A mid-market support leader I spoke with last quarter showed me her renewal quote from one of the big AI customer service platforms. Year one had been $480,000. Year three's projected number was $1.4 million. The volume hadn't tripled. The price had. Her CFO wanted to know why a tool that "automates customer service" was now the third-largest line item in the operations budget.
This is the question every buyer of AI automation eventually has to answer, and almost no vendor will help you answer it honestly: what does this actually cost over three years, and what do I own at the end?
The market has settled into three distinct buying models, and they have wildly different cost trajectories. If you only compare the first-year price, you will pick the wrong one. Here is the math nobody publishes.
The Three Models You Are Actually Choosing Between
When a CFO asks "what are our options for AI customer service," the answers fall into three categories. Vendors blur the lines on purpose, but the underlying economics are very different.
Model A: Consumption-Based Platforms. Sierra, Ada, Decagon, Intercom Fin, and most of the standalone AI agent vendors. You pay per conversation, per resolution, or per interaction. Costs scale linearly with volume - or worse, because tiered pricing often punishes growth. Annual contracts typically carry minimums of $30K to $300K or more. You are renting access. Stop paying and the agent disappears.
Model B: Ecosystem-Native AI. Salesforce Agentforce, Zendesk AI, Microsoft Dynamics Copilot, HubSpot Breeze. The AI is bundled into a platform you already use. Pricing flows through per-user licenses, AI credit packs, or new tiered plans. Integration with the parent platform is deep; integration with anything else is limited. You are extending a system you already pay for, with new cost layers on top.
Model C: Custom-Built AI. A flat monthly retainer for design, build, and ongoing optimization. The agent is built around your specific workflows, deployed in infrastructure you control, and integrated with whatever stack you already run. At the end of the engagement, the software is yours. This is the model OpenNash uses, and it is also how teams like Klarna and LinkedIn approach their internal agent builds, as documented in LangChain's enterprise case studies.
The mistake most buyers make is comparing Model A's monthly bill to Model C's monthly retainer, deciding the platform is "cheaper," and never running the actual three-year math. Let's do that now.
The 3-Year TCO Breakdown at 30K Interactions Per Month
Here is the scenario: a mid-market company handling 30,000 AI-managed customer interactions per month. Not a small SMB. Not a Fortune 100. The kind of business that has real volume, real complexity, and a CFO who reads contracts.
Model A: Consumption-Based Platform
| Cost Component | Monthly | Annual | 3-Year |
|---|---|---|---|
| Per-interaction fee ($1.50 avg) | $45,000 | $540,000 | $1,620,000 |
| Platform license fee | $5,000 | $60,000 | $180,000 |
| Implementation (Year 1) | - | $50,000 | $50,000 |
| Total | $1,850,000 |
That $1.50 average per interaction is not aggressive. Reinventing.ai's 2026 SMB pricing analysis puts most enterprise AI agent platforms in a $0.99 to $2.50 range per resolved conversation, with successful resolutions often priced higher. If volume grows to 50K per month by Year 3 - a reasonable assumption for a growing business - the three-year total approaches $2.5M.
Model B: Ecosystem-Native (Salesforce Agentforce)
| Cost Component | Monthly | Annual | 3-Year |
|---|---|---|---|
| AI conversations ($2/conversation) | $60,000 | $720,000 | $2,160,000 |
| Underlying platform licenses (20 users) | $8,000 | $96,000 | $288,000 |
| Implementation and admin | $6,250 | $75,000 | $150,000 |
| Total | $2,598,000 |
Agentforce is more expensive than standalone platforms because you are paying for the AI layer plus the underlying CRM seats it requires. Salesforce's Flex Credit model can soften the per-conversation cost, but only by adding contract complexity and consumption forecasting that most operations teams aren't equipped to manage.
Model C: Custom-Built (OpenNash Retainer)
| Cost Component | Monthly | Annual | 3-Year |
|---|---|---|---|
| Build and optimization retainer | $7,995 | $95,940 | $287,820 |
| Infrastructure (your environment) | $500 | $6,000 | $18,000 |
| Total | $305,820 |
At 30K interactions per month, custom-built costs roughly 12-15% of consumption platforms over three years. The delta only widens as volume grows, because the marginal cost of an additional interaction in a custom build is essentially the LLM API call - cents, not dollars. And at the end of the engagement, you keep the software.
The AlphaCorp 2026 pricing guide puts mid-complexity custom builds in the $50K-$180K range for one-time development, which lines up with the retainer math when you amortize across three years of ongoing improvements.
Where the Crossover Actually Sits
Custom is not always cheaper. The honest version of this analysis names the volume range where each model wins.
Below 2,000 interactions per month, consumption platforms are usually the better financial choice. Total spend is small enough that the upfront engineering investment in a custom build doesn't amortize meaningfully. You are also probably moving fast and iterating on what your support function even looks like. Pay-per-use makes sense.
Between 2,000 and 10,000 interactions per month is the contested zone. The math depends heavily on your per-interaction rate, your platform's overage policies, and how complex your workflows are. A simple FAQ deflection bot at 5K conversations per month might still be cheaper on a platform. A multi-system workflow that touches CRM, billing, and inventory at the same volume usually isn't.
Above 10,000 interactions per month, custom builds win on a three-year basis in nearly every scenario I have seen. The variable cost of platform pricing simply outruns the fixed cost of a build. By 30K per month you are not in the same conversation. By 100K per month you are looking at platform bills that exceed the cost of an in-house engineering team.
Aisera's build-vs-buy analysis lands on a similar crossover, noting that organizations with predictable, high-volume use cases consistently get better unit economics from owned systems than from per-conversation pricing.
The kicker most cost models miss: even below the crossover, custom gives you an asset. Platform spending is rent. Custom spending builds equity in software your business owns.
Costs Nobody Includes But Should
Both sides of this debate have hidden costs that don't show up in the sales pitch. Honest comparison requires naming them.
Hidden costs of platforms
- Volume overages during demand spikes. Holiday seasons, product launches, PR crises - exactly the moments when you need the system to work, and exactly the moments when consumption pricing punishes you. Some contracts cap overages; many don't.
- Escalation charges. When the AI can't resolve a conversation and hands off to a human, several platforms still charge for the AI portion of the interaction. You pay for the failure.
- Integration costs. Connecting the platform to your other systems - CRM, billing, knowledge base, internal tools - is usually a separate professional services line item, often $25K-$100K depending on complexity.
- Migration costs. If you switch vendors in Year 3, you start from zero. No data portability, no workflow portability, no model improvements come with you. The Quickchat AI 2026 buyer's guide flags this as the single most underestimated cost in platform selection.
- Roadmap dependency. Your strategy is constrained by what the vendor decides to build next. If your business needs a feature their roadmap doesn't include, you wait or work around it.
Hidden costs of custom
I will not pretend custom builds are free of tradeoffs. The honest version:
- Longer time-to-value. Platforms can be live in days. A well-built custom agent takes weeks for an MVP and months to reach production-grade reliability. If you need automation running by next Tuesday, custom is not the answer.
- Partner dependency. A custom build requires a partner who actually understands your business. That relationship has to work, and the partner has to stay engaged for ongoing optimization. This is real risk.
- Ongoing maintenance is not free. LLMs change. APIs change. Your business changes. Someone has to keep the agent current. The retainer model handles this, but it is a real cost to budget.
- Less plug-and-play. Turning on a platform feature is a checkbox. Building a new capability in a custom system requires actual engineering work.
The honest framing: platforms trade long-term cost for short-term speed. Custom trades short-term speed for long-term cost advantage and ownership. Neither is universally better. The question is which tradeoff matches your business.
What Doesn't Show Up in a Spreadsheet
Cost is the easiest part of this decision to model and the least interesting. The real strategic differences are harder to quantify.
Ownership as exit value. Custom AI software is an asset on your balance sheet. Platform subscriptions are an operating expense. If you sell the company, the acquirer pays for what you own; what you rent disappears at renewal. This matters more than most operators realize until they are in due diligence.
Competitive differentiation. Every company in your industry can buy the same Sierra or Ada deployment. Nobody else has software built specifically for your operations, your data, your edge cases. This is the same logic that drove early SaaS buyers to eventually build custom internal tools - and it applies harder to AI, where the system is making decisions on your behalf.
Vendor independence. Platform pricing changes. Vendors get acquired. Roadmaps shift. As Simon Willison has documented in his work on agent architectures, the underlying primitives of agent systems - tool use, retrieval, planning loops - are increasingly commodified. Paying enterprise pricing for primitives that are becoming cheap is a strategic error.
Flexibility across use cases. A platform built for customer support doesn't extend to internal operations, compliance review, scheduling, or voice. Each new use case means another vendor, another contract, another integration. A custom architecture extends to new use cases at marginal cost.
A Decision Matrix You Can Actually Use
If you read nothing else, use this:
| Your Situation | Best Model |
|---|---|
| Low volume (under 2K interactions/month), need to launch fast | Consumption Platform |
| Already deep in Salesforce/Zendesk, basic AI needs | Ecosystem-Native |
| High volume (over 10K interactions/month), cost-sensitive | Custom-Built |
| Complex workflows spanning multiple systems | Custom-Built |
| Regulated industry with specific compliance needs | Custom-Built |
| Building IP or planning a future acquisition | Custom-Built |
| Need automation beyond customer support | Custom-Built |
| Small team, no engineering, simple use case | Consumption Platform |
The pattern is clear once you list it out. Platforms win on speed and simplicity. Custom wins on volume, complexity, ownership, and strategic flexibility. Most mid-market and up companies eventually land in the custom column - the only question is whether they spend three years funding a vendor's growth before they figure that out.
Don't start the conversation with "which platform should I buy?" Start with "Do I want to rent automation or own it?" If you handle significant volume, your workflows are specific to your business, and ownership matters to your long-term strategy, the math almost always points to custom.
Platforms know this. It is why they don't publish three-year TCO comparisons. We just did.
Want to see what custom AI agents would look like for your business? Talk to us.