Pricing transparency in AI services is rare. Most studios refuse to publish ranges because their numbers vary too widely, their average customer is too sophisticated to need anchoring, or — frankly — they do not want competitors knowing their pricing. We are publishing ours.
The headline ranges (US, 2026)
For context, we are a specialist studio with 5-12 people on active engagements. Larger consultancies (Accenture, Deloitte, BCG X) typically quote 2-4x these numbers. Freelancer marketplaces (Toptal, Upwork) typically quote 0.3-0.6x. Independent senior engineers cost similar to us per hour but you bear all the project management overhead.
| Engagement type | Typical cost (USD) | Timeline | What you get |
|---|---|---|---|
| Discovery | $5,000 – $15,000 | 1-2 weeks | Written problem statement, recommended approach, build/buy reasoning, cost range, stack proposal |
| Focused first project | $25,000 – $80,000 | 4-8 weeks | One workflow, end-to-end, measurable success metric |
| AI feature in existing product | $40,000 – $150,000 | 6-12 weeks | Production-ready AI feature with eval harness and cost controls |
| Platform build | $100,000 – $400,000 | 3-6 months | Multi-workflow system, custom AI infrastructure, integrations |
| Enterprise transformation | $400,000 – $2M+ | 6-18 months | Multiple platforms, organizational change support |
| Ongoing partnership | $8,000 – $40,000 / month | Rolling | Continued development, optimization, support |
Now let us break down what moves you within those ranges.
The five scope drivers that actually move the price
Most AI projects do not blow their budgets because of bad estimating — they blow because one of these five drivers was underestimated during scoping.
1. Integration count and depth
Each integration adds engineering time and ongoing risk. Approximate ranges:
- Native API integration with a major platform (Salesforce, HubSpot, Slack, Stripe): 1-2 weeks each
- Custom integration with a smaller SaaS or in-house system: 2-4 weeks each
- Legacy system integration (older databases, mainframes, on-prem systems): 4-12 weeks each
A project requiring 6 integrations is typically 3-4x the cost of a project requiring 2, all else equal. If your scoping conversation is not asking detailed integration questions, you are getting estimated for an unrealistic version of the project.
2. Data quality and accessibility
The single largest cost surprise in AI projects is data preparation. Estimates assume your data is "reasonably clean" — when it is not, costs grow:
- Clean, structured data with API access: baseline
- Mixed structured + unstructured, mostly accessible: +20-40%
- Heavy unstructured (PDFs, scanned documents, emails) in production volume: +40-100%
- Multi-language, multi-format, scattered across legacy systems: +100-200%
We will not estimate firmly until we have looked at sample data. Any partner who quotes without seeing your actual data is guessing.
3. Compliance requirements
Compliance does not just add features — it adds review cycles, vendor selection constraints, documentation, and architectural choices. Approximate adders:
- Standard B2B SaaS (no specific compliance): baseline
- SOC 2 Type II requirements: +10-20%
- HIPAA / BAA requirements: +20-40%
- GDPR with data-residency requirements: +15-30%
- Defense / FedRAMP environments: +50-150%
Compliance also constrains your vendor choices for AI providers, which can increase ongoing operating costs (private model deployments cost more than commercial APIs).
4. Real-time vs batch
Real-time AI has materially higher engineering and operating costs than batch:
- Batch processing (overnight jobs, hourly runs): baseline
- Near-real-time (5-60 second response): +20-40%
- Real-time conversational (chat, voice): +50-100%
- High-volume real-time (thousands of concurrent users): +100-200% plus higher ongoing infrastructure costs
If your use case actually tolerates batch, embrace it. We have seen projects double in cost solely because someone said "real-time" when the actual user experience needed nothing tighter than a 30-second response.
5. Custom AI features vs orchestration
There is a big difference between "we orchestrate calls to OpenAI / Anthropic" and "we build custom models or significantly fine-tune existing ones".
- Pure orchestration (prompts, tools, RAG over a vector DB): baseline
- Light fine-tuning or extensive prompt engineering with eval harness: +30-60%
- Significant fine-tuning or domain-specific model training: +100-300%
- From-scratch model development: typically a separate engagement entirely
For 90% of business applications, orchestration is what you actually need. If a partner is pushing you toward fine-tuning or custom model training, ask "what does this give us that careful orchestration of frontier models cannot?" The answer should be specific.
Ongoing operating costs (not just build cost)
People focus on build cost and forget operating cost. For AI systems, ongoing costs typically split into three buckets:
LLM inference costs
Highly variable by use case. Rough monthly ranges for production systems:
- Internal-only AI assistant (single team, occasional use): $50-500/mo
- Customer-facing chatbot (modest traffic, well-cached): $500-5,000/mo
- High-volume processing (thousands of documents per day): $2,000-20,000/mo
- Enterprise AI platform with thousands of users: $10,000-100,000+/mo
These numbers can be cut 40-70% with intentional architecture (caching, model routing, batching). They can also be 2-5x higher than expected if naive prompts go unchecked. Cost monitoring is not optional in production.
Infrastructure (hosting, observability, third-party services)
For a typical AI integration project: $200-2,000/mo. For platform builds: $2,000-15,000/mo. For enterprise systems: $15,000+/mo.
Maintenance and continued development
Most engagements continue at 10-30% of build cost annually as ongoing partnership — keeping the system running, improving it, and adapting as the underlying AI providers change. Some clients move maintenance in-house after launch (we deliberately ship stacks teams can operate); others retain us as their AI engineering function.
What "$50,000" actually buys (a worked example)
Concretely: what does a $50,000 first project look like? This is roughly the size of project we run most often.
Engagement shape: 6 weeks, one focused workflow, two senior engineers part-time + a project lead, fixed scope.
Typical breakdown:
- 1 week discovery (problem definition, data sample review, stack proposal): ~10% of cost
- 2 weeks design and integration architecture: ~20% of cost
- 2 weeks core build and AI integration: ~40% of cost
- 1 week testing, monitoring setup, deployment, training: ~25% of cost
- 5% buffer for edge cases (held back unless used)
What you get:
- Production-ready AI workflow handling one specific business process
- Source code in your repo
- Documentation your team can maintain from
- Eval harness with test cases
- Cost monitoring and alerts
- 2-week post-launch support included
What you do not get:
- A second workflow (separate engagement)
- Ongoing changes after the first 2 weeks (separate retainer)
- Integration with systems we did not scope upfront
This is a real shape, not a marketing example. About 35% of our annual engagements look exactly like this.
How to estimate your own project before talking to anyone
Before your first sales call with any studio, do this back-of-envelope:
1. Identify the one workflow you would automate first. Not three. Not "AI for our company". One workflow.
2. Count its integrations. Be honest. Each integration is real engineering work.
3. Assess your data. Is it clean and structured? Or is it PDFs and emails?
4. Establish your latency requirement. Real-time? Within an hour? Overnight?
5. List your compliance constraints. SOC 2? HIPAA? Nothing specific?
For a workflow with 2-3 integrations, reasonably clean data, no special compliance, and tolerance for non-real-time response: budget $30-60k for the first project.
Add 50-100% if compliance applies. Add 50-100% if data is heavily unstructured. Add 50-100% if real-time is genuinely required. Add per-integration if integration count climbs.
If you arrive at a number that feels too high, your scope is too big — pick the most valuable single workflow and start there. If you arrive at a number that feels too low, you are missing scope drivers — pressure-test what you have not accounted for.
What to do next
If you want a real estimate against your specific project, book a 30-minute consultation. We will run the five scope drivers against your project on the call, and either give you a same-day rough range or schedule a paid discovery if the project warrants it.
If you are still in the "do I even need this" phase, our free business assessment walks through the high-level decision in about 10 minutes.