The architecture, in plain English
Every business agent has five parts. The model (Claude, Gemini) provides reasoning. Tools give it hands: function calls into your email, CRM, database, calendar. Grounding gives it your truth: retrieval over your docs and data so it answers from reality, not vibes. Memory/state lets multi-step work survive across turns. And the control layer — the part that separates products from demos — decides what it may do alone, what needs human sign-off, and what happens when it's uncertain.
Buyers don't need to implement this; they need to interrogate vendors about part five, because it's where corners get cut.
What drives cost up or down?
Four multipliers. Integration surface: each system the agent touches (CRM, helpdesk, ERP) adds connection, permission and failure-mode work. Reliability bar: an internal research assistant tolerates errors a customer-facing refund agent cannot — evals and guardrails scale with stakes. Data readiness: clean, accessible data is a discount; scattered tribal knowledge is a surcharge. Compliance: regulated data (health, finance) adds audit trails and access controls.
Market rates for production builds run $25k–100k. Ongoing: model usage typically $100–1,000/month per workflow at SMB scale, plus maintenance as your tools and policies change.
- Each integrated system: +cost, +failure modes
- Customer-facing stakes: eval suites become the project
- Messy data: fix it before or during — never after
- Regulated data: audit trails from day one
Build in-house, buy SaaS, or commission custom?
Buy off-the-shelf when a vertical product already nails your exact workflow (support triage has credible SaaS). Build in-house when agents are your core product and you're hiring for it permanently. Commission custom when the workflow is yours alone — your systems, your policies, your edge cases — and you want ownership without building an ML team.
The decision heuristic: if you can describe your workflow to a SaaS demo and it fits 90%, buy it. If every vendor demo requires 'well, except for us…' three times, it's a custom build. Chalk Labs does the audit as the first step of every agent engagement — sometimes the honest answer is a $200/month subscription, and we'd rather say so than bill you $50k to rediscover it.
Questions we hear about this
For buyers, the framework question is mostly a distraction — model APIs' native tool-calling plus thin orchestration covers most business agents, and heavyweight frameworks add churn risk. Judge vendors on eval methodology and escalation design, not framework name-dropping.
Layered defense: retrieval grounding (answers cite your data), constrained action spaces (it can only do what tools allow), confidence thresholds routing uncertain cases to humans, and eval suites replaying hundreds of real scenarios before and after every change. Zero hallucination is a lie; contained, detected error is the standard.
The workflow's actual artifacts: policy docs, canonical FAQs, historical tickets/emails with resolutions, and API access to systems of record. A surprising amount of 'agent readiness' is just writing down rules that currently live in someone's head.
Baseline the workflow first: volume, human minutes per case, error/escalation rates. Post-deploy, track containment rate (handled without humans), quality on sampled outputs, and cycle time. Most single-workflow agents at $15k–50k pay back within two to six months on labor math alone.