When Off-the-Shelf SaaS Beats a Custom AI Agent

Buy the off-the-shelf product when the job is generic and a vendor has already solved it for a thousand companies that look like yours. Build a custom agent when the workflow is specific to your business, touches data no vendor connects to, and is part of why customers pick you. Everything between those two cases is arithmetic, and the crossover sits lower than most buyers assume.

Here is that arithmetic in one line. Microsoft’s Microsoft 365 Copilot pricing page, fetched 30 August 2026, lists the Copilot Business add-on at $18 per user per month on an annual commitment. At 44 seats that is $9,504 a year, every year. An AI Workforce Sprint is $9,500 once, for one agent running on your own infrastructure. The two are not the same product. That is the point: they stop being comparable at a headcount you can reach with two hands twice over.

The rule, stated plainly

Buy when the workflow is generic, the vendor’s roadmap is close enough to your roadmap, and the worst outcome is wasted subscription spend. Build when the workflow is yours, the data sits in systems no vendor integrates with, or a published pricing cap will collide with your growth. Most businesses need both, and in that order.

Buying is not the risk-free option

The pitch for buying is that it removes delivery risk. It replaces it with a different risk, and that one is measurable.

Zylo’s 2026 SaaS Management Index, published 29 January 2026 from more than 40 million SaaS licenses and $75 billion in spend under management, found companies leaving 36% of their SaaS licenses unused. Median SaaS spend ran $9,455 per employee. Of the 218 IT leaders it surveyed, 78% reported unexpected charges from consumption-based or AI pricing models, and 61% said unplanned SaaS cost increases forced them to cut projects. Spend on AI-native applications rose 108% year over year.

That is the honest shape of the buy risk. The software does not fail. You pay for it whether or not anyone opens it, and the meter moves without anyone making a decision. Buying wrong is survivable and quiet. Building wrong is expensive and loud. Quiet failures are the ones that run for three years.

The crossover is a seat count, not a feeling

Model tokens are no longer the expensive part of building. OpenAI’s pricing page, fetched 30 August 2026, lists gpt-5-mini at $0.25 per million input tokens and $2.00 per million output, and gpt-4o-mini at $0.15 and $0.60. An internal agent handling a few thousand jobs a month spends less on models than a single seat of almost anything costs.

So the build cost is not compute. It is scoping, integration, and the upkeep after launch. Price those honestly, then compare them against seat price times headcount times the number of years you expect to still be doing this job. Three years is the fair horizon, because that is how long a subscription you forget about keeps billing.

Signal	Buy the SaaS	Build the agent
The workflow	Same as every company in your sector	Specific to how you win
The data	Already inside the vendor’s product	In your database, ledger, or ERP
Headcount	Few users, or growing unpredictably	Enough seats that annual license cost passes a one-time build
Vendor limits	You fit inside the published caps	You have hit a cap, a seat ceiling, or a tier change
Failure you can afford	Wasted subscription spend	Delivery risk on a written scope
Time to value	Days	Four to six weeks

Microsoft’s own page carries a worked example of the fourth row: the Copilot Business add-on is limited to up to 300 users and requires a separate qualifying license underneath it. A buy decision that expires at a known headcount is still a good decision, as long as you made it knowing the date.

Where a custom build actually pays

The Widening AI Value Gap, BCG’s September 2025 study of more than 1,250 firms worldwide, puts about 70% of the potential value from AI in core business functions: sales and marketing, manufacturing, supply chain, and pricing. Research and innovation together account for 15%.

Read that as a build map. The functions holding the value are the same functions where your process differs from your competitor’s, which is exactly where a vendor product has to work for everyone and therefore has to stay shallow. The same study reports that agents already account for 17% of total AI value in 2025 and are expected to reach 29% by 2028, with 51% of North American firms experimenting with or deploying them against 41% in Europe.

None of that says build. It says build in one place. If the job you are considering sits outside your core, in a function where you look like every other company in your sector, the vendor has already amortised the work across their customer base and you cannot win that trade.

The option most buyers skip

The question is rarely buy or build for the whole job. It is usually buy or build for one step of it.

Take a typical distributor. It has a decent CRM already, bought, working, and nobody wants to replace it. What it does not have is anything that reads the incoming inquiries, checks stock and pricing against the actual system of record, and writes the reply. That gap is four to six weeks of work, not a platform migration, and it makes the software they already pay for worth more per seat.

Test it this way. Name the step people actually hate. If a vendor sells that exact step, buy it this week. If the vendor sells the ten steps around it and calls the middle one an integration, that middle one is your build.

The failure both options share

Federal Reserve FEDS Notes, published 3 April 2026, reports 40.7% of the workforce using generative AI in some form as of November 2025, against 12% using it daily at work. Roughly two thirds of the people with access are not touching it in the run of a normal day.

A license nobody opens and an agent nobody trusts fail identically. Both were bought or built without a named owner, a named metric, and a date on which somebody decides whether it stays. That is a governance answer, not a procurement answer, and it applies the same way to an $18 seat and a $9,500 build.

What we do about it

We buy the generic layer and build the seams. Email, documents, spreadsheets, helpdesks, CRMs: all bought, none of it worth building. The connective work between them, the part that reads your actual numbers and acts on them, is where a tomte earns its price. A tomte is our word for one production agent with one defined job, scoped narrow enough that you can tell in a week whether it works.

One of ours reconciles what the ad platforms claim against what the ledger actually shows, because no dashboard on the market reads our ledger. We published what it found on the ad-spend truth engine page.

Before you price a build, price it properly: we broke the numbers down in what a custom AI agent costs in 2026. If you are not yet sure which job qualifies at all, start with what to automate first.

Send us the job and the software you already pay for. You get a written answer within one business day telling you to buy it, build it, or leave it alone, with the arithmetic shown. No meetings. Start async.