Gains:
- Turning model selection into a structured decision tree with seven key questions
- Ability to propose models by applying the decision tree to real business scenarios
- Evaluating the advantages and burdens of establishing a model portfolio instead of dependence on a single model
We collected the pieces across eight units: providers, implicit/explicit weight, context, multimodality, cost, tiers, and data sovereignty. Now it's time to combine these pieces into a single decision mechanism. The purpose of this unit is to answer the question "which model?" You can no longer answer the question with intuition, but with a structured decision tree. You'll also see why mature organizations gravitate towards a portfolio of models, not just one. Once the unit is finished, you will be able to apply the decision tree to real business scenarios and draft your own portfolio.
Seven Question Decision Tree
We can reduce model selection to seven questions. Answering these questions in order will almost automatically lead you to the right candidate.
- Is the data sensitive? If you are processing personal/confidential data → data sovereignty (clear weight, regional, zero retention) determines first. If not, your flexibility will increase.
- How difficult is the task? Simple/formidable → light level. Complex/multi-step → strong step + reasoning.
- How much is the volume? Very high volume → speed and cost come to the fore, lean towards the light range. Low volume → you can invest in quality.
- How big is the document/context? Long documents → wide context window is a must. Short texts → small window is enough.
- Is there visual/audio? If available → multimodal model mandatory. Otherwise, the text model is sufficient.
- What is budget sensitivity? If high → step optimization, cache, batch; maybe open weight. If low → priority on quality.
- How is the team and infrastructure? If there is no technical team → closed weight, ready interface. If there is a strong team → open weight option is on the table.
Tip: Once you answer these seven questions, the pool of candidate models usually narrows down to one to two models. The purpose of the decision tree is not to impose a single "right answer" on you, but to quickly eliminate options and make it easier for you to test the remaining ones with a little pilot.
Putting the Decision Tree in a Table
Question
Answer → Redirect
Is the data sensitive?
Yes → open weight / regional / zero before storage
Is the task difficult?
Yes → strong + reasoning / No → mild
Is the volume high?
Yes → speed & cost priority (light)
Is the document large?
Yes → wider context
Is there a visual?
Yes → multimodal mandatory
Is the budget limited?
Yes → optimization + maybe open weight
Is there a technical team?
None → closed/ready interface
This table works like a checklist: tick each row, assemble the resulting profile. For example, the “data sensitive + task difficult + low volume + technical team available” profile will likely direct you to a strong openweight model running on your own server.
Not a Single Model, a Portfolio
Experienced institutions realize after a while that doing everything with a single model is neither the cheapest nor the best way. Instead, they build a portfolio of models: using different models depending on the task. The logic is simple:
- Simple, bulky jobs → light and cheap model.
- Complex, critical work → powerful model.
- Jobs processing confidential data → open weight/regional model.
- Visual works → multimodal model.
The portfolio approach has two major benefits. The first is cost optimization: you have each job done by a model as powerful as it needs, and you do not pay for more. Second, provider independence: you are not dependent on a single provider; If someone raises prices, changes policy, or cuts service, you have an alternative ready. This is called "avoiding vendor lock-in".
Caution: The portfolio approach is not free; Managing and monitoring multiple models and ensuring consistent quality adds overhead. If you're a small team, it's often smarter to start with a single balanced model and move to a portfolio as your business matures. Establishing the portfolio earlier than necessary may create management confusion.
A Portfolio Example
Below is a sample portfolio that a mid-sized service company might build:
business type
sample task
Selected model type
Why
High volume sorting
Comment/ticket tagging
slightly off
Cheap, fast, sufficient
General content production
Email, draft description
balanced off
good balance
complex analysis
Contract risk review
Strong off (wider context)
Accuracy critical
Confidential data processing
Financial/patient data analysis
Open weight (own server)
Data does not go out
visual processing
Invoice/photo reading
multimodal
visual requirement
This table is based on the one-sentence question "which model do you use?" It shows that the question is actually the wrong question. The correct answer is: "Which model for which job?"
Three Realistic Cases: Applying the Decision Tree
Case 1 — Small law firm. An office of five people wants to summarize the contract. Decision tree: data sensitive (yes, but manageable with explicit consent and contract) → task difficult (yes) → volume low (~100 per month) → document large (yes, 40+ pages) → no images → budget medium → no technical team. The result: leveraging a large-context, powerful, closed-weight model from the off-the-shelf interface with zero retention contracts. The low volume makes the cost of the powerful stage manageable.
Case 2 — E-commerce operation. Team of twenty people, they have three jobs: product description (high volume), customer review classification (very high volume), data extraction from invoice image (medium volume, visual). Portfolio instead of single model: balanced model to statements, light model to comments, multimodal model to invoices. Thus, the appropriate cost and talent for each job falls; the total bill comes down to roughly a third of what it would be if they had the powerful model do everything.
Case 3 — Public institution. A public institution wants to summarize citizen applications, but the data is personal and is not intended to go abroad. The decision tree crashes at the first question: data is sensitive and sovereignty is mandatory → explicit weight/regional model comes first. Since they have a technical team, they run an open weight model on their own infrastructure. Other questions (difficulty, volume) fine-tune which tier they choose.
Weak Prompt / Strong Prompt
Weak prompt:
Which model do you recommend us?
Powerful prompt:
Your role: AI solution architect. I answered these seven questions for my job:1) Is the data sensitive: [ANSWER]2) Task difficulty: [ANSWER]3) Volume: [ANSWER]4) Document size: [ANSWER]5) Visual/audio: [ANSWER]6) Budget sensitivity: [ANSWER]7) Technical team: [ANSWER]Task: Narrow down the candidate model type based on these answers (stage, weight type, context, multimodality). Suggest 2 concrete candidates. Then justify whether I should solve this problem with a single model or a portfolio of 2-3 models.
The powerful prompt carries the answers of the decision tree directly to the model; Thus, the model does not make guesses, it narrows down according to your real profile.
Copiable Templates
1. Running a decision tree:
Let me define my job: [DESCRIPTION]. Ask the seven decision questions (data sensitivity, task difficulty, volume, document size, image, budget, team) for me in order, narrow down the candidate pool after each answer, and finally recommend 2 model types.
2. Portfolio outline:
I have the following jobs: [JOB LIST]. Apply the decision tree for each job, match the appropriate model type and create a portfolio table (job / model type / justification). Evaluate the management burden of the portfolio in a single sentence.
3. Single model or portfolio:
My team size [NUM], my work diversity [LOW/MEDIUM/HIGH], my monthly AI budget[RANGE]. Should I start with a single balanced model or move on to a portfolio immediately?Compare the pros and cons, suggest a phased path.
4. Addiction risk assessment:
We are currently dependent on a single provider [PROVIDER]. What are the risks of this dependence (price, policy, disruption)? Which secondary provider should I keep ready as a backup to reduce risk? Suggest 3 measures that will make the transition easier.
Common mistakes
- Choosing by intuition: Skipping the decision tree and choosing a model because "it sounds good"; Later, incompatibility arises.
- Leaving the data question for last: In sensitive data, the question of sovereignty comes first; Putting it at the end leads to locking into the wrong model.
- Delegating every job to a single model: Losing either quality or budget by having jobs of different difficulty and volume done by a single model.
- Establishing the portfolio too early: Creating unnecessary complexity by trying to manage three or four models as a small team.
- Not thinking about dependency at all: Being completely dependent on a single provider and being left with no alternative in price/policy changes.
In summary
- Model selection can be boiled down to seven questions (data, challenge, volume, context, visual, budget, team); these questions quickly narrow the candidate pool.
- In sensitive data, the question of sovereignty comes first and frames other decisions.
- Rather than a single model, mature organizations establish a portfolio of different models depending on the task; This both optimizes cost and reduces dependency.
- The portfolio has a management overhead; Small teams should start with a single balanced model and move gradually.
Application task
Bring together all the notes you made in previous units (needs profile, weight preference, token/cost estimate, tier diagnostics, data map). With pattern 1 in this unit (running a decision tree), answer seven questions for your business and arrive at two candidate model types. If you have more than one type of work, create a portfolio table with template 2 (portfolio outline). This draft will be the input to the evaluation and validation study in the final unit.
checklist
- [ ] By applying the seven decision questions in order, I can narrow down the candidate pool.
- [ ] I know why the question of sovereignty comes first when it comes to sensitive data.
- [ ] I can apply the decision tree to a real business scenario and propose a model.
- [ ] I can evaluate the advantages and management burden of the model portfolio.
- [ ] I know the risk of dependence on a single provider and ways to reduce it.