Short answer: buy the general assistant, build only the tools that carry your own workflow, data or rules, and decide each tool with the same seven questions. Buy when the job is generic (drafting, summarising, meeting notes, searching files), many people need it, and the vendor's data terms and permission model pass your review. Build when the job is specific to how your company works, when the data cannot go to a vendor, or when a wrong answer costs real money. Most companies end up with both. The cost gap is usually people and upkeep, not model tokens.
This is engineering guidance, not procurement or legal advice. Prices below are US list prices, ex-tax, checked on 20 November 2026. They can change, and local pricing varies.
What are you actually choosing between?
"Build or buy" hides four choices. Each leaves you owning a different amount of work.
| Option | Examples | What you still own |
|---|---|---|
| 1. Buy a general assistant seat | Examples: Microsoft 365 Copilot, ChatGPT Business or Enterprise, Claude Team or Enterprise | What you still own: Rollout, permissions, training, usage review |
| 2. Buy a specialist SaaS tool | Examples: Invoice capture, contract review, a support desk with AI built in | What you still own: Integration, vendor review, exit plan |
| 3. Build on a model or cloud API | Examples: Your code calling a model provider or a document-extraction service | What you still own: Evaluation, review queue, monitoring, upkeep, security review |
| 4. Build and self-host the model | Examples: An open-weight model on your own servers or cloud account | What you still own: Everything in option 3, plus GPUs, model updates and uptime |
The choice is per use case, not per company. A firm can buy option 1 for everyone and still build option 3 for its one unusual process. List your use cases and rank them first. Our use-case prioritisation matrix shows one way.
What seven questions decide build or buy?
| Question | Leans buy | Leans build |
|---|---|---|
| 1. Is the task generic? | Leans buy: Drafting, summarising, meeting notes, document search | Leans build: A step only your company does, in your order, with your rules |
| 2. Does your own data or logic decide it? | Leans buy: The answer comes from general knowledge | Leans build: The answer depends on your price lists, approval limits or customer records |
| 3. Where can the data go? | Leans buy: A vendor's terms, retention and region settings pass your review | Leans build: The data must stay on your servers, or no vendor option passes review |
| 4. How many people use it, and how often? | Leans buy: Many people, light use each | Leans build: Few people, heavy use, or a step inside a pipeline with no human user |
| 5. How bad is a wrong answer? | Leans buy: A person reads and edits every output | Leans build: A wrong answer posts a payment, sends an order or reaches a customer |
| 6. Who owns it in a year? | Leans buy: Nobody on your team wants to carry it | Leans build: You can name an owner with time, plus a backup |
| 7. How easy is it to leave? | Leans buy: Exports are simple and the contract is monthly or annual | Leans build: A bought tool would hold your workflow data with no export |
Count the answers. Four or more on one side is a signal, not a verdict. Two questions can overrule the rest. If the data cannot leave (question 3) and no vendor can meet that, you build. If nobody can own the tool (question 6), you buy, because an unowned internal tool decays within months.
Question 2 is the one teams skip. A general assistant does not know your approval matrix or how your finance team codes a cost. You can feed it that context in prompts and files, and for a few pages of rules that works. Once the rules must be exact, tested and the same every time, you are building a tool.
What does buying cost?

Seat prices are public for most of the main products. These are list prices as of 20 November 2026, ex-tax.
| Product | Published list price | Notes |
|---|---|---|
| Microsoft 365 Copilot | Published list price: $30.00 per user per month paid yearly, or $31.50 paid monthly on an annual commitment | Notes: Needs a separate qualifying Microsoft 365 plan |
| Microsoft 365 Copilot Business | Published list price: $18.00 per user per month paid yearly, down from $21.00 | Notes: For plans of up to 300 users. The discount runs 1 July to 31 December 2026, needs an annual commitment and applies to the first year only. Some advanced features bill by usage (Copilot Credits) on top |
| Copilot Chat | Published list price: No additional cost with eligible Microsoft 365 subscriptions | Notes: Web-grounded. The paid licence adds Microsoft Graph work data |
| Claude Team, Standard seat | Published list price: $20 per seat per month billed annually, $25 billed monthly | Notes: 2 to 150 seats. Premium seat: $100 annually |
| Claude Enterprise | Published list price: $20 per seat per month billed annually, 20 seat minimum | Notes: Usage is billed separately at API rates |
| ChatGPT Business | Published list price: See per-user pricing on OpenAI's site | Notes: Starts at 2 users, billed monthly or annually |
| ChatGPT Enterprise | Published list price: No public price | Notes: OpenAI's pricing page says to contact sales |
The seat price is not the whole cost. Someone has to run the rollout, fix permissions and train people. And a seat covers many tasks, so if only a handful of people would use a feature, compare a build against those few seats, not the whole company.
What does building cost?
A build has four cost lines. Model usage is usually the smallest. People are usually the largest.
| Cost line | What it covers | Public reference point |
|---|---|---|
| People | What it covers: Design, build, testing, then upkeep | Public reference point: US BLS median annual wage, May 2025: $135,980 for software developers and $140,300 for computer and information research scientists (closest BLS categories; all experience levels; excludes benefits and equity). India: AmbitionBox shows a typical AI engineer range of ₹14.8 to 16.4 lakh a year (self-reported, 1 to 6 years' experience, updated 19 November 2026) |
| Model calls | What it covers: Tokens in and out, per document or per question | Public reference point: Public list prices, per 1 million tokens, checked 20 November 2026. Anthropic Sonnet 5.5: $2 in, $10 out. OpenAI GPT-6.2 Sol: $2 in, $10 out. Batch use is 50% off at both |
| Hosting | What it covers: Only if you self-host a model | Public reference point: E2E Networks NVIDIA H100 80GB, ex-GST: ₹379.05 an hour on demand, ₹347.70 an hour committed for 3 months or more |
| Review, evals, monitoring | What it covers: Work that makes output safe to rely on | Public reference point: No public price. Estimate as engineering time |
How big is the model bill really?
Take an illustrative intake tool that reads 10,000 documents a month. Assume each document uses about 3,000 input tokens and 500 output tokens. That is 30 million input tokens and 5 million output tokens a month. At $2 and $10 per million, the bill is $60 for input and $50 for output, about $110 a month. The assumptions are ours, so use your own volumes. For a typical internal tool, the token bill is small beside the salary of the person who maintains it.
When does self-hosting pay?
One H100 running all month at the on-demand rate is about ₹2.73 lakh (379.05 times 720 hours), before any second GPU for resilience. Against a model bill of around $110 a month, that is not a saving. Self-host when the data cannot go to a provider or volume is very high, not because it feels cheaper. See our private LLM versus API cost model in rupees.
How do you compare build cost with seat cost?
Use one formula for each use case:
- Build side, over three years: build and upkeep effort, plus model calls, plus hosting if any.
- Buy side, over three years: seats that would use the feature, times list price, times 36 months, plus rollout and training effort.
- Adjust for coverage. A seat covers many tasks. A build covers one. Compare only if the seat does the same job to the same standard.
For scale, one US median software developer salary of $135,980 equals the yearly list price of about 380 Copilot seats at $360 each. If a tool serves 20 people and one engineer must look after it, buying usually wins on cost alone. If it removes hours of repeated work for an expensive team, or carries a process no vendor sells, building can win.
What do the studies say about the value of bought seats?
Buying does not buy adoption. Two public UK government evaluations of Microsoft 365 Copilot show why.
- Department for Business and Trade pilot. 1,000 licences from October to December 2024, about 70% volunteers and 30% randomly assigned. Mean daily active use was 21% (30% excluding weekends and bank holidays) and mean weekly active use was 64%. Satisfaction was 72%. The report found no evidence that time savings had led to improved productivity.
- Cross-government experiment. 20,000 civil servants in 12 organisations, 30 September to 31 December 2024, published 2 June 2025. Users reported saving an average of 26 minutes a day. That figure is self-reported. Satisfaction averaged 7.7 out of 10, and 82% said they would not want to go back.
Microsoft also cites a Forrester study of 100%+ return with payback in 10 months. It was commissioned by Microsoft and models a composite organisation over three years, so treat it as vendor material.
The lesson: seats pay back when the task is chosen, permissions are clean and people are shown how to use the tool. Budget for that work. Our 90-day Copilot adoption plan covers how to measure it.
What should you check before you buy a seat?
Run the same short review for every assistant. Public statements from the three main vendors, as of 20 November 2026, differ on the points that matter:
| Check | Microsoft 365 Copilot | ChatGPT Business and Enterprise | Claude Team and Enterprise |
|---|---|---|---|
| Used to train models? | Microsoft 365 Copilot: Microsoft says prompts, responses and Microsoft Graph data are not used to train foundation LLMs | ChatGPT Business and Enterprise: OpenAI says it does not train on your data by default, and you own your inputs and outputs where allowed by law | Claude Team and Enterprise: Anthropic says commercial inputs and outputs are not used to train by default. Feedback and opt-in are exceptions |
| Who can see what? | Microsoft 365 Copilot: Copilot only surfaces organisational data a user already has at least view permission to | ChatGPT Business and Enterprise: Check workspace and connector permissions | Claude Team and Enterprise: Check workspace and connector permissions |
| Where is data processed? | Microsoft 365 Copilot: Outside the EU, queries may be processed in the US, EU or other regions. Do not assume India-only processing | ChatGPT Business and Enterprise: Enterprise content can be stored at rest in India for new Enterprise and Edu workspaces. Model inference on that content does not run in India. Metadata, logins, billing and external integrations may sit outside the region | Claude Team and Enterprise: The pricing page makes no residency claim. Ask the vendor in writing |
| Retention controls | Microsoft 365 Copilot: Admins can set retention for Copilot interactions in Microsoft Purview | ChatGPT Business and Enterprise: Retention control applies to Enterprise, Edu and Healthcare plans | Claude Team and Enterprise: Custom data retention controls are Enterprise only |
Copilot follows your existing SharePoint and OneDrive permissions. If those are messy, Copilot makes the mess easy to search. Fix permissions before rollout. The same logic applies to any assistant you connect to shared drives.
If staff already use free consumer tools, a bought seat with proper terms can reduce risk. A shadow AI audit shows what you would be replacing.
What does the build side look like in practice?

Take reading supplier invoices into your finance system. You could buy a specialist product. If your rules are unusual (your own approval limits, your own purchase-order matching), you might build on a cloud extraction service. The parts are public, but most of the work sits around the model:
- Extraction. A cloud document service such as AWS Textract or Azure Document Intelligence returns a confidence score with each extracted value.
- A review queue you build yourself. The vendors' built-in human review tools are no longer an option for new builds: Amazon A2I is no longer open to new customers, and Google Document AI Human-in-the-Loop has been deprecated since 16 January 2024 (provider docs, checked 20 November 2026).
- Validation after the model. Schema-valid output is not the same as correct output. Your code still checks that totals add up, the tax number has a valid format and the purchase order matches.
A built tool is a model call in the middle with a lot of ordinary engineering around it. Our guide to a document intelligence pipeline covers the stages, confidence thresholds, the review screen and an accounts payable example in detail.
When is a hybrid the right answer?
For most firms, yes. Here is how we would sort common use cases. This is judgement, not a rule.
| Use case | Usual pick | Why |
|---|---|---|
| Drafting, summarising, meeting notes | Usual pick: Buy a seat | Why: Generic task, many users, a human edits the output |
| Search across shared files | Usual pick: Buy first, build if it fails | Why: Fast to try, but only as safe as your file permissions |
| Invoice or form intake with your own rules | Usual pick: Build on a cloud API, with your own review queue | Why: Your validation rules are the value |
| A customer-facing or approval workflow | Usual pick: Build | Why: A wrong answer has a cost, so logic must be tested |
| Work on data that cannot leave your environment | Usual pick: Build, possibly self-hosted | Why: No vendor option passes review. Consider a private AI setup |
What does a build commit you to owning?
A built tool is a product. Before you start, agree who owns each of these:
- An owner and a backup, named, with time set aside each month.
- An evaluation set. Real examples with known answers, run before every prompt or model change.
- Monitoring. Error rates, review-queue volume, cost per document, and model retirements.
- Access control and logging. Who can use it, what it can see, what is kept, and erasure if personal data is involved.
- A review date. Many internal tools outlive their need.
If you cannot fill in that list, buy.
How do you write the decision down?
Keep a one-page decision record for each use case: the task, the seven answers, the option chosen, the cost comparison and its assumptions, the vendor review result, the owner and a review date. Revisit it at least yearly. Vendors add features and prices change, so what was worth building can become a bought feature.
Where can Mono help?
Mono helps companies decide which AI tools to buy and which to build, then supports the rollout or the build. That includes a use-case review, a vendor data review and an adoption plan. See our AI adoption service. Legal questions about vendor contracts go to your counsel.
Sources: Microsoft 365 Copilot enterprise and Copilot Business pricing pages (microsoft.com), checked 20 November 2026. Microsoft Learn, Microsoft 365 Copilot privacy, updated 4 November 2026. OpenAI, Enterprise privacy (updated 8 January 2026), ChatGPT pricing, and Data residency help article (updated November 2026). Claude pricing page (claude.com/pricing) and Anthropic privacy centre, model training article, dated 18 August 2026. UK Department for Business and Trade, Microsoft 365 Copilot pilot evaluation report. UK Government, Microsoft 365 Copilot experiment: cross-government findings report, 2 June 2025. US Bureau of Labor Statistics, Occupational Outlook Handbook, May 2025 median pay for software developers and for computer and information research scientists. AmbitionBox, AI Engineer salary in India, 19 November 2026. E2E Networks pricing. Anthropic and OpenAI API pricing pages, checked 20 November 2026. AWS documentation for Textract AnalyzeDocument and Amazon A2I. Google Cloud documentation for Document AI Human-in-the-Loop. Last reviewed 20 November 2026.
FAQ
Should we build or buy AI internal tools?
Buy general assistants for generic tasks such as drafting, summarising and meeting notes, where many people need light use and a human edits the result. Build tools that carry your own rules, depend on your own data, sit inside a process where a wrong answer costs money, or must keep data in your environment. Most companies end up with a bought assistant for everyone and a few built tools. Decide each use case separately.
How much does Microsoft 365 Copilot cost per user?
As of 20 November 2026, Microsoft lists Microsoft 365 Copilot at $30.00 per user per month paid yearly, or $31.50 paid monthly on an annual commitment, and it needs a separate qualifying Microsoft 365 plan. Copilot Business, for plans of up to 300 users, is listed at $18.00 per user per month paid yearly under a first-year discount that runs from 1 July to 31 December 2026. Some advanced features bill by usage on top. These are US list prices, ex-tax.
Is building an AI tool cheaper than buying seats?
Not usually, for generic work. Model usage is often a small bill, but people and upkeep dominate. One US median software developer salary of $135,980 (BLS, May 2025) equals the yearly list price of about 380 Copilot seats at $30 a month. Building tends to pay when it replaces many seats' worth of spend or a costly manual process, or when no vendor product fits your rules or data constraints.
Can I keep data in India with ChatGPT Enterprise, Copilot or Claude?
It depends on the product, and you should confirm in writing. OpenAI says new ChatGPT Enterprise and Edu workspaces can store content at rest in India, but model inference does not run in India (OpenAI help centre, checked 20 November 2026). Microsoft says that outside the EU, Copilot queries may be processed in the US, EU or other regions. Anthropic's pricing page makes no residency claim. The DPDP Act does not itself require data to stay in India. If a contract or sector rule requires processing in India, a build hosted in India, such as a self-hosted open-weight model, may be needed. Take the legal question to counsel.
Why do bought AI seats go unused?
Because buying a licence does not create a habit. In a UK Department for Business and Trade pilot of 1,000 Copilot licences, mean daily active use was 21% and weekly active use was 64%. Seats get used when each team has a named task, permissions are clean, and people are shown how to use the tool on real work. Plan that rollout effort as part of the cost of buying.
