Independent AI consultant or agency? Choose by who owns the decision.

Hire an independent consultant when your startup needs one senior person to own the technical decisions. These decisions include the model, the vendors, the features to build and the features to stop. Hire an agency when you have made these decisions and you need more engineers to build the product. Use a talent marketplace when you need a list of candidates quickly and you can examine them yourself.

A wrong choice usually causes one problem. The person who makes the decisions is not the person who gets paid for the result.

I am an independent consultant. Read this page with that fact in mind. I also tell you where the independent model is weak.

This page compares the types of provider. If you do not know whether you need outside help, read the signs you need a fractional AI CTO first. If you selected the type of provider and you must now select a person, use the AI CTO interview and selection guide.

What are the options for a generative AI startup?

All four types can do good work. Select the type that matches your problem now. A startup at seed stage usually has one of two problems:

• Nobody can decide what to build. This is a problem of judgment.
• The team agrees on the plan, but nobody has time to build it. This is a problem of capacity.

Why is the incentive more important than the resume?

Each type of provider earns money in a different way. The method of payment has an effect on the advice that you get.

An agency earns money for each hour of build work. This is correct when you need a build. It is a problem when the correct advice is “do not build this now.” It is also a problem when an existing API and a scheduled script can do the work. Most agency engineers are honest. But the business model pays for more scope. Remember this when you read an agency proposal.

An independent consultant on a monthly block gets the same payment for one feature or for five features. Thus, it is easier for this person to tell you to make the scope smaller. The risk is different. The consultant can give advice and never test it. Ask the consultant what they will build or measure in the first month.

A marketplace earns a margin on each match. Its work stops after the introduction. The quality of the work after that point depends on the person that you select.

A consulting firm earns money on a program. The program helps many teams in a large company agree on one plan. A startup with ten people does not usually have this problem.

What is different for a generative AI startup?

Generative AI adds decisions that other software startups do not have. Ask each provider how they manage these four areas.

Model choice. Model providers release and retire models on their own schedule. They also decide how much notice you get. One person must own the selection of the model and the plan to change it. That person must not get a referral fee from a model vendor or a platform.

Evals. An eval is a test that you can repeat. It compares the model output with examples that have a known good answer. Without evals, nobody can tell if a prompt change or a new model made the product better or worse. I think that the eval suite is the main asset of a generative AI product. You can replace the model. You cannot easily replace the tests.

Cost per task. Each user action has a cost in a generative AI product. A popular feature can lose money each time a user uses it. One person must measure the cost of each task. The total monthly API bill does not give this information.

Data and safety. Customer data goes into prompts, logs and search indexes. Agents can use tools and do actions in other systems. One person must decide which data the model can see. That person must also decide which actions need approval from a human.

An agency can do this work in a fixed scope. But these decisions do not stop after one project. They occur again each time that the product or the model changes. For this reason, many generative AI startups need a technical owner before they need a larger build team.

Where is an independent consultant strong?

One person knows the full design. Agents can now write code faster than a team can review it. Thus, the important skill is to keep the architecture consistent. An independent consultant who reviews each important change can do this. An agency team that changes its engineers frequently cannot always do this.

Advice that does not depend on a vendor. An independent consultant does not have a partner program. The consultant can recommend an open-weight model in your own cloud, a hosted API or no model.

Direct access. You speak directly to the person who makes the decision. No account manager stands between you and the technical decision.

Public evidence. Many independent consultants work in public. You can read their code and their articles before the first call. This evidence shows how they think. A sales presentation does not show this as well. My public work includes agentic_security and the articles on this site. agentic_security is an open-source tool that tests LLM agents against attacks.

Where is an independent consultant weak?

Capacity. One person has one calendar. Maybe your problem is “we need four engineers for three months.” In that case, an independent consultant is the wrong hire. The consultant can plan the work and review it. The consultant cannot do all of the work.

Continuity. The person can become sick or leave. Then the knowledge can leave with the person. Prevent this with the contract. Put all decisions in a written log. Keep all code in your repositories. Keep all credentials in your accounts. Do not depend on the memory of one person.

Range of skills. One person is not an expert in all areas. A good independent consultant knows the limits of their skills. This person brings in a specialist when necessary. Examples are a penetration test, a legal review or a special integration.

When is an agency the better choice?

Select an agency when three conditions are true:

• You have a clear specification.
• A technical owner on your side can review the work.
• The deadline needs more than one or two engineers.

An agency is strong in staffing and delivery process. Give it a clear scope. Also give it an owner on your side who can stop changes to the scope.

Here is a hypothetical example. A startup tested a document-review assistant with ten pilot customers. The team agreed on the architecture, the model and the evals. Now the startup needs a web app, admin tools and integrations in one quarter. This is a problem of capacity. An agency is a good choice. An independent AI CTO can also review the agency work for the startup.

Many startups use both types. The independent consultant owns the decisions and the eval suite. The agency builds the product to these decisions. This method works when both sides write down their roles before the work starts.

When is a talent marketplace the better choice?

A marketplace is useful when you must see several candidates in one week. It is also necessary that you can examine the candidates yourself. The marketplace does the search for you. It does not do the selection for you.

Use the same selection process that you use for a senior hire:

1. Write a role brief.
2. Review examples of real work.
3. Speak to references.
4. Give a short paid assignment.

The selection guide explains each step.

How do I examine an independent AI consultant?

Ask these questions on the first call. Clear answers are more important than impressive answers.

1. What will you own, and what will my team own? Get the answer in writing. “Advise on AI” is not a scope.
2. Do you get referral fees or partner commissions from model, cloud or tool vendors? If the answer is yes, ask the consultant to show the fee in each recommendation.
3. Do you write code? If the answer is yes, ask where the code goes. The code must go into your repositories. Your company must own it.
4. What did you put into production, and what failed after the launch? People who operated generative AI systems remember the failures in detail.
5. How many clients do you have at the same time? How quickly do you reply? A part-time role needs a defined reply time for urgent problems.
6. What occurs when we stop the work? Ask for a handoff plan. The plan includes the decision log, the eval suite, the removal of access and a final review.

What must the contract include?

Keep the contract short. Include these items:

• IP and code ownership. All code, prompts, eval data and documents are the property of the company.
• Access. The consultant works in your accounts. You can remove all access in one step.
• Scope and hours. The contract gives a monthly block with a set number of days. It also tells how the consultant replies to urgent requests.
• Conflicts. The consultant tells you about each vendor relationship that can have an effect on a recommendation.
• Notice period. The notice period is short. Both sides can stop the work without a dispute.
• Handoff. The contract lists the documents and reviews that the consultant gives you at the end.

When must I not hire an independent consultant?

Do not hire an independent consultant in these conditions:

• You have a technical leader who put a generative AI system into production, and this person has time to own the new system.
• The work is one narrow task. Examples are a security test, a data migration or one integration. A specialist does this work better.
• You need a technical cofounder who stays for many years. This is a different role, with equity and long-term ownership. The fractional CTO vs technical cofounder guide compares the two roles.

Next step

Write your main problem in one sentence. Then use this rule:

• If the problem is about decisions, find an independent technical owner.
• If the problem is about capacity, find a build team. Also find a person on your side who reviews the work.
• If the problem is about both, hire the decision owner first. Then the build team gets a specification.

Do you want an independent consultant for this role? Read how my fractional AI CTO engagement works and how I set rates. For an early-stage team, read AI CTO for early-stage startups. That page explains how the role changes from pre-seed to seed.