TCO of a Custom AI Agent vs Per-Seat SaaS: The 3-Year Math

Comparing the TCO of a custom AI agent vs per-seat SaaS over three years comes down to the shape of the curve: the agent front-loads the spend and then charges maintenance, while SaaS starts cheap and climbs every time the team grows. In a 15-20 user company the two models typically cross between month 14 and month 24 — below that point the licence wins, above it the owned agent does.

This is the calculator we run internally before recommending either option. It isn't a marketing comparison: assumptions are stated, third-party figures are dated, and the last section covers the cases where a custom agent does not pay off. We run our own businesses this way — we're our own first client — and we've made both calls more than once.

What actually belongs in each side of the TCO?

Total cost of ownership is every euro that leaves the company to keep that capability running, not just the vendor invoice. Half of it is missing from most comparisons clients show us: implementation, integrations, training, internal admin hours, and the cost of getting out if something changes.

These are the lines that belong in the table, on both sides:

• Entry cost. Building the agent, or signup plus onboarding for the SaaS. On the SaaS side this is usually a one-off implementation fee that doesn't surface until the final proposal.
• Recurring cost. Agent maintenance (we work with 10-20% per year of the build cost) or subscription per seat per month.
• Inference cost. Applies to the owned agent only: what the model calls cost. In SaaS it's baked into the price, or billed separately as "AI credits".
• Integrations. Connecting to your ERP, CRM and email. Both models spend here, and often spend the same.
• Internal hours. Who administers users, permissions, templates and data. It's the cost nobody budgets and the one that scales hardest with headcount.
• Exit cost. Migrating data, rewriting automations, retraining. A SaaS holding your data captive has a real exit cost even though it never appears on an invoice.

What do the numbers actually look like?

For custom agents we work with ranges we've been applying for two years: a single-process agent in a small or mid-sized company builds for €15,000-40,000 depending on integrations and volume, with annual maintenance running 10-20% of that figure. A basic chatbot — answering FAQs, taking no actions — lands at €1,500-3,000.

For SaaS you have to check the specific vendor's published list price, because the spread is enormous. Two reference points from the sales-tooling segment, checked in August 2026 and quoted in US dollars: HubSpot's Sales Hub Professional is listed at $90 per seat per month billed annually ($100 monthly), plus a one-time $1,500 onboarding fee; Salesforce Sales Cloud Enterprise is listed at $175 per user per month. These prices move, so use the list in force the day you run your numbers, not ours.

The inference cost of an owned agent usually surprises people by how small it is. Anthropic's published API rates in August 2026 are $5 per million input tokens and $25 per million output tokens on their high-end model, and $1 / $5 on the fast model. An agent handling a few thousand interactions a month rarely exceeds low double-digit euros monthly, unless it works with long documents. If a quote puts inference forward as the headline cost, something is badly sized.

Three-year comparison table (15-user company)

Stated assumptions: 15 users in year one, 20 by year three; one main process automated; SaaS at $90 per seat per month with a one-off onboarding fee; custom agent at €22,000 with 15% maintenance; currency treated at parity to keep the comparison clean. Figures are in euros and rounded.

That result doesn't mean SaaS is expensive: it means that at 15-20 seats, per-seat is no longer the cheap option. At 4 users the table flips completely — recurring drops to €4,320 a year while the agent still costs €22,000 to build. That's the crossover, and you should run it with your seat count before deciding anything.

Where does your crossover point fall?

The crossover is the month when cumulative SaaS spend overtakes cumulative agent spend. It's a simple division: agent build cost divided by monthly SaaS cost minus monthly agent cost. With the numbers above — €22,000 entry against roughly €1,350 a month in licences — the crossover lands around month 20; at 5 users it never crosses within a sensible horizon, and SaaS is the right answer there.

Three factors move that point hard:

1. Seat count and its growth. The dominant factor. A team doubling in 18 months pulls the crossover forward aggressively; a flat team pushes it back.
2. Process fit. If your process is exactly what the SaaS does out of the box, the licence almost always wins. Every customisation you need pulls the crossover closer.
3. Expected lifespan. If the capability is needed for 18 months, the agent rarely amortises. If it's needed for five years, it almost always does.

A fourth factor is less visible: SaaS raises prices and you don't control when. Double-digit renewal increases have been common across the sector in recent years. With an owned agent, the recurring cost is something you negotiate and it's proportional to what actually gets maintained.

What about the hybrid model?

The hybrid is what we end up recommending in most companies: SaaS for the standard, an owned agent for what differentiates you. Your CRM, your accounting and your email are commodities — buying those is correct; the specific process that makes your business money rarely ships in a licence.

In practice this means an agent that sits on top of your SaaS stack: it connects via API, runs the logic you need, and writes the result back into the system everyone already uses. You then pay seats only for the people who genuinely use the SaaS interface, not for everyone who needs the output — and that alone typically cuts licence count in half. We go deeper on the decision in our guide to build vs buy in AI.

What should NOT go into this calculation?

Hypothetical savings and invented opportunity costs. If your justification rests on "we'll save 30 hours a week" without having measured today's hours, you don't have a TCO calculation — you have a slide deck.

• Hours saved with no baseline. Measure the process as it stands first. If you can't measure it, you won't be able to prove the improvement either.
• Headcount savings nobody will act on. If there's no real intent to reduce staff — the usual case — the saving is in capacity, not payroll. Equally valid, but it doesn't book the same way.
• Comparing list price against negotiated price. If you're going to negotiate the SaaS, compare against the negotiated number; if not, against list. Mixing the two skews the result.
• Ignoring your own time. Internal admin hours are a real cost even when they never leave the bank account.

How to run your own numbers in an afternoon

1. Count the real seats today and estimate them 36 months out. Not the ones you'd like: the ones you'll pay for.
2. Ask for the full SaaS list price, including implementation, contract minimums and what falls outside the tier you're looking at.
3. Ask for an itemised agent quote: build, integrations, annual maintenance and an inference estimate. If maintenance isn't there, ask for it — we break it down in our guide to AI agent maintenance cost.
4. Build the 36-month table with both columns and run the cumulative month by month until you find the crossover.
5. Apply a 20% error margin to the agent column. Custom projects overrun; SaaS overruns too, just on the other axis (seat growth).

If the crossover lands inside the first 24 months, the custom agent is defensible with no further argument. If it lands beyond month 30, start with SaaS and revisit when the team grows. And if none of it matters because your process is unusual and no SaaS covers it, you already have your answer. The starting ranges are in our guide to how much a custom AI agent costs, and if you'd rather run the numbers together on your real figures, that's how we work through our AI consulting service.

Frequently asked questions

At how many users does a custom agent start to pay off?

With typical market ranges the threshold sits between 10 and 15 seats for a mid-tier SaaS, dropping to 5-8 if the SaaS is premium. Below that the licence almost always wins, unless your process doesn't fit what the SaaS does out of the box.

Can inference cost run away?

It can, which is why you bound it at design time: pick the right model per task, cache what repeats, and set spend limits. In a well-built internal-process agent, inference is a minor line next to maintenance. If it worries you, ask for the quote to include an estimated monthly ceiling at your real volume.

What happens if the agent vendor disappears?

That's the right question, and the answer lives in the contract: require that the code, the prompts and the integrations are yours and sit in a repository you can access. A properly delivered custom agent can be maintained by another team. If a vendor refuses to hand over the code, you're buying SaaS dressed up as development — the worst of both worlds.

Do I count licence savings as agent return?

Only if you're genuinely going to cancel those licences. It's common for an agent to reduce the seats needed while the company keeps the SaaS for other functions; in that case the saving is partial and should be counted partially. Be strict here — it's where internal business cases inflate the most.

How often should I redo this calculation?

Once a year, and any time seat count changes materially or the vendor raises rates. A TCO calculation is a snapshot built on assumptions; when the assumptions move, the decision can move with them.