Build, buy or partner

What you will be able to do

Decide how Larkfield gets an AI capability: make it, buy it, or make it with a specialist. You compare the three ways on six points and count the people’s time that no price list shows. Then you choose part by part, not for the whole product at once.

Before you startWhat to know and to have readyAbout 30 minutes

What you need to know

• The four number labels (recorded, price, assumption, constructed), from Read the evidence pack.
• What Larkfield’s support reply assistant does, from the pilot summary.

What you need

• A browser and a document tool. Open the vendor comparison template. You use its criteria in this lesson and fill it in at the end of the module.

You need no account and no money. You run nothing.

Dinner for twenty guests

You invite twenty people to a family celebration. You have three ways to feed them.

You can cook it yourself. You choose every dish, and you know what is in the food. But you need two days, a big kitchen and the skill to cook for twenty. You can buy ready meals from a shop. It is fast, and the shop has the skill. But you get the shop’s menu, and you cannot ask how it made the sauce. Or you can hire a caterer, who cooks your family’s recipes in your kitchen with your help. You share the control and the cost.

No way is always the best. For a weekday dinner, you cook. For a wedding, you hire help. And a smart host mixes them: a home-made main course and a bought dessert.

The problem: the price list hides the biggest cost

Leaders face the same choice for AI work, and it has a name: build versus buy. Build means that your own team makes and runs the system. Buy means that a vendor sells you a finished product or service. A third way is to partner: a specialist company builds it with you, often in your own systems, and hands it over.

The usual mistake is to compare the prices on the price lists. The system design and cost course shows this with a database for Larkfield. A managed database, which the provider runs, has a list price of US$17.55 a month. A database on your own rented server costs US$53.58 a month: about three times as much, but still small.

Then count the people. The managed database needs about 1 hour of the team’s work a month: reading alerts, testing a restore. Your own server needs about 6 hours: updates, backups, monitoring. These hours are an assumption, priced at US$60 an hour.

The price list (list) against the whole cost (all), for Larkfield's database. Buy (managed): US$17.55 on the price list, US$77.55 with 1 hour of people's time. Build (your own server): US$53.58 on the price list, US$413.58 with 6 hours. Prices from the system design course; the hours and the US$60 rate are assumptions.

Text description of the chart

A column chart with four columns, in US$ a month. Buy, list (price list only): 17.55. Buy, all (whole cost) with 1 hour of people’s time at US$60: 77.55. Build, list: 53.58. Build, all, with 6 hours at US$60: 413.58.

The price list showed a gap of about US$36 a month. The whole cost shows a gap of US$336.03. People’s time is the biggest part, and it is on no price list. Most AI work is the same: the model fees are small. The cost of the people who run, check and change the system is large. Estimate value and return found the same: the model costs 0.03% of the assistant’s value.

The idea: six questions, asked part by part

A good build, buy or partner decision compares the three ways on the same six points. The matrix below shows how each way usually scores. Your case may differ, so treat it as a list of questions, not as answers.

The build, buy or partner matrix. Green is usually a strength, amber depends on the deal, red is usually a weakness. Each cell is a question to check in your own case.

Text description of the diagram

The matrix compares three ways to get an AI capability on six points.

• Control. Build: full; you change anything at any time. Buy: low; the vendor decides what changes. Partner: shared, as the contract says.
• Speed to a first result. Build: slow, months to hire and build. Buy: fast, weeks, if the product fits as it is. Partner: medium; the partner brings the skills.
• Data access. Build: your data stays with you. Buy: the vendor holds it, so read the terms. Partner: the partner’s staff get access, so set limits.
• Skills you need inside. Build: many; you hire and keep them. Buy: few; enough to test the product and manage the vendor. Partner: some, to take the work over at the end.
• Lock-in. Build: to your own people and code. Buy: high, to the vendor’s model, data and contract. Partner: medium; it depends on the hand-over.
• Total cost. Build: high at the start, and people every month. Buy: low at the start, then fees for as long as you use it. Partner: project fees, then your own people.

Two terms need a definition. Vendor lock-in is the cost of leaving a vendor. It includes the data you cannot export, the contract months still to pay, and the work to replace the product. Control is your power to change the system’s behaviour and to see why it failed.

The most useful idea is the last one: choose part by part. An AI product has several parts, and you can build, buy or partner for each part. Build the parts that make your company different. Buy the parts that many companies need in the same way.

Larkfield’s assistant, part by part

Larkfield already made this choice once, without calling it a decision. Look at the parts of its support reply assistant.

Larkfield's support reply assistant, part by part. The model is bought. The prompt, the test tickets and the release gate are built, because they hold Larkfield's own policy and knowledge. The agents are Larkfield's people.

Text description of the diagram

A customer’s ticket goes to the language model. The model is bought: a hosted model that Larkfield pays per use. The prompt with Grace’s written policy also goes to the model; the team built it. The model gives a draft reply, a team and a decision: does a person need to handle this ticket? An agent reads the draft, edits it and sends it. Separately, the team’s labelled test tickets and its release gate check every change to the prompt before it goes live.

The model is bought. To train a language model costs far more than a shop can spend, and many companies need the same model. The prompt with Grace’s policy, the labelled test tickets and the release gate are built. The release gate is the list of written tests that every change must pass before it goes live. Run a fair pilot shows it line by line. They hold what only Larkfield knows: its rules, its customers’ questions, and the mistakes that it cannot accept.

The bought part is cheap. At 822 tickets a day, the hosted model costs about US$8.53 a month in model fees, recorded in the fine-tuning course. The built parts cost people’s time, and they are the reason that the assistant sends every ticket that needs a person to a person.

Discuss: should Larkfield build its own model?

Nadia, Larkfield’s head of customer operations, must answer a proposal. Tomás, the AI engineer, ran an experiment in the fine-tuning course: he trained a small model of Larkfield’s own on 900 example tickets. He wants it to replace the hosted model.

• Tomás: “Our own model answers in 0.24 seconds a ticket, and the tickets never leave our account. We stop depending on a vendor.”
• Omar, the product manager: “Why build anything? A help-desk vendor sells a drafting add-on. We could switch it on next month.”
• Kwame, the solution architect: “Or we bring in a specialist firm. They know models; we know our tickets.”
• Grace, the support lead: “I care about one thing first. Every ticket that needs a person must reach a person.”

Before you read on, answer these:

1. Which part of the assistant would each proposal build, buy or give to a partner?
2. What does the evidence say about the quality of Tomás’s own model?
3. What would each option cost a month, with people’s time?

The course’s answer. Keep buying the model, and keep building the parts that hold Larkfield’s knowledge. The evidence is clear on quality and on cost:

	Hosted model (bought)	Tomás’s own model (built)	Label
Tickets that need a person, sent to a person (development tickets)	23 of 23	0 of 23	recorded
The same, holdout tickets	24 of 24	0 of 24	recorded
Cost a month at 822 tickets a day	US$8.53 in model fees	US$1,241 to host it (US$1.70 an hour), before any people’s time	recorded / price

Tomás’s model is fast and private, but it sent no ticket that needed a person to a person. It also costs about 145 times more a month to run. The fine-tuning course reached the same verdict: do not ship it. A partner does not change this. It would build the same thing for a fee, and no partner has shown a result on Larkfield’s tickets.

Omar’s idea is different: it buys the whole product, also the parts that hold Larkfield’s knowledge. That can be right. But a bought product must prove itself on Larkfield’s own tickets, against today’s assistant. The next two lessons show how to ask for that evidence and how to test it fairly.

What would change the answer: a law that keeps tickets inside Larkfield’s systems would remove the hosted model as an option. A volume many times larger would change the cost of hosting. If nobody on the team could look after the prompt and the release gate, buying the whole product, or a partner, would become more attractive.

When to build, buy or partner, and what each costs

Build a part when it holds your own knowledge or makes your product different. You also need people to keep it running, now or after you hire them. Building costs the most people’s time, and those costs continue every month.

Buy a part that many companies need in the same way, when a product passes a fair test on your own examples. Buying is the fastest start. It costs fees for as long as you use it, and some control over your data and your future.

Partner when you need skills for a limited time, and you plan to take the work over at the end. A partner costs project fees, and a hand-over that you must plan from the first day.

In every case, write down the people’s time as an assumption with a range, next to the prices. The system design and cost course shows how to build that cost model in detail.

Common mistake

Comparing the price lists only

Symptom: a proposal says “our own server costs three times the managed one, but we can manage it”. Three months later, the team spends six hours a month on updates and backups that the provider did before.

Cause: the comparison counted the list prices and left out people’s time. For the database, that time was most of the whole cost: US$360 of US$413.58 a month.

Fix: ask for the whole cost of each option: the price list plus people’s hours. Write the hours as an assumption with a range. Ask who does the work today that the new option moves to your team.

See it end to endBuild, buy or partner for invoice extractionCase study

In this case study, you compare an in-house build, two vendor products and a partner for invoice extraction. The option with no fees is the dearest over 36 months, because of the wrong values that reach the books.

Knowledge check

Check your understanding

Define build, buy and partner in one sentence each. Then say which of the three Larkfield uses today for its language model, and which for its prompt and release gate.

Show the answer

Build: your own team makes and runs the system. Buy: a vendor sells you a finished product or service. Partner: a specialist company builds it with you, often in your own systems, and hands it over. Larkfield buys the language model (a hosted model, paid per use). It builds the prompt with Grace’s policy, the labelled test tickets and the release gate. The decision is made part by part.

Check your understanding

The managed database has a list price of US$17.55 a month and the own server US$53.58. A colleague says: 'The difference is US$36 a month, so it does not matter.' What is missing, and what is the real difference in the base case?

Show the answer

People’s time is missing. The managed database needs about 1 hour a month, the own server about 6 hours (assumptions, at US$60 an hour). The whole cost is US$77.55 against US$413.58, a difference of US$336.03 a month. Most of the cost of the own server is people, and no price list shows it.

Check your understanding

Tomás says: 'Our own model is 0.24 seconds a ticket and private. That is enough reason to build it.' Give two pieces of recorded evidence against him, and one change that would make building worth another look.

Show the answer

First, quality: his model sent 0 of 23 development tickets and 0 of 24 holdout tickets that needed a person to a person. The hosted model sent all of them. Second, cost: hosting his model costs US$1,241 a month (US$1.70 an hour, a list price). The hosted model costs US$8.53 a month in model fees. Building would deserve another look if a law kept tickets inside Larkfield’s systems, or if the volume grew many times larger. Even then, only a new model that passes the same tests could replace the hosted one.

Recap

• Build (your team makes it), buy (a vendor sells it) or partner (a specialist makes it with you). Compare them on control, speed, data access, skills, lock-in and total cost.
• The price list hides people’s time. For one database, the price lists say US$17.55 against US$53.58. The whole costs, with assumed hours, are US$77.55 against US$413.58.
• Choose part by part. Larkfield buys the model and builds the prompt, the test tickets and the release gate, because they hold its own knowledge.
• Tomás’s own model sent 0 of 24 holdout tickets that needed a person to a person and costs US$1,241 a month to host. The evidence says: keep buying the model.
• Recorded numbers are a snapshot of fast-changing models; the method lasts.

Next, you ask vendors the questions that show what a product really offers.

Next: Choose vendors and models as a leader