Buyer guide · Updated October 2026
The model is rarely the expensive part. The integrations, the review steps and the operating costs are. This guide shows where the money goes so you can compare quotes on the same workload.
Get a scoped estimateDirect answer
For HummingAgent, a focused automation build is budgeted at $7,500 to $20,000; a custom AI agent integrated with your systems at $15,000 to $50,000; multi-system and enterprise builds start at $50,000. Discovery starts at $2,500 and is usually credited against the build. Ongoing support and improvement, when you want it, starts at $1,500 a month.
Those are planning ranges, not quotes. The number for your business depends on five things covered below: how many systems the agent touches, how clean the data in them is, how many actions it is allowed to take without a person, where it has to run, and how much it will be used after launch.
A custom AI agent is three layers. The model does the language work: reading a request, extracting fields, drafting a reply. Software around the model does the business work: identity, permissions, state, validation, retries, logging, the review screen. Integrations connect that software to the systems you already run. In a typical build the model layer is the smallest share of the engineering time. The software and integrations are most of it.
This is why two quotes for "an AI agent" can differ by a factor of five and both be reasonable. One is a chatbot that answers questions from a document set. The other reads invoices from a mailbox, checks them against purchase orders in your ERP, queues exceptions for a person and posts approved ones through the accounting API. Same phrase on the proposal, very different work.
| Driver | Why it moves the price | What makes it cheaper |
|---|---|---|
| Number of systems | Each CRM, phone system, document store or accounting package is its own authentication, data model, rate limits and failure modes. | Start with one system of record and one source. Add the second integration after the first is stable. |
| Data quality | If two systems disagree about a customer, price or status, the agent amplifies the disagreement. Cleanup or an explicit sync rule has to happen first. | Pick a workflow where the source of truth is clear and already trusted by the team. |
| Allowed actions | Reading is cheap. Writing records, sending messages or moving money needs validation, duplicate protection, an approval step and an audit trail. | Launch read-and-draft first. Let the agent prepare the action and a person click it. Expand autonomy from evidence. |
| Where it runs | Deploying in your own cloud or tenant adds infrastructure setup, secrets management, network rules and a security review on your side. | Decide this at discovery, not after the build. A late infrastructure change is the most common overrun. |
| Usage after launch | Model calls, call minutes, telephony, hosting and monitoring are metered. A fixed build price does not make usage unlimited. | Ask every vendor for the operating cost at your expected monthly volume and at twice that volume. |
Rough relative effort for the integrations we see most often. Modern SaaS with a documented API sits at the low end. Anything that only exposes a screen, an export file or a legacy database sits at the high end, because the integration has to be built and then defended against changes the vendor never announces.
| System | Typical integration | Relative effort | Common surprise |
|---|---|---|---|
| CRM (HubSpot, Salesforce, Zoho) | Create or update contacts, deals, notes; read pipeline | Low to medium | Custom fields and duplicate rules differ per account; Salesforce API limits on lower tiers. |
| Phone and SMS (Twilio, RingCentral, your carrier) | Answer calls, transcribe, route, send confirmations | Medium | Number porting, call-recording consent by state, and after-hours routing rules nobody wrote down. |
| Calendars and scheduling | Read availability, book, reschedule, confirm | Low to medium | Multiple calendars per person, buffer rules, and time zones for field teams. |
| Email and documents (Microsoft 365, Google Workspace, SharePoint) | Watch a mailbox, read attachments, file results | Medium | Admin consent for mailbox access and retention policies that delete what the agent needs later. |
| Accounting and ERP (QuickBooks, NetSuite, Sage, Dynamics) | Create bills, match POs, post journal entries | Medium to high | Sandbox access, approval hierarchies, and period-close locks. |
| Field service and practice management | Create jobs, update status, pull customer history | Medium to high | Partial or paid API access; some vendors require a partner agreement. |
| Legacy or no API | Screen automation or file exchange | High, and ongoing | Breaks silently when the vendor changes a screen. Budget for maintenance or replace the system. |
Every custom agent has a monthly bill made of metered services. For a voice receptionist that is call minutes, speech-to-text, the model and telephony. For a document workflow it is model usage per page, storage and the hosting that runs the review screen. None of these are HummingAgent revenue; they are paid to the providers, and you should see them itemized.
Two questions settle most of the uncertainty. What is the operating cost at my expected monthly volume? What is it at twice that volume? If a vendor cannot answer both with the billable unit named, the quote is not finished.
A home-services company with two offices wants an AI receptionist that answers every call, qualifies the request, books into the right technician calendar and texts a confirmation. Systems: the phone carrier, Google Calendar for four technicians, a field-service platform for the job record, and SMS.
Discovery produces the call flows the owners actually want, the service-area rules, the emergency definition, and the handoff rules for callers who ask for a person. The build connects telephony and speech, reads calendar availability with buffer rules, creates the job in the field-service platform through its API, and sends the confirmation. Exceptions, which here means an angry caller, an address outside the service area, or a calendar conflict, transfer to the office with the transcript attached.
This sits in the custom-agent range rather than the focused-build range because of the four systems and the write actions (booking, job creation, SMS). The operating cost is dominated by call minutes and speech, which is why the estimate is built from the company's actual monthly call volume rather than a flat subscription. The metrics agreed before launch: answered-call rate, booked-call rate, transfers per hundred calls, and minutes of office time per booking.
A subscription chatbot answers questions from content you upload and lives inside the vendor's platform. A custom agent is integrated with your systems, takes actions in them under rules you set, runs where you choose, and is owned by you. You are paying for the software and integrations around the model, which a subscription product does not build for your business.
Discovery starts at $2,500 and produces the workflow map, architecture and estimate. Some teams use that alone to decide whether to build or to fix a process first. It is credited against the build if you proceed with HummingAgent.
The engineering to build an integration is one-time. What recurs is the operating cost of the connected services (API tiers, telephony, model usage) and any maintenance when a vendor changes its API. Modern APIs change rarely; screen-based integrations change often, which is why we advise against them.
In our experience: a late decision to deploy in a different environment, discovering during the build that two systems disagree about the data, and adding a second or third integration before the first is stable. All three are avoidable at discovery.
No. You own the code and the deployment. Ongoing support and improvement is optional and starts at $1,500 a month. If you cancel it, the system keeps running; you take over the monitoring and the vendor bills.
We are model-agnostic and choose per step. Extraction, classification and routing usually run on smaller, cheaper models; drafting or complex reasoning may use a larger one. If your security team approves only a specific provider, we design to that. Model choice changes the operating cost, not the build cost, in most projects.
Bring one workflow and the systems it touches. A discovery session maps it, names the integrations and gives you a scoped estimate with operating costs itemized.
Schedule a discovery call