When you should hire us, and when you shouldn’t
Every comparison page you’ve read concludes that the company who wrote it is the answer. This one doesn’t.
Straight answer: sometimes you shouldn’t hire us. If a platform already serves your workflow, buy it; if you have engineers with agent experience, build in-house. This page maps all four options honestly — including the conditions under which each one beats us.
Last reviewed August 2026 · Sources cited inline
Your four real options
| Option | Best when | Worst when | Typical cost |
|---|---|---|---|
| Build in-house | You have engineers with agent experience and spare capacity | Your team has never shipped an eval suite | Salary + opportunity cost |
| Horizontal platform | Your workflow is standard: support deflection, scheduling, CRM | Your workflow is specific to your industry | $0.99–$2.00/resolution + $39–$149/seat/mo (Intercom, 2026; Zendesk, 2026) |
| Vertical product | A funded product already serves your exact vertical | Your operation differs materially from the norm | $200–$500/month typical |
| Custom agency (us) | Specific workflows, real integration depth, no product fits | You need something a $200/month product already does | Fixed from audit |
Build in-house
Choose it when you have engineering capacity, the workflow is core to your competitive position, and you want the knowledge to stay inside.
The thing people underestimate: not the build. It’s the evaluation infrastructure and the ongoing maintenance. Models get deprecated. APIs change. Agents drift quietly. Teams that ship a great first agent frequently have no regression suite six months later and can’t tell whether quality has degraded.
Honest note: if you’re going to build in-house anyway, our audit is still useful and cheaper than discovering the integration constraints yourselves. Plenty of clients take the document and build it themselves. That’s fine.
Horizontal platforms
Choose them when your need is a solved problem. Customer service deflection is the most mature agent category there is, and per-resolution pricing has standardised around $0.99–$2.00 per resolved conversation — Intercom Fin at $0.99, Zendesk AI at ~$1.50–$2.00 (Intercom, 2026). You will not beat that economically with a custom build.
Where they break: the moment your workflow needs to know something specific to your industry. A platform that understands tickets doesn’t understand permit-to-work.
Vertical products
Choose them when a funded product already serves your exact vertical well. Home services, real estate scheduling and insurance claims all have well-capitalised vertical products now, and they’re often the right answer.
We’ll name them in the audit. If a $500/month product does 80% of what you need, we’ll say so rather than quoting you a build.
Where custom wins: when you’re the operation the product wasn’t designed for — multi-site, unusual compliance, non-standard systems, or a workflow the product treats as an edge case.
Offshore development shops
Choose them when you need volume engineering at a low rate and you have the internal capacity to specify and manage tightly.
Where it goes wrong: agent projects fail on integration judgement and evaluation discipline, not on code volume. Those are the hardest things to specify remotely and the easiest to skip.
Where we’re genuinely the wrong choice
- You want the cheapest possible option
- Your workflow is standard and a product already covers it
- You need 20+ engineers on a programme
- You want someone to take full accountability for an outcome we don’t control
- You’re looking for a proof of concept to satisfy a board rather than something that runs
We’d rather tell you in the first call than in month three.
Frequently asked questions
Isn’t it against your interest to send us elsewhere?
Short-term, yes. Clients we shouldn’t have taken churn, complain and cost more to service than they pay. We’d rather have the referral.
How do we evaluate any AI agency, including you?
Ask: can you reach our systems of record, or only read from them? Show me an eval suite from a real project. Who owns the accounts? What happens when the model we build on is deprecated? What’s your rollback path? Vague answers to any of those are the signal.
What if we’re not sure which option fits?
That’s what the audit answers, and it explicitly includes the disqualification list.