A $2,000 ceiling can support a small hosted knowledge pilot when the buyer supplies a clean, permissioned source set and accepts platform constraints. The value should come from citations, refusal behavior, evaluation, and handoff—not from pretending the budget buys a custom production RAG platform.
Use one public website audience and a compact source collection with clear owners and update dates. Keep account actions, private customer records, complex access control, multiple channels, and enterprise support outside the pilot.
Related service: AI Chatbot Development services- 01Curate and approve one compact knowledge set
- 02Configure retrieval, citations, fallback, and handoff
- 03Build a 60-question held-out evaluation
- 04Pilot on one website section with analytics
- 05Review transcripts and decide whether custom architecture is justified
SECTION 01
A worked $2,000 budget: hours, delivery, and reserve
Allocate 40 hours at an assumed $40 per hour: $1,600 for the work below and $400 held for corrections. The rate is an illustrative configuration assumption within Clutch's global $25–$49 AI range, not a U.S. agency average, a confirmed quote, or a Catapult rate. Market source checked September 20, 2026.
For comparison, the same 40 hours at Clutch's listed U.S. AI-company range of $50–$99 per hour would be $2,000–$3,960 in labor alone. A provider may need more hours or a minimum engagement. Forty hours assumes an existing chatbot platform and one approved source set, not a newly engineered chatbot product.
| Work | Hours | At the assumed $40/hour |
|---|---|---|
| Agree source boundaries and answer rules | 8 | $320 |
| Configure the knowledge and intake flow | 16 | $640 |
| Evaluate unsupported answers and escalation | 12 | $480 |
| Handoff, logs, and source maintenance | 4 | $160 |
| Correction reserve—not extra features | — | $400 |
| Maximum planning envelope | — | $2,000 |
SECTION 02
Constrain the $2,000 chatbot to one curated public knowledge domain
The buyer provides an approved set such as up to 30 short pages or documents, removes duplicates and obsolete versions, names the canonical source for conflicts, and supplies 60 questions with expected answers or safe outcomes. One hosted platform and one website surface are assumed.
The provider configures ingestion, retrieval settings available in the product, citations, refusal, handoff, basic event analytics, evaluation, and operator documentation. Custom vector infrastructure, data pipelines, identity, private account context, multilingual coverage, and continuing optimization are excluded.
| Included deliverables | Explicit exclusions |
|---|---|
| One curated-source hosted knowledge chatbot | Custom RAG stack, model training, or bespoke chat application |
| Citation, refusal, and human-handoff behavior | Authentication, role-based sources, or customer-account access |
| Sixty-question evaluation with severe-error review | CRM, booking, transaction, helpdesk, or omnichannel actions |
| Basic analytics, operator guide, and limited stabilization | Unlimited documents, revisions, traffic, support, or performance guarantee |
SECTION 03
Worked $2,000 example: a cited public policy and service assistant
A U.S. professional association curates 22 public pages covering membership types, application steps, events, refunds, accessibility, and contact routes. The chatbot answers with a link to the supporting page and refuses to interpret policy beyond the source.
Questions about an individual's status, payment, legal eligibility, or exceptions move to staff. The pilot runs on one site section, measures 60 held-out questions, samples live transcripts, and ends with an expand, revise, migrate, or stop recommendation.
- 01
Buyer removes stale pages and resolves source conflicts
- 02
Provider configures hosted ingestion, citations, and fallback
- 03
Evaluation covers answerable, absent, conflicting, sensitive, and adversarial questions
- 04
Limited traffic launch records outcome and handoff events
- 05
Team reviews evidence and vendor limits before granting more scope
SECTION 04
How to check the results before launch
Treat retrieved text as untrusted input, restrict the corpus to approved material, protect system instructions and secrets, and require a safe path when sources do not support an answer.
Canonical sources
Every included page has a permission basis, owner, freshness date, and conflict rule.
Supported-answer check
Review whether the cited passage supports the specific answer, not merely whether a link appears.
No account action
Personal status, payments, exceptions, and transactions move to authenticated human service.
- 01
Hold out 60 questions across common, rare, absent, conflicting, and hostile cases
- 02
Score claim support, citation usefulness, refusal, handoff, tone, latency, and cost
- 03
Classify unsupported confident answers as severe even when plausible
- 04
Re-run the set after source, prompt, model, or platform changes
SECTION 05
What you will pay to run it
Knowledge assistants need content operations as well as software operation. Budget document updates, re-indexing, transcript review, model and platform usage, analytics retention, and response ownership.
| Cost item | Budget basis | Control |
|---|---|---|
| Hosted knowledge platform | Documents, storage, messages, domains, seats, and features | Track limits and exportability before content grows |
| Model and retrieval | Query volume, context size, reranking, and retries | Report cost per evaluated and live conversation |
| Knowledge maintenance | Source changes, conflicts, removals, and re-indexing | Assign source owners and review dates |
| Quality and support | Transcript sampling, incident handling, and vendor changes | Set a review cadence and separately scope ongoing support |
SECTION 06
Questions to ask your developer
Ask candidates to show how they test unsupported answers and source conflicts. Retrieval demos should not substitute for a held-out evaluation and an operator who owns content after launch.
SECTION 07
Frequently asked questions
Can a conventional U.S.-based agency deliver this AI chatbot scope for $2,000?
Ask for a quote against the exact configuration or pilot in this guide. The worked budget uses an assumed $40/hour, while Clutch lists U.S. AI-company rates at $50–$99/hour. At the higher rate, fewer hours fit. Neither benchmark guarantees that a provider can deliver the proposed scope for $2,000.
Are these Catapult package prices?
No. These examples help you plan a brief. Contact us for pricing based on your workflow, data, integrations, and support needs.
Is the $2,000 chatbot a custom RAG system?
No. This scenario uses a hosted product and its available retrieval features on a curated source set. Custom ingestion, vector infrastructure, model orchestration, identity, permissions, observability, and scaling require a different budget and architecture.
Do citations guarantee a correct chatbot answer?
No. A citation can be irrelevant, incomplete, stale, or contradicted by another source. Evaluation should check whether each material claim is supported and whether the source was approved and current.
PRIMARY REFERENCES
Sources and further reading
These references cover the standards, platforms, or published prices discussed in the guide. Worked examples and checklists are our editorial guidance.
- AI development pricing: market data checked September 20, 2026Clutch
- AI Risk Management FrameworkNational Institute of Standards and Technology
- OWASP Top 10 for LLM and GenAI applicationsOWASP GenAI Security Project
EDITORIAL METHOD
About this guide
We use AI to assist with drafting and editing. Catapult AI Work is responsible for the published content. Examples illustrate possible approaches; they are not client case studies unless identified as such.
Budget examples are not Catapult package prices. Check linked provider pages for current fees and plan limits before making a purchase.
Read the editorial policy