General AI chat subscriptions like Microsoft 365 Copilot charge $18 to $32 per seat, per month, so the bill grows every time you hire. A custom build, a single-workflow agent or an internal copilot, runs $3,000 to $12,000 to set up and $120 to $600 a month regardless of headcount. Below roughly 15 to 20 seats, buying is usually cheaper. Above that, the per-seat math starts to lose.
What "off-the-shelf" actually means in 2026
Off-the-shelf no longer means one generic chatbot. It means a per-seat subscription to a general assistant. Microsoft 365 Copilot, the market's most visible example, is sold as an add-on starting at $18 a user a month billed annually, on top of a qualifying Microsoft 365 license you already pay for separately, or bundled with a Business Premium plan for $32 a user a month, according to Microsoft's own Copilot pricing page. Every seat gets the same thing: a chat window that drafts, summarizes, and answers questions from whatever it already knows, plus whatever document you paste in.
That is genuinely useful for a wide range of everyday work. It is not the same thing as a tool that knows your pricing sheet, checks your CRM, or takes an action on its own. Off-the-shelf tools read and write text. They do not, by default, do anything with your systems.
What a custom build actually means
A custom build in this context is not a from-scratch AI lab. For a small or mid-size business it is almost always one of two things: a single-workflow agent that handles one job end to end (qualify a lead, triage a ticket, draft a proposal from a template), or an internal copilot wired into the tools a team already has open, Slack, a CRM, a ticketing system, that answers and acts using your actual data.
Both are scoped, contained builds, not open-ended R&D. A single-workflow agent typically costs $3,000 to $12,000 to design and connect, then $150 to $600 a month to run, the range we walk through in our AI agent cost guide. An internal copilot for a 15 to 40 person team runs $4,000 to $10,000 to set up and $120 to $500 a month after that. Neither number moves with how many people use it. It moves with how many systems it touches and how often it runs.
The real cost line: per seat vs per workflow
This is the part most comparisons skip: the two pricing models scale on different axes entirely.
| Off-the-shelf (per seat) | Custom build | |
|---|---|---|
| Setup cost | $0 to a few hundred | $3,000 to $12,000 |
| Ongoing cost | $18 to $32 per seat, monthly | $120 to $600 flat, monthly |
| Scales with | Headcount | Usage, not headcount |
Run the numbers for a 20-person team on the Copilot add-on tier at $18 a seat: $360 a month, $4,320 a year, forever, and the number only goes up as the team grows. Run the same team against a single-workflow custom agent at the midpoint of its range, $7,500 to set up and $375 a month to run: year one costs about $12,000, but year two and beyond cost $4,500 a year flat, whether the team is 20 people or 40. The crossover happens fast, right around 20 seats, once the AI is doing one repeatable job rather than general assistance.
Where buying off-the-shelf is still the right call
Under about 15 seats, or when the job is genuinely general (drafting emails, summarizing meetings, brainstorming copy), buying wins outright. There is no setup cost, no integration work, and no waiting three to four weeks for a build. A five-person team paying $90 a month for a Copilot add-on is not going to beat that with a custom project, and it should not try.
Off-the-shelf also wins when the task has no fixed shape. A general assistant is good at open-ended requests that change every time. A custom build is good at the same request, over and over, with the same inputs and the same rules. If the request changes shape constantly, a rigid custom workflow will fight the job rather than help it.
Where a custom build pays for itself faster
The math flips once three things are true at the same time: the team is large enough that per-seat cost has real weight, the job is the same job every time, and the answer depends on data a generic assistant does not have, your CRM records, your ticket history, your own documents. A generic chatbot can talk about your business in the abstract. It cannot pull the actual status of the Morrison account or check whether a specific customer already asked this question last week.
That gap shows up in the numbers. McKinsey's November 2025 State of AI survey found adoption is now close to universal, yet only 39% of organizations report AI's impact reaching the enterprise bottom line. The pattern behind that gap is familiar: broad, general tool access rolled out to everyone, with no specific workflow it was built to change. A scoped build tied to one measurable job (fewer manual lookups, faster ticket triage, more qualified leads reaching a rep) is what turns AI spend into a number a business can actually track.
The hybrid setup most businesses land on
Almost no business we work with picks one model exclusively. The workable pattern is a small number of general seats for drafting and research, a subscription like Copilot for the people who actually need it, paired with one or two scoped custom builds for the specific jobs that repeat every day and touch real data: a lead-qualification agent, a support copilot that reads the ticket history before it answers.
That split keeps the per-seat bill small (most teams do not need every employee on an $18-a-seat plan) while putting the custom build budget where it actually earns it back: the one or two workflows that run constantly and would otherwise eat someone's afternoon.
A short checklist before you decide
Ask these before signing a subscription or scoping a build:
- How many people actually need this daily, not "might find it useful"? Under 15, lean toward buying.
- Does the job repeat in the same shape every time, or does it change with every request? Repeatable favors building.
- Does the answer depend on your own data, a CRM record, a ticket, a document, rather than general knowledge? If yes, a generic tool cannot do this well regardless of price.
- Can you name the one metric this should move (fewer manual lookups, faster response time, more qualified leads)? If not, a subscription is the lower-risk starting point until you can.
Our AI Strategy & Discovery sprint is built for exactly this decision: a two-week look at where AI actually pays off in your business, with a ranked recommendation instead of a guess. If the answer comes back "buy," we will tell you that too. Get in touch and we will scope it against your actual team size and workflow.
Frequently asked questions
Is a custom AI agent always more expensive than a subscription?
No. Below roughly 15 to 20 users, or for general tasks like drafting and summarizing, a per-seat subscription at $18 to $32 a month is almost always cheaper than a $3,000-plus custom build. The crossover favors building once headcount grows or the task is the same repeatable job every time, not before.
Can I start with an off-the-shelf tool and move to custom later?
Yes, and it is the most common path. Most businesses start every employee on a general assistant, then identify the one or two workflows that run constantly enough to justify a scoped build. Starting broad and narrowing to a custom agent later avoids overbuilding before you know which job actually needs it.
What makes a workflow a good candidate for a custom build instead of a subscription?
Look for a job that repeats in the same shape daily or weekly, depends on your own data (a CRM, a ticket system, internal documents), and currently costs a person real time to do by hand. Lead qualification, ticket triage, and status lookups across systems are common examples. A task that changes every time is a poor fit.
Does Microsoft 365 Copilot require an existing Microsoft license?
Yes. Copilot is sold as an add-on starting at $18 a user a month billed annually, layered on top of a qualifying Microsoft 365 plan you already pay for separately, such as Business Standard, Business Premium, E3, or E5. The advertised add-on price is not the full cost if you do not already hold that license.
How long does a custom AI build actually take?
A single-workflow agent or internal copilot typically takes three to four weeks from discovery to a working pilot: one week to scope the job and the data it needs, two weeks to build and test against real requests, and a final week piloting with a small group before full rollout. That is faster than most teams expect and far faster than an enterprise-scale AI project.
Do custom builds use the same AI models as general chat subscriptions?
Custom builds run on leading frontier models, the same class of production-grade language models behind consumer tools, configured against your own data and connected accounts rather than a generic public interface. The difference is not the underlying model, it is what the system is allowed to see and do with your information.