Every major platform now ships an AI agent product. Salesforce has Agentforce, Microsoft has Copilot agents, HubSpot, Zendesk and Intercom all have their own. The pitch is consistent and genuinely appealing: your data is already here, so the agent is a configuration exercise rather than a project.
Sometimes that's exactly right. Sometimes it produces a bill that scales with your success and a system you can't change.
This is the comparison without a thumb on the scale. We build custom agents, so read the recommendation with that in mind — but the honest answer is that the platform wins a substantial share of cases, and knowing which ones saves you from an expensive mistake in either direction.
The structural difference
| Platform agent (Agentforce et al.) | Custom build | |
|---|---|---|
| Time to first working version | Days | Weeks |
| Pricing model | Consumption — per conversation or action | Infrastructure + model tokens |
| Cost as volume grows | Scales linearly, roughly forever | Flattens; mostly fixed |
| Model choice | Whatever the vendor supports | Any provider, switchable |
| Data location | Vendor's environment | Wherever you put it |
| Customising behaviour | Within configuration limits | Unbounded |
| Channel reach | Vendor's supported channels | Any channel with an API |
| Who maintains it | Vendor | You or your partner |
| If you leave | Rebuild from scratch | Portable |
| Compliance posture | Vendor's certifications | Your responsibility |
Two rows decide most real decisions: how cost behaves as volume grows, and whether you already live inside that ecosystem.
Where consumption pricing bites
Platform agents are typically priced per conversation or per action. The precise figures change often enough that quoting them is a disservice — check the current rate card. What matters is the shape, and the shape is consistent across vendors: you pay per unit of usage, indefinitely.
At low volume this is clearly correct. A few hundred conversations a month costs less than any custom build, and it works immediately.
The maths inverts somewhere, and the inversion point is lower than people expect. Run your own numbers: take your realistic monthly conversation volume, multiply by the vendor's per-conversation rate, annualise it, and compare against a one-off build plus hosting and token costs. For a business handling meaningful daily volume, the payback period on a custom build is often measured in months rather than years.
Two adjustments that make the comparison honest, both of which favour the platform more than people assume:
Count all conversations, not just useful ones. Consumption billing usually includes the ones that go nowhere. If a third of your inbound is spam or wrong numbers, you're paying for that third.
Include maintenance on the custom side. A custom agent is software. It needs updates, monitoring, and someone to fix it at 2am. Comparing a build cost against a subscription without accounting for ongoing ownership is how custom projects get approved and then resented.
Where the platform genuinely wins
You're already deeply invested in it. If your entire sales operation runs in Salesforce, your reps live in it, and your data model is mature there, the integration advantage is real and large. A custom agent would need to replicate connections you already have.
You need it working this month. Configuration beats construction on timeline, always.
You have no engineering capacity. A custom agent you can't maintain is a liability with a countdown on it. This is the most common reason custom builds fail, and it has nothing to do with the code.
Compliance is someone else's problem. Enterprise vendors carry certifications — SOC 2, HIPAA, regional data residency — that are expensive and slow to obtain independently. If your procurement process requires them, that alone can decide it.
Volume is low or unpredictable. Consumption pricing is genuinely efficient when you're not sure what you need yet.
Where a custom build wins
Your volume is high and growing. This is the clearest case. Predictable high volume is exactly where fixed costs beat per-unit costs.
Your channel isn't well served. Most platform agents are built around web chat, email, and their own ecosystem. WhatsApp is frequently a second-class integration or absent entirely. If WhatsApp is your primary sales channel — as it is across much of South Asia, the Middle East, and Latin America — the platform may simply not do what you need.
Your process is genuinely unusual. Configuration handles common patterns well. If your qualification logic involves your own scoring model, or a custom pricing engine, or an inventory system nobody has heard of, you'll spend longer fighting the configuration limits than you would writing it directly.
Data residency or privacy requires it. Some obligations can only be met by controlling where the data physically lives.
You want to control the model. Model quality, pricing, and capability shift constantly. Being able to switch providers — or run different models for different tasks — is worth real money over a multi-year horizon.
The middle path most people miss
Build-versus-buy is a false binary, and the pragmatic answer is frequently neither extreme.
Use the platform for what it's genuinely good at — CRM, records, reporting, the workflows your team already knows — and build a thin custom agent on your high-volume channel that writes into it through the API. You get channel control and predictable costs where volume lives, without rebuilding a CRM you already have and like.
That's the architecture we most often end up recommending, and it's the one that survives contact with reality. A WhatsApp agent that qualifies leads and pushes structured records into an existing CRM is a few weeks of work, doesn't touch the systems your team depends on, and stops the per-conversation meter running on your highest-volume channel.
How to actually decide
Four questions, in order:
- What's your realistic monthly conversation volume in twelve months? Not today's — the number you're planning for. Multiply by the vendor's per-unit rate and annualise.
- Is your primary channel a first-class citizen on the platform? If it's WhatsApp, verify this carefully rather than assuming.
- Can you maintain software? In-house or through a partner you trust. If the honest answer is no, take the platform and don't look back.
- Does anything hard-require data control? Regulation, contract, or procurement. If yes, that decides it.
If volume is high, your channel is poorly served, and you can maintain software — build. If any of those is false, buy, and revisit in a year.
The genuinely bad outcome is choosing on ideology rather than arithmetic: paying consumption pricing at a volume where it's indefensible, or maintaining custom software you don't have the capacity to own.
Want the arithmetic run properly?
We build custom AI agents, usually alongside an existing CRM rather than replacing it. If you want a straight assessment of which side of this line you're on — including when the answer is "stay on the platform" — talk to us about AI chatbot development.