access only once a process has run well for a few weeks.
02Shared accounts
Have each person connect with their own login. Otherwise you can’t tell who did what.
03Writing without approval
Keep write tools on “Needs approval”, even when it gets tedious.
04No owner
Every connection has one person who decides and follows up.
05Forgotten limits
Ten records per bulk change is a safety limit. Plan large clean-ups separately.
06Messy data
The AI won’t tidy your CRM. A duplicate contact stays a duplicate. Clean the key fields first.
05
5. What it costs
All prices below are the official USD prices per user per month, excluding VAT, as listed on vendors’ websites on 24 October 2026. Euro amounts are indicative, converted at the ECB reference rate of 23 October 2026 (1 euro = 1.1193 dollars). Prices may vary by region, and UK customers may be billed in sterling. Always check your own billing page.
5.1
Price per user
| Plan | Annual billing | Monthly billing | Note |
|---|---|---|---|
| Claude Team, Standard | $20 (≈ €17.87) | $25 (≈ €22.34) | 2 to 150 seats |
| Claude Team, Premium | $100 (≈ €89.34) | $125 (≈ €111.68) | higher usage, Claude Code |
| Claude Enterprise | $20 (≈ €17.87) + usage at API rates | annual contract, SCIM, audit logs | |
| Grok Business | $30 (≈ €26.80) | no training, bring your own MCP server | |
| ChatGPT Business, Standard | $20 (≈ €17.87) | $25 (≈ €22.34) | minimum 2 seats |
| ChatGPT Business, Premium | $100 (≈ €89.34) | $125 (≈ €111.68) | |
| Microsoft 365 Copilot Business | $21 (≈ €18.76) | no monthly option | on top of a Microsoft 365 licence, up to 300 users |
Until 31 December 2026 Microsoft offers a first-year promotion at $18, for eligible existing customers only. We leave it out of the calculations.
per user per month
View as table
| Annual USD | Monthly USD | Annual EUR | Monthly EUR | Source | |
|---|---|---|---|---|---|
| Claude Team · Standard | $20 | $25 | €17.87 | €22.34 | S3,S4 |
| Claude Team · Premium | $100 | $125 | €89.34 | €111.68 | S3,S4 |
| Claude Enterprise | $20 | annual only | €17.87 | annual only | S3 |
| Grok Business | $30 | annual only | €26.80 | annual only | S6 |
| ChatGPT Business · Standard | $20 | $25 | €17.87 | €22.34 | S8 |
| ChatGPT Business · Premium | $100 | $125 | €89.34 | €111.68 | S8 |
| Microsoft 365 Copilot Business | $21 | annual only | €18.76 | annual only | S9 |
5.2
API credit with Claude Team
Claude Team includes a pooled API credit. Anthropic gives its own example: three Standard and two Premium seats give $260 a month. That works out at $20 per Standard seat and $100 per Premium seat, capped at $500 a month. You can use the credit for processes that run through the API.
5.3
What a process costs in tokens
An automated process is paid for by the token: what the model reads (input) and what it writes (output). Take a process that runs 200 times a month, with 20,000 input tokens and 2,000 output tokens each time.
| Model | Price per million tokens (in / out) | Cost per month |
|---|---|---|
| Claude Haiku 5.5 | $0.50 / $2.50 | $3.00 |
| Claude Sonnet 5.5 | $2 / $10 | $12.00 |
| Claude Opus 5.5 | $4 / $20 | $24.00 |
| Grok 4.3 | $1.25 / $2.50 | $6.00 |
| Grok 4.7 | $2 / $6 | $10.40 |
The formula: runs × (input tokens × input price + output tokens × output price) / 1,000,000.
View as table
| in / out (USD / 1M) | per month | Source | |
|---|---|---|---|
| Claude Haiku 5.5 | $0.50 / $2.50 | $3.00 | S3 |
| Claude Sonnet 5.5 | $2.00 / $10.00 | $12.00 | S3 |
| Claude Opus 5.5 | $4.00 / $20.00 | $24.00 | S3 |
| Grok 4.3 | $1.25 / $2.50 | $6.00 | S7 |
| Grok 4.7 | $2.00 / $6.00 | $10.40 | S7 |
5.4
Two worked examples
These are examples, not quotes. Put your own numbers into the calculator.
Example A: 25 staff. Claude Team on annual billing, 23 Standard and 2 Premium seats, three processes on Sonnet 5.5.
- Seats: 23 × $20 + 2 × $100 = $660 a month, or $7,920 a year (≈ €7,076).
- Tokens: 3 × $12 = $36 a month. That falls within the $500 API credit, so net $0.
- Kickstart: €4,500 excl. VAT, one-off.
- Year 1: ≈ €11,576. Year 2: ≈ €7,076.
Example B: 100 staff. Claude Team on annual billing, 92 Standard and 8 Premium seats, five processes on Sonnet 5.5.
- Seats: 92 × $20 + 8 × $100 = $2,640 a month, or $31,680 a year (≈ €28,303).
- Tokens: 5 × $12 = $60 a month, within the credit, so net $0.
- Kickstart: €4,500 excl. VAT.
- Year 1: ≈ €32,803. Year 2: ≈ €28,303.
The Kickstart is the first step. Further build after it is scoped separately and is not included in these examples.
For comparison, seats only for 25 users on annual billing: Claude Team $6,000, ChatGPT Business $6,000, Microsoft 365 Copilot Business $6,300 (plus the Microsoft 365 licence), Grok Business $9,000.
Results
- Seats per month
- $660
- Seats per year
- $7,920 (≈ €7,076)
- Token cost per month
- $36.00 (≈ €32.16)
- Claude Team API credit per month
- $500
- Net token cost per month
- $0.00 (≈ €0.00)
- Year 1 (with Kickstart)
- ≈ €11,576
- Year 2
- ≈ €7,076
Source S20
- Standard seats≈ €4,932
- Premium seats≈ €2,144
- Tokens≈ €386
- Minus API credit≈ €-386
- Kickstart, one-off≈ €4,500
- Year 1≈ €11,576
View as table
| USD | EUR | |
|---|---|---|
| Standard seats | $5,520 | ≈ €4,932 |
| Premium seats | $2,400 | ≈ €2,144 |
| Tokens | $432 | ≈ €386 |
| Minus API credit | $-432 | ≈ €-386 |
| Kickstart, one-off | ≈ €4,500 | |
| Year 1 | ≈ €11,576 |
5.5
What do you set against it?
What you spend today on separate subscriptions, consultants, agencies or extra staff varies widely. That is why the calculator leaves those amounts blank. Add up your current AI subscriptions, estimate how many hours a week the chosen processes take now, and compare that with the figures above.
06
6. Ten processes to start with
We score each process on three things, each from 1 (low) to 3 (high): how much work it takes to set up, how much risk there is if it goes wrong, and how much time it gives back. The scores are our own judgement from our work, and they depend on the state of your data.
| # | Process | Effort | Risk | Impact | Rights | Who approves |
|---|---|---|---|---|---|---|
| 1 | Client meeting preparation | 1 | 1 | 2 | read CRM, read calendar | nobody, read only |
| 2 | CRM update after a call | 2 | 2 | 3 | write notes and fields | account owner |
| 3 | Weekly pipeline check | 1 | 1 | 2 | read CRM | nobody, read only |
| 4 | Quote in house style | 2 | 2 | 3 | read CRM, write documents | salesperson before sending |
| 5 | Lead qualification | 2 | 2 | 3 | read CRM, write fields | sales manager |
| 6 | Content repurposing | 1 | 1 | 2 | read documents | marketing |
| 7 | Commentary on monthly figures | 2 | 2 | 2 | read accounts | finance lead |
| 8 | Invoices and missing documents | 3 | 3 | 2 | read and write accounts, read email | bookkeeper |
| 9 | First read of contracts and NDAs | 2 | 3 | 2 | read documents | lawyer or managing director |
| 10 | Support ticket preparation | 2 | 2 | 3 | read tickets, write drafts | support lead |
Quick wins. Processes 1, 3 and 6 take little work, carry little risk and save time straight away. They only need read access. This is usually where we start.
Biggest gains. Processes 2, 4, 5 and 10 return the most, but they need write access and a set approval step.
Handle with care. Processes 8 and 9 touch money and legal terms. A person always reads along.
Ten processes to start with
View as table
| # | Department | Effort | Risk | Impact | Rights | Who approves | |
|---|---|---|---|---|---|---|---|
| 1 | Client meeting preparation (Quick win) | Sales | 1 | 1 | 2 | read CRM and calendar | nobody, read only |
| 2 | CRM update after a call | Sales | 2 | 2 | 3 | write notes and fields | account owner |
| 3 | Weekly pipeline check (Quick win) | Sales | 1 | 1 | 2 | read CRM | nobody, read only |
| 4 | Quote in house style | Sales | 2 | 2 | 3 | read CRM, write documents | salesperson before sending |
| 5 | Lead qualification | Marketing and sales | 2 | 2 | 3 | read CRM, write fields | sales manager |
| 6 | Content repurposing (Quick win) | Marketing | 1 | 1 | 2 | read documents | marketing |
| 7 | Commentary on monthly figures | Finance | 2 | 2 | 2 | read accounts | finance lead |
| 8 | Invoices and missing documents | Finance | 3 | 3 | 2 | read/write accounts, read email | bookkeeper |
| 9 | First read of contracts and NDAs | Legal and operations | 2 | 3 | 2 | read documents | lawyer or managing director |
| 10 | Support ticket preparation | Customer service | 2 | 2 | 3 | read tickets, write drafts | support lead |
Effort versus impact
x = effort · y = impact · colour = risk
View as table
| # | Effort | Impact | Risk | |
|---|---|---|---|---|
| 1 | Client meeting preparation | 1 | 2 | 1 |
| 2 | CRM update after a call | 2 | 3 | 2 |
| 3 | Weekly pipeline check | 1 | 2 | 1 |
| 4 | Quote in house style | 2 | 3 | 2 |
| 5 | Lead qualification | 2 | 3 | 2 |
| 6 | Content repurposing | 1 | 2 | 1 |
| 7 | Commentary on monthly figures | 2 | 2 | 2 |
| 8 | Invoices and missing documents | 3 | 2 | 3 |
| 9 | First read of contracts and NDAs | 2 | 2 | 3 |
| 10 | Support ticket preparation | 2 | 3 | 2 |
07
7. From first connection to a team that uses it
A typical project looks like this.
From first connection to a team that uses it
Day 0
We agree which processes to build, who owns them and which rights they need. You make sure users have access to the workspace and the CRM.
Within 72 hours
The first process is live, with read access only. The 72 hours start once users have access.
Week 1
Two or three processes are running. We adjust based on what the team tells us. Write access is added where needed, always with approval.
Week 2
All agreed processes are running. The team gets a short training session and we record who had which training. You receive a simple AI policy.
After that
You own it. You know how to adjust a process or add a new one.
7.1
What the EU AI Act asks
If you operate in the EU, Article 4 of the AI Act has applied since 2 February 2025. It asks businesses that use AI systems to take measures to support their staff’s AI literacy. Since the Digital Omnibus amendment (mid-July 2026), no specific level is mandated. National market surveillance authorities supervise and enforce it from 2 August 2026.
According to the European Commission you don’t need a certificate. An internal record of training and other initiatives is enough, and no dedicated AI officer or governance board is required. UK businesses that only operate in the UK are not covered, but the same record is good practice.
EU AI Act, Article 4
Applies
2 February 2025
Digital Omnibus
mid-July 2026: no specific level mandated
Supervision and enforcement
2 August 2026
View as table
| Applies | 2 February 2025 |
|---|---|
| Digital Omnibus | mid-July 2026: no specific level mandated |
| Supervision and enforcement | 2 August 2026 |
08
8. Checklist: twenty points
Checklist: twenty points
0 of 20 done
8.1
Template: a one-page AI policy
Template: a one-page AI policy
- Purpose
- Why we use AI, and what we don’t use it for.
- Approved tools
- The workspace we chose. Other tools only with approval.
- What never goes into an AI tool
- For example passwords, payment details, health data, and data a client has not consented to.
- Approval
- Who approves changes in the CRM and other systems.
- Meetings
- When recording and summarising is allowed, and that participants know beforehand.
- Training
- Who gets which training and where we record it.
- Owner and review
- Who maintains the policy and when we review it.
09
9. Do it yourself or get help?
Much of what this paper describes you can do yourself. Five questions help you decide. If you answer no to all of them, do it yourself. One yes or more usually means help is worth it.
Do it yourself or get help?
- 1Does the AI need to change anything in your CRM or accounts?
- 2Does the process span more than one system?
- 3Does it need to run without anyone starting it?
- 4Does it need to work within a month?
- 5Is there nobody in-house with a few hours a week to manage it?
9.1
How NUVAI works
Most build work runs on our own delivery agents. A senior reviews everything and signs it off before it reaches you. You pay a fixed price and know in advance what you will get.
9.2
The Kickstart
- Price. €4,500 excl. VAT, fixed.
- Duration. Two weeks.
- Scope. Three to five processes on Claude or Grok, connected to your CRM.
- First result. The first process goes live within 72 hours of users having access.
- You also get. A permissions overview, a one-page AI policy and a training record.
Book a Kickstart at nuvai.eu/kickstart.
Want this running on your own accounts in two weeks? The Kickstart is a fixed €4,500 excl. VAT for three to five processes. The first one runs within 72 hours of your users getting access.
See the KickstartAppendix A
Appendix A. How we calculate
- Prices. Official USD prices per user per month, excluding VAT, checked on 24 October 2026.
- Exchange rate. 1 euro = 1.1193 dollars, ECB reference rate of 23 October 2026. Euro amounts are rounded and indicative.
- Token cost per process. runs × (input tokens × input price + output tokens × output price) / 1,000,000.
- Claude Team API credit. min(500, 20 × Standard seats + 100 × Premium seats) dollars a month, derived from Anthropic’s example.
- Net token cost. max(0, token cost − API credit), only for processes that run through the API.
- Year 1. 12 × seats + 12 × net token cost + Kickstart. Year 2 is the same without the Kickstart.
- What we leave blank. Your current costs for consultants, agencies and hiring. You fill those in.
Appendix B
Appendix B. Glossary
- MCP (Model Context Protocol). An open standard for how an AI assistant talks to other software.
- MCP server. The part of your CRM or tool that receives the AI’s requests and carries them out.
- Connector. A ready-made link in the AI workspace, often built on MCP.
- OAuth. Signing in to a service without handing your password to the AI.
- Token. A piece of text, roughly three quarters of a word. AI models charge per token.
- Shadow AI. AI tools staff use without the business having chosen them.
- AI literacy. Enough knowledge to use AI sensibly, as meant in Article 4 of the AI Act.
Appendix C
Appendix C. Calculator
The calculator in the web version uses the formulas in Appendix A. You enter: number of staff, Standard and Premium seats, plan, billing period, number of processes, runs per month, tokens per run and model. You can also add your current costs. Example values are marked as examples.
Appendix D
Appendix D. Sources
See the list of sources at the end. Every source was consulted on 24 October 2026 unless stated otherwise.
- S1Eurostat, Artificial intelligence by size class of enterprise (isoc_eb_ai), indicator E_AI_TANY, % of enterprises. https://ec.europa.eu/eurostat/databrowser/view/isoc_eb_ai/default/tableDataset updated 15 June 2026. Values read from the Eurostat database API. NL 2025 flagged (b) break in series.accessed 24 October 2026
- S2Microsoft & LinkedIn, 2024 Work Trend Index Annual Report: AI at work is here. Now comes the hard part. https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part31,000 respondents, 31 countries, published 8 May 2024.accessed 24 October 2026
- S3Anthropic, Claude pricing (Team, Enterprise, API). https://claude.com/pricingaccessed 24 October 2026
- S4Anthropic, What is the Team plan?. https://support.claude.com/en/articles/9266767-what-is-the-team-planaccessed 24 October 2026
- S5Anthropic, How is my Team plan bill calculated? (pooled API credit example). https://support.claude.com/en/articles/9267289-how-is-my-team-plan-bill-calculatedaccessed 24 October 2026
- S6xAI, Grok Business. https://x.ai/grok/businessaccessed 24 October 2026
- S7xAI, Models and pricing (API). https://docs.x.ai/developers/pricingaccessed 24 October 2026
- S8OpenAI, ChatGPT Business: pricing and seats. https://help.openai.com/en/articles/8792828accessed 24 October 2026
- S9Microsoft, Microsoft 365 Copilot Business pricing. https://www.microsoft.com/en-us/copilot/pricing/businessaccessed 24 October 2026
- S10Model Context Protocol, Security best practices (token passthrough). https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/security_best_practicesaccessed 24 October 2026
- S11HubSpot, Set up and use the HubSpot connector for Claude. https://knowledge.hubspot.com/integrations/set-up-and-use-the-hubspot-connector-for-claudeaccessed 24 October 2026
- S12HubSpot, Integrate AI tools with the remote HubSpot MCP server. https://developers.hubspot.com/docs/apps/developer-platform/build-apps/integrate-with-the-remote-hubspot-mcp-serveraccessed 24 October 2026
- S13Salesforce, Set up Claude connection to your Salesforce org (Trailhead). https://trailhead.salesforce.com/content/learn/modules/salesforce-in-claude-for-admins/set-up-salesforce-in-claudeaccessed 24 October 2026
- S14Pipedrive, Pipedrive MCP server. https://www.pipedrive.com/en/features/mcp-serveraccessed 24 October 2026
- S15Zoho, MCP for Zoho CRM. https://www.zoho.com/crm/developer/mcp.htmlaccessed 24 October 2026
- S16Microsoft Learn, Connect to Dataverse with Model Context Protocol. https://learn.microsoft.com/en-us/power-apps/maker/data-platform/data-platform-mcpaccessed 24 October 2026
- S17Microsoft Learn, Connect agents to Dynamics 365 Sales MCP server. https://learn.microsoft.com/en-us/dynamics365/sales/connect-agents-to-model-context-protocolaccessed 24 October 2026
- S18Teamleader, Teamleader Orbit MCP. https://www.teamleader.eu/blog/teamleader-orbit-mcpPage modified 23 October 2026.accessed 24 October 2026
- S19European Commission, AI Literacy: Questions & Answers. https://digital-strategy.ec.europa.eu/en/faqs/ai-literacy-questions-answersLast update 20 October 2026.accessed 24 October 2026
- S20European Central Bank, Euro foreign exchange reference rates, USD (EXR.D.USD.EUR.SP00.A). https://data.ecb.europa.eu/data/datasets/EXR/EXR.D.USD.EUR.SP00.A1 EUR = 1.1193 USD on 23 October 2026.accessed 24 October 2026
- S21NUVAI, Kickstart. https://nuvai.eu/kickstartNL: nuvai.eu/nl/kickstart, FR: nuvai.eu/fr/kickstart.accessed 24 October 2026
Want this running on your own accounts in two weeks? The Kickstart is a fixed €4,500 excl. VAT for three to five processes. The first one runs within 72 hours of your users getting access.
See the KickstartDownload the PDF PDF, 27 pages
See also: How we work · Kickstart
Claude and Anthropic are trademarks of Anthropic, PBC. Grok is a trademark of SpaceXAI LLC. ChatGPT is a trademark of OpenAI. Microsoft 365 and Copilot are trademarks of Microsoft. HubSpot, Salesforce, Pipedrive, Zoho and Teamleader are trademarks of their respective owners. NUVAI is independent of these vendors.
NUVAI is a brand of Quest2S BV · BE 0544.737.251 · nuvai.eu