Week 4 | Enterprise track
Build, Buy Or Partner
- Week 4 of 12: AI integration for companies
- Promise: you will choose a route with a clear, explainable reason.
Outcomes
What You Will Be Able To Do
- Compare build, buy and partner
- Classify vendors by type
- Explain foundation models and open versus closed
- Estimate total cost of ownership
- Reduce lock-in risk
Overview
Three Routes
- Build: your team creates it
- Buy: you license a finished product
- Partner: an outside firm co-delivers
- Most companies use all three
| Route | Best when | Main risk |
|---|
| Build | Unique advantage | Slow start, skills |
| Buy | Common function | Lock-in, limited fit |
| Partner | Skills gap | Knowledge leaves |
Build rule
When To Build
- Build where it gives an advantage rivals cannot copy
- Unique data or expertise is the usual source
- Common functions are cheaper to buy
- Check you can keep the team to maintain it
Landscape
Five Vendor Types
- Embedded AI in tools you already use
- Specialist applications for one job
- Model providers selling API access
- Cloud platforms with models and tools together
- Integrators and consultancies
Models
What A Foundation Model Is
- Large model trained on broad data
- Adapted to many tasks
- Companies rarely train their own from scratch
- Use by prompting, retrieval or fine-tuning
Pricing basics
Size, Speed And Tokens
- Larger models: harder tasks, higher cost
- Smaller models: faster and cheaper
- Route simple requests to small models
- Pricing often by tokens: pieces of text
Trade-offs
Open-Weight Versus Closed
- Closed: service only, often most capable, little setup
- Open-weight: you run it, control data location
- Licenses differ, read them
- Many companies mix both
| Question | Closed | Open-weight |
|---|
| Where does data go? | To provider | Your environment |
| Setup effort | Low | Higher |
| Control over changes | Provider | You |
| Capability | Often highest | Varies |
Cost
Total Cost Of Ownership
- Count three years, not the first invoice
- Fees, integration, data preparation, security review
- Training, support, monitoring, staff time
- Include the cost of exit
Cost traps
Ask For Pricing At Your Volume
- Demo volume is not your volume
- Ask the price at twice your volume
- Check how fees grow with users
- Week 9 covers running costs and return
Risk
What Lock-In Looks Like
- Data stored in a vendor format
- Workflows and prompts tied to one model
- Staff trained on one interface
- Long contract commitments
Mitigation
Reducing Lock-In
- Keep your data and outputs in open formats
- Put a thin layer between apps and the model
- Test a second provider on real tasks
- Negotiate export rights and notice of price changes
Application
|
Thin model layer <- change the model here only
|
Provider A Provider B (tested fallback)
Method
The Decision Table
- Rows: your realistic options
- Columns: fit, speed, security fit, TCO, lock-in, skills
- Score 1 to 5 and discuss the gaps
- Write a reason a manager can repeat
Due diligence
Questions To Ask Every Vendor
- Where is our data stored and processed?
- Is our data used to train your models?
- Can models change without notice?
- How do we export our data and leave?
Summary
Recap
- Match route to situation: build for advantage, buy for common needs
- Know the five vendor types
- Compare open and closed on data, control and effort
- Count full cost and plan for exit
Week 5
Next Week
- Security, privacy and compliance
- KVKK, GDPR and the EU AI Act at a high level
- Data leakage and shadow AI
- Writing an acceptable-use policy
Wrap up
Questions And Exercise
- Exercise: compare three routes for your use case
- Estimate three-year cost lines as ranges
- Send seven questions to a real vendor
- Questions from the room