All insights

Build vs buy

Build vs buy AI features: a guide for SaaS teams

Buy what every company needs. Build what only your product can do.

Co-founder and CEO, Datics AI7 min read

In brief

In the build vs buy AI decision for one feature, build when the feature is core to what customers pay for and must work inside your own data, workflows and permissions. Buy when the job is generic, such as meeting notes or internal search, and a vendor already does it well. Partner when the feature must be built into your product but your team lacks the AI experience or the time.

On this page
  1. 01Build vs buy AI: what each path means for one feature
  2. 02The build vs buy decision table
  3. 03How to decide for one feature
  4. 04Three worked examples
  5. 05What we see in production
  6. Frequently asked questions

01Build vs buy AI: what each path means for one feature

Most build vs buy advice is written for company strategy: whether to hire an AI team, which platforms to standardize on, where to place long bets. We cover that in AI strategy: build, buy or partner?. This article is narrower. You have one feature on the roadmap, and you need to decide how it gets made.

Build means your engineers write the feature into your product. It almost never means training a model. You still call a model from OpenAI, Anthropic or an open-weight family. What you build is everything that makes the answer correct for this user: retrieval over your records, the permission checks, the screens, the tools the AI may call, and the test set that tells you whether it works.

Buy means a vendor product does the job, and your team configures it instead of coding it. A meeting notes tool, a support search widget, or an assistant that comes with software you already license all count. You get speed and someone else's roadmap. You give up control of how it fits your data and your screens.

Partner means an outside team builds the feature inside your codebase, on your cloud and your model choice, and hands it over. It is still a build. The difference is who does the work and how fast it starts. The code should be yours from the first commit.

02The build vs buy decision table

Score your feature against each row. When most answers fall in one column, the decision is usually clear. When they split, the first three rows weigh most.

QuestionPoints to buildPoints to buyPoints to partner
Is the feature core to what customers pay for?Yes, it sets you apartNo, every company needs itYes, and it is needed this year
How sensitive is the data it reads?Customer data under contract termsInternal or public data, or the vendor meets your termsCustomer data, and the team has not built these controls before
How closely must it follow your workflows and permissions?Every user sees only what they could see in the productA separate screen or tool is acceptableClosely, inside existing screens
Does your team have the skills?Engineers who have shipped and tested an AI featureNot needed beyond setupStrong product engineers, little AI experience
How soon must it ship?The roadmap has roomWeeksWeeks, and hiring would take months
Who carries testing and upkeep?Your team, with a test set run on every releaseThe vendor, but you still test it on your casesThe partner first, then your team with the handed-over tests
What does leaving cost?Nothing; you own itData export and retraining usersNothing, if the code and tests are in your repositories

Two rows deserve a second look. Permissions decide more of these choices than teams expect: in our project records, 13 of 52 AI projects had to respect each user's permissions, and a bought tool that cannot see your permission model either shows too much or answers too little. And upkeep is larger than it looks. The authors of Hidden Technical Debt in Machine Learning Systems showed that the model code is a small part of a real system, and that the infrastructure around it, such as data collection, configuration and monitoring, is where most of the work lies.

03How to decide for one feature

This is the order we follow with a client. It fits in one working session if product, engineering and security are in the room.

  1. 01

    Write the feature as one user outcome.

    Name the screen, the person and the step it serves. "AI for support" cannot be decided; "agents see a suggested reply with its source in the ticket view" can.

  2. 02

    Check what you already pay for.

    A tool your company licenses may already do the job, with your access controls in place.

  3. 03

    Score the table above.

    Write the answer to each row, not only the column it points to.

  4. 04

    Test before you sign or build.

    Run twenty or more real questions with known good answers through the vendor tool or a quick prototype. A demo on clean sample data hides the problems; our guide to running an AI POC that reaches production explains why real data matters. If the feature reads customer data, review it against the OWASP Top 10 for LLM applications, whether you build it or buy it.

  5. 05

    Price the whole life of the feature.

    Count model use, testing, monitoring and model upgrades, not only the license or the first build. The NIST AI Risk Management Framework treats measuring and managing an AI system as continuing work, and your budget should too.

  6. 06

    Write down what would change the decision.

    A new enterprise contract, a price change, or a vendor that cannot meet a data rule are common triggers.

If you have more than one candidate feature, rank them first. Our free worksheet, Pick your first AI feature, walks each idea through four questions and a five-factor score. In its worked example, the top idea needed no AI at all.

04Three worked examples

These are illustrative examples on fictional products, not client work. Each shows one path.

Buy: call summaries for an internal sales team. A property management SaaS wants its account executives to stop writing call notes by hand. The calls are internal, the data is the company's own, and every sales team has the same need. A meeting notes tool already approved by IT does the job in a week. Building it would spend engineering time on something that does not set the product apart.

Build: explaining a payroll run. A payroll SaaS wants customers to ask why this month's run differs from last month's and get an answer tied to their own records. The feature is core to the product, reads data under contract terms, and must show each user only the employees they may see. The team has a platform group that has already shipped search and tested it. They build it, on a managed model inside their own cloud account.

Partner: feedback on student writing before the school year. An EdTech platform needs AI feedback on student writing inside its existing lessons. District contracts require student data to stay in its cloud, every teacher must see only their own classes, and no one on the team has tested an AI feature before. Hiring would take months. A partner builds it in the platform's own repository, with the test set handed over, and the team takes it from there.

05What we see in production

The honest answer is sometimes "keep what you bought." OCCI, a U.S. industrial contractor, had already built a Copilot Studio agent over its reporting database and needed proof that its figures could be trusted. We tested it inside OCCI's own Microsoft tenant on live job cost data and checked each of 20 answers against our own SQL query. Copilot Studio already carried OCCI's access controls, security posture and connections to SQL and Power BI. When OCCI asked about Claude or custom code, we advised staying, because custom code adds flexibility but also security overhead. OCCI chose to stay on Copilot and pilot its weekly labor report first.

When a feature has to work inside the product, the integration is the main job. In our records, 36 of the 52 tagged AI projects, about seven in ten, needed integrations with other systems, and connecting the AI was where most of the time went. That is why a bought tool that sits beside your product often answers less than buyers expect.

The partner path works best when the team stays in charge. At CleanHaus, our engineers worked inside its team, in its own repositories, AWS accounts and logins, reviewed by its CTO and shipped through its own pipelines. Tests run in tiers, from unit tests on every push to LLM-as-judge evals, so a model upgrade became a routine change for its team. That is what a partner should leave behind.

If you are still choosing which feature to build, our Planning Sprint ranks the candidates in two weeks for $5,000 and ends with a fixed quote for the first wave.

Frequently asked questions

Is buying an AI tool cheaper than building one?

At the start, almost always. Over the life of a feature, it depends on how much the tool must be adapted to your data and permissions. A tool that fits as delivered stays cheaper. A tool that needs workarounds, exports and a second permission model can cost more than building, and still answer less.

What does "build" mean if we still use OpenAI or Anthropic models?

Building rarely means training a model. You call a hosted or open-weight model and build what makes its answers right for your users: retrieval over your records, permission checks, the screens, the tools it may call, and the test set. That surrounding work is what customers experience as the feature.

How do we keep a bought AI tool from locking us in?

Keep your own test cases, so you can measure any replacement on the same questions. Confirm you can export your data and configuration, and check where the vendor processes your data before you sign. Write down which change, such as a price rise or a new contract term, would make you revisit the choice.

When should we use a partner instead of hiring?

When the feature is core but needed sooner than you can hire, or when you want one feature shipped while you learn what an AI team should look like. Insist that the code and test set live in your repositories from the first day, so your team can take over without a gap.

How Datics helps

Datics specializes in brownfield AI: production AI built into established software, one capability at a time, in your codebase and under your name.

›References (3)
  1. NIST: AI Risk Management Framework
  2. Sculley et al.: Hidden Technical Debt in Machine Learning Systems (NeurIPS 2015)
  3. OWASP: Top 10 for Large Language Model Applications

Related work

Case studies from our projects on the same subject.

OCCI

Other software · Enterprise data assistant evaluation

OCCI kept Copilot once its live job cost answers were checked against SQL

Leaders wanted plain language answers from their own job cost data. We tested their Copilot agent in their tenant and checked each answer against SQL.

Brownfield AIForward deployedEvaluated before launch
Read case study

CleanHaus

Property Management · Turnover operations SaaS

CleanHaus took its app and AI agent live, our engineers inside its team

CleanHaus needed AI and mobile engineers without hiring a department. Our engineers worked in its code, cloud and logins until both went live.

Forward deployedBrownfield AIInside a live product
Read case study

Your customers already expect AI from your product.

Let us make it real inside the product you run today.