Four Underestimated Costs in Custom AI Agent Quotes
In custom AI Agent quotes, models and hosting look loudest. What actually slips schedules and forces change orders is usually four quieter lines: data cleaning, ACL design, joint integration, and gray-release ops. If procurement only compares “which model, how many tokens,” you discover too late that scans will not parse, contractor accounts can see HQ knowledge, the counterparty has no test env, and nobody owns weekly sampling. GeonAI recommends line-iteming assumptions and person-days for all four in the SOW; /agents are capability references only (not a public trial of 363+ presets). Email [email protected]. Artifacts: /blog/custom-agent-delivery-artifact-checklist. ROI: /blog/ai-agent-roi-calculator.
1. Data cleaning and parsing
- Scans, broken layouts, mixed languages, duplicate policies, expired attachments
- No cleaning plan = garbage in retrieval, hallucinations and bad citations out
- Line-item: parse-failure list, dedupe, version labels, ingestion acceptance report
Knowledge is not “dump into a vector store = delivered.” RAG and version ownership: /blog/enterprise-rag-knowledge-base-agent, /blog/custom-agent-more-than-a-prompt.
2. Permissions and classification
- Roles, contractors, multi-entity setups, HQ vs regional knowledge isolation
- ACL is a design problem, not “add a login”
- Line-item: ACL matrix, refuse policy, ACL-leak eval samples
Boundary tables as contract annex: /blog/custom-agent-scope-boundary-in-contract. Skipping ACL design is cheaper on paper and far more expensive after a cross-store leak.
3. Joint integration (联调)
- Counterparty test env, field dictionary, rate limits, error codes, idempotency, receipts
- Calendar time for the other team—not only “we finished the client”
- Line-item: integration person-days, blocker escalation path, joint-test log as an acceptance annex
Tool allowlists and ERP/CRM patterns: /blog/agent-function-call-mcp-integration. If there is no test env, write “environment readiness” and delay rules into quote assumptions.
4. Gray-release operations
- On-call, sampling, rollback, weekly review, scale gates—go-live is not project end
- Line-item gray support for 4–8 weeks (adjust for traffic and harm metrics)
- Pilot metrics and “do no harm” gates: /blog/custom-agent-30-day-pilot-metrics, /blog/poc-to-production-agent-checklist
Quote structure that catches the leaks
| Line item | Often missed? | How to write it |
|---|---|---|
| Scenario design & boundaries | Less often | Include intent table + boundary drafts |
| Model / hosting | Less often | Clear usage assumptions |
| Data cleaning & parsing | Often | Tier by doc volume/formats |
| ACL / classification design | Often | By roles and isolation layers |
| Integration person-days | Often | By systems + calendar windows |
| Gray ops 4–8 weeks | Often | Include sampling + weekly gates |
Fixed-price packages are fine—if document volume, system list, and test-env availability are written assumptions. Overruns become change orders, not verbal “add a little.”
How to brief GeonAI
Share doc volume and formats, system list, whether a test env exists, and whether the entry is external or internal. Email [email protected], /pricing, or Live chat. We quote acceptable line items—we do not bury these four under “misc.”
Frequently asked questions
Can we fixed-price all four items?
Yes, with written assumptions on doc volume, systems, and test env. Over those assumptions → written change orders, not infinite scope.
Can the customer do data cleaning themselves?
Yes—if the contract defines ingestion acceptance criteria and reject/return rules. Otherwise dirty data still lands on the delivery team.
How do we quote integration without a test environment?
Mark environment readiness as a precondition; agree in advance whether wait time is billable.
Is four weeks of gray ops enough?
It depends on traffic and harm-sampling volume. Low traffic can extend the window—but do not skip gates because “we have no samples.”
How does this relate to the pricing-factors article?
That post covers pricing dimensions; this one names four frequently omitted SOW lines. Use both as a checklist: /blog/custom-ai-agent-pricing-factors.
Is “model API only” a full custom quote?
No. A quote without cleaning, ACL, integration, and gray ops will usually reappear as those four invoices after launch.