Custom AI Agent vs Off-the-Shelf SaaS:A Decision Guide

Consider a founder weighing whether to pay us $10,000 to build a support agent or just turn on the one inside their help desk. The first thing I'd ask is how many conversations a month it would handle. Say the answer is about 400.

At 400 conversations a month, I'd tell them to turn on the one in the help desk. Our own price is the reason. A custom agent at our rates costs $100,000 over three years before a single model token is billed. The per-resolution SaaS they already had would cost a small fraction of that at their volume.

That's an odd thing for an agency to write in public, I know. But this article is the version of that conversation I'd want to read if I were on the other side of it: the real published prices, checked today (September 30, 2026), the arithmetic done out loud, and the handful of reasons that should push you to custom even when the arithmetic says otherwise.

If you want the broader framework first (build, buy, or build on a platform), I set that out in the build versus buy decision framework. This piece is narrower. It's one choice, between a subscription product and a commissioned agent, with numbers attached.

The belief most competent buyers hold goes like this: custom software is the premium option, SaaS is the budget option, so you pick based on budget. It sounds right. It's how people buy cars.

It doesn't hold for agents, bc the two options are not the same product at different price points. A SaaS agent is a very good answering machine attached to a knowledge base and a list of connectors somebody else chose. A custom agent is a piece of your operations that happens to talk. One is renting a furnished apartment; the other is building a house on your own lot. You don't decide between those on monthly cost alone. You decide on whether the apartment has the rooms you need.

So the order I use is: first, can the SaaS do the job at all? Second, is there a non-price reason it can't be the answer (compliance, ownership, a rule it can't express)? Only third, at your volume, which is cheaper over three years?

Why this order? Because the first two questions have yes or no answers that don't move, and the third has an answer that moves every time a vendor updates a pricing page. If you start with price, you end up re-deciding every six months. If you start with capability, price only settles the cases capability left open.

There's a second reason. Price comparisons between these two options are unusually easy to get wrong, bc the units don't match. The SaaS charges per resolution, per conversation, per credit, per call or per seat. Custom charges a fixed fee plus tokens. Comparing them means converting both to the same unit at the same volume, which is exactly what most comparison pages skip. I'll do that conversion in full below, but only after the question it answers has been earned.

Most small businesses stop at the first question with a yes, and they should buy. The rest of this article is for the ones who get a no, or a "sort of."

I checked every price below on the vendor's own page today. Where a vendor doesn't publish a number, I say so rather than fill the gap.

Intercom Fin. Intercom's pricing page lists Fin AI Agent at $0.99 per outcome, charged once per conversation when the customer confirms the answer resolved it, doesn't ask for more help, or a workflow completes (Intercom pricing). You also pay seats, from $29 on Essential to $132 on Expert, with a one full seat minimum. If you run Fin on top of another help desk there's a monthly minimum commitment. Two corporate changes matter here: the company renamed itself Fin in May 2026, and on June 15, 2026 Salesforce signed a definitive agreement to acquire it for about $3.6 billion, expected to close in the fourth quarter of Salesforce's fiscal 2027 (MarTech). I'll come back to why that matters.

Zendesk. Zendesk's pricing page says AI agents are included in Suite plans and are billed per automated resolution, with Suite Team at $55 and Suite Professional at $115 per agent per month on annual billing (Zendesk pricing). Its help center defines an automated resolution as a customer issue resolved without human help, verified by an LLM, and lists 5, 10 or 15 included resolutions per agent per month depending on tier (Zendesk Help). What it does not publish is the dollar price per resolution beyond that allowance. A third-party analysis puts committed packs around $1.20 to $1.50 with higher overage (eesel AI). I use $1.50 below as a scenario, not as Zendesk's rate. Get yours in writing.

Salesforce Agentforce. The Agentforce pricing page lists $2 per conversation for customer-facing agents, or Flex Credits at $500 per 100,000 credits with 20 credits per action (30 for voice), which is $0.10 an action (Agentforce pricing). Per-user add-ons start at $125 a month and Agentforce 1 editions at $550 per user a month. Note that $2 is per conversation, resolved or not. That's a different unit from Fin's, and it changes the maths a lot.

HubSpot Customer Agent. HubSpot's marketplace listing says resolving one text conversation uses 50 HubSpot Credits (HubSpot Marketplace). HubSpot's billing documentation lists pay-as-you-go overage at $0.010 per credit (HubSpot Knowledge Base). Fifty times one cent is $0.50 a resolution, the cheapest unit price in this table, once your included credits run out.

The smaller tools. Lindy sells per user: Plus at $29.99, Pro at $99.99 and Max at $199.99 a month, with Enterprise custom and listing a signed BAA (Lindy pricing). For phones, Upfirst runs from $24.95 for 30 calls to $299 for 600, with overage between $1.50 and $0.70 a call and BAAs only on its Custom plan (Upfirst pricing). Smith.ai's AI receptionist has a free plan with 25 calls, Pro at $150 and Enterprise at $500, working out to roughly $1.60 to $3.00 a call depending on plan and volume (Smith.ai pricing).

To be clear, none of those prices is the whole bill. Seats, add-ons, minimums and a help desk you may already pay for all sit around them. But they're the part that scales with volume, and volume is what decides the comparison.

Here's ours, in full, because a comparison with only one side priced isn't a comparison.

A Frenchy Digital agent starts at $5,000 for the agent plus $5,000 for setup, so $10,000 to start. That can go up with complexity and with the number of API endpoints and integrations. After launch there's a $2,500 monthly baseline. Model tokens, telephony and third-party APIs are billed to you directly, at provider rates, on accounts you control.

Out loud, then. Year one is $10,000 plus 12 times $2,500, which is $10,000 plus $30,000, or $40,000. Each later year is 12 times $2,500, or $30,000. Three years is $40,000 plus $30,000 plus $30,000, which is $100,000. All plus usage.

Spread over 36 months, that's $100,000 divided by 36, about $2,778 a month. That's the number every SaaS price below has to beat or lose to.

Now usage, which is an estimate and I'll show exactly how I built it. Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, with Haiku 4.5 at $1 and $5 (Claude pricing). Suppose a support conversation, including its system prompt, retrieved help articles, history and a couple of tool calls, sends about 20,000 input tokens across all its turns and gets back about 2,000 output tokens.

• Input: 20,000 tokens at $2 per million is $0.04.
• Output: 2,000 tokens at $10 per million is $0.02.
• Total: about $0.06 per conversation on Sonnet 5.5, before prompt caching, which lowers it.
• Per resolution: if the agent resolves, say, 60% of conversations (a scenario, not a benchmark), you pay usage on every conversation but only count the resolved ones, so $0.06 divided by 0.6 is $0.10 per resolution.

A voice agent adds phone minutes. Twilio lists inbound local calls at $0.0085 a minute and a local number at $1.15 a month (Twilio pricing), so a three-minute call is about 2.6 cents of carrier time. Speech recognition and synthesis sit on top of that, and I didn't verify those prices for this piece, so I don't fold them into a number.

The full breakdown of where agent money goes, including the parts people forget, is in what an AI agent really costs in 2026.

The break-even is where the SaaS bill for a month equals the custom bill for a month. The custom bill is a fixed $2,778 plus usage that grows with volume. The SaaS bill is a unit price times volume. So every unit saves you the gap between the SaaS price and your usage per unit, and you need enough units for those gaps to add up to $2,778.

Written as one line: break-even units equals $2,778 divided by (SaaS price per unit minus your usage per unit).

Against Fin. $0.99 minus $0.10 is $0.89. $2,778 divided by $0.89 is about 3,121. So over three years, owning only beats Fin above roughly 3,100 resolutions a month. If you treat the $10,000 setup as already spent and look only at the running rate, it's $2,500 divided by $0.89, about 2,809.

A correction, then. In our earlier AI agent cost article, I wrote that owning beats Fin only above about 2,679 resolutions a month. Fin's price hasn't changed; it was $0.99 then and is $0.99 today. The error was mine: I treated usage as a fixed $152 a month instead of letting it grow with volume, and I left the $10,000 setup out. Redone properly, the run-rate figure is about 2,800 and the three-year figure is about 3,100. The conclusion (below that, rent) stands. The number moves up.

Against Agentforce. Here the unit is a conversation, not a resolution, so the usage figure is $0.06. $2 minus $0.06 is $1.94. $2,778 divided by $1.94 is about 1,432 conversations a month. That's less than half the Fin figure, bc Agentforce charges twice as much per unit and charges for every conversation.

Against HubSpot. $0.50 minus $0.10 is $0.40. $2,778 divided by $0.40 is about 6,944 resolutions a month. HubSpot's unit price is low enough that a custom agent almost never wins on price alone for a HubSpot shop.

Against Zendesk. Using the $1.50 third-party scenario, the gap is $1.40, and $2,778 divided by $1.40 is about 1,984. If your negotiated rate is lower, the number rises. Plug in your own quote.

Against a receptionist tool.Take Upfirst Scale: $299 covers 600 calls, then $0.70 each. If a custom voice agent cost as little as $0.10 a call in usage (an assumption I haven't verified, since I didn't price speech), the break-even is where $299 plus $0.70 times (calls minus 600) equals $2,778 plus $0.10 times calls. That solves to about 4,831 calls a month. Most small businesses take nowhere near that. For a standard answering job, buy the receptionist.

And the same numbers as three-year totals, so you can see the curves cross. Fin is priced on resolutions; Agentforce on conversations, assuming the same 60% resolution scenario; custom is $100,000 plus the usage estimate. Seats are left out of all three, since you'd need a help desk either way.

At 500 resolutions a month, Fin costs about $17,820 over three years ($0.99 times 500 times 36) and custom costs about $101,800 ($100,000 plus 833 conversations times $0.06 times 36). That's a factor of more than five. At 6,000 resolutions a month, the picture flips: about $213,840 on Fin against about $121,600 custom.

How sensitive is this to my usage guess?Fairly, and it's worth seeing why. Suppose my token estimate is three times too low and a resolution really costs you $0.30 in usage. Against Fin the gap shrinks from $0.89 to $0.69, and $2,778 divided by $0.69 is about 4,026. Against HubSpot the gap falls to $0.20 and the break-even doubles to about 13,890, which is to say never, for almost any small business. Against Agentforce, tripling $0.06 to $0.18 per conversation leaves a gap of $1.82 and a break-even of about 1,526. Per-conversation pricing is the most forgiving of a bad usage estimate, bc the unit price is so much larger than the usage.

Now the other direction. With prompt caching, or with a smaller model like Haiku 4.5 at $1 and $5 per million tokens handling the simple turns, usage could come in well under my $0.06. At zero usage (the theoretical floor) the Fin break-even is $2,778 divided by $0.99, about 2,806. So even with free tokens, you need somewhere around 2,800 resolutions a month before owning beats Fin over three years. That floor is the most useful number in this article for a Fin customer. Below it, no amount of clever engineering makes custom cheaper.

And the resolution share matters as much as the tokens. If your agent resolves only 40% of conversations instead of 60%, usage per resolution rises from $0.10 to $0.15, and the Fin break-even rises to about 3,307. If it resolves 80%, usage per resolution drops to about $0.075 and the break-even falls to about 3,036. The honest summary: somewhere between 2,800 and 4,000 resolutions a month, depending on things you'll only know after running an agent for a while.

One more thing the arithmetic hides. The SaaS unit price only charges you for success (Fin per outcome, HubSpot per resolution, Zendesk per verified resolution). Your custom usage charges you for every conversation, including the ones that fail and go to a person. That's a quiet advantage for the per-resolution products, and it's why I divide usage by the resolution share instead of comparing per-conversation costs directly. Agentforce's per-conversation price doesn't have that advantage, which is a large part of why it loses at a much lower volume.

Most of the custom builds I'd actually recommend are decided here, not in the table above. Here are the seven that come up.

Security is its own subject, and it applies to both options: any agent reading untrusted input is exposed to prompt injection, which nobody has solved. I cover how to limit the blast radius in our companion guide to ten security controls before you connect your data, linked at the end of this article.

Most small businesses with standard needs should buy.Here's what standard looks like.

• The agent mainly answers questions from content you already have: a help center, policies, product pages.
• The actions it takes are ones a native connector already does: look up an order in Shopify, book into a common calendar, create a ticket, capture a lead.
• Volume is under about 1,500 resolutions a month on Fin or HubSpot, or under a few thousand calls on a receptionist tool.
• Your buyers or regulators require a vendor SOC 2 report.
• You need it live in weeks, not months.
• You don't want to manage a technical relationship.

Consider a dental practice taking 600 calls a month that wants calls answered, appointments booked in a common scheduling system, and messages taken after hours. A receptionist product priced per call does that for a few hundred dollars a month. Our baseline alone is $2,500. I'd tell that practice to buy, and I'd tell it to trial two or three before signing.

Or consider an online shop answering 1,000 "where is my order" chats a month. On Fin that's about $990 a month plus seats, provided its connectors reach your store platform (check that in the trial). Custom would cost nearly three times as much for the same job. Buy.

If you're in this bucket, our comparison of the top AI agent platforms for small business, linked below, is a better next read than anything we sell.

Custom earns its price in four situations, and usually you need two of them together.

1. 1.The agent has to write to your own systems in ways no connector supports, and those writes are the point of the agent.
2. 2.Your rules are specific enough that a settings page can't express them, and getting them wrong is expensive.
3. 3.Volume is above the break-even for the SaaS you'd otherwise use, and growing.
4. 4.Owning the code and the data matters, for portability, for a future sale, or for policy.

The one real agent build I can point to is Beyond Points AI. The case study describes an agent that searches award and cash travel inventory, transfers points between programs and completes bookings through a browser agent, with a person confirming every irreversible step (transfer, book, pay, cancel). No SaaS support agent does that, bc the job is writing to loyalty programs and booking systems on a user's behalf, behind the product's own confirmation rules. That's situations one and two together.

Consider, as a scenario, a regional service company taking 5,000 support conversations a month where most questions need a lookup in an in-house dispatch system with no public connector. On Agentforce at $2 a conversation that's $10,000 a month before anyone writes the dispatch integration. Custom at about $2,778 a month plus roughly $300 in usage is cheaper and does the actual job. That's the shape of a good custom case.

If that sounds like you, the service is described on our AI agent creation page, and the packaged starting points are on the AI agents page, including a customer support agent that competes directly with the products priced above.

This isn't a binary, and the best answer for a lot of mid-sized businesses is both.

Keep a SaaS agent for the common questions it handles well and cheaply. Then build a small custom piece for the one workflow it can't do: an action that writes to your in-house system, a rules engine that decides refunds, a check that runs before anything irreversible. Where the platform can call an external API (confirm this with the vendor, as it varies by plan), the custom piece can sit behind the SaaS front end without replacing it.

The advantage is obvious: you pay per resolution for the easy questions and own the hard part. The downside, which I'd rather say before you find it: you now have two things to maintain, two vendors to manage, and a seam between them that can break. If the SaaS changes how its actions work, your custom piece needs updating.

Consider, as a scenario, an insurance brokerage already on Zendesk. Its AI agent answers policy questions from the help center well. What it can't do is start a certificate of insurance request in the brokerage's agency management system, which has no connector the help desk supports. A hybrid keeps the Zendesk agent for questions and adds one custom action that validates the request, creates it in the management system and hands the certificate to a person to approve before it goes out. The brokerage pays Zendesk for the questions and owns the one piece that makes the agent useful to its business.

How is that priced? The custom action is still a build, so our $10,000 start and monthly baseline apply if we do it, and the scope is smaller than a whole agent. Whether that's worth it depends on what the workflow is worth to you, which is a question the break-even table can't answer. Ask it of the workflow, not of the agent.

There's also a sequencing version of the hybrid, which I recommend more than any other. Start on SaaS. Run it for three to six months. Export the logs. You'll then know your real volume, your real resolution share and exactly which questions it fails on. That's the data a custom build should be scoped against, and it makes the break-even arithmetic above use your numbers instead of my scenarios.

Go down the left column. If most answers land in the middle, buy. If two or more land on the right, and one of them is about writes or rules, custom is worth a proper look.

A shortcut for the volume row: take your monthly resolutions, multiply by your SaaS unit price, and compare with $2,778. If the SaaS number is lower, price is on the SaaS side.

On the SaaS side:

• No published unit price, and a sales rep who won't put one in writing before you sign.
• A resolution definition you can't audit. If the vendor decides what counts as resolved, ask to see how.
• No conversation export, or an export that leaves out the tuning and workflows you built.
• A resolution or deflection rate quoted as proof. Every one I've seen is vendor-reported.
• A minimum commitment far above your actual volume.

On the custom side:

• An agency that won't tell you when the SaaS is the better fit.
• No written statement of who owns the code, the prompts and the data at the end.
• Model and telephony usage billed through the agency with a markup instead of on your own accounts.
• A cost range with no method behind it.
• Claims of SOC 2 or other certifications that the vendor can't show you a report for.
• No plan for who maintains the agent after launch, or what that costs each month.

There's a longer list of agency warning signs in our AI agent developer red flags guide.

Articles on this topic usually lean on a few figures. They're not here, and here's why.

HubSpot's listing says its agent can resolve 70% of conversations, and press coverage of the acquisition repeats a claim that Fin closes about 76% of requests on its own. Both come from the seller. They might be right for some customers. They aren't a neutral measurement of what either agent will do for you, so I don't use them in the arithmetic. My 60% resolution share is labeled as a scenario for that reason.

Zendesk's per-resolution dollar price: Zendesk doesn't publish it, so I cite the third-party figure as a third-party figure and tell you to get your own quote.

The familiar lines about AI cutting support costs by 30%, or about most AI projects failing (the MIT 95%, the 85%, the 87%, Gartner's 40% cancellations) don't trace to measurements that support how they're used. None of them helps you decide this question anyway. Your volume and your systems do.

• Usage per conversation is an estimate built from Anthropic list prices and an assumed 20,000 input and 2,000 output tokens. Your conversations may be shorter or much longer.
• The 60% resolution share is a scenario, not a benchmark. Your share changes every break-even figure that uses resolutions.
• Speech recognition and synthesis prices for voice agents were not verified, so voice break-evens are rough.
• Zendesk's per-resolution price is not published; the figure used is third-party.
• Seats, add-ons, minimums and negotiated discounts are left out of the break-even to keep it comparable. They move real bills.
• Our own price can rise with complexity and number of integrations, which raises every break-even figure.
• Prices were checked on September 30, 2026. Pricing pages change without notice, and Fin is mid-acquisition.

1. 1.Pull last month's conversation count from your help desk or phone system. Multiply by the unit price of the SaaS you'd use and compare with $2,778. That one number tells you which side price is on.
2. 2.Write one page listing what the agent must read, what it must write, and where. Circle every write that no native connector covers.
3. 3.If nothing is circled and the SaaS number is lower, start a trial this week. If something is circled, get a written quote for both options against the same page.

That's it. Most people will end up buying, and that's the right answer for most people.