Enterprise AI Copilots that Hold Up in Production
from $60K. 4 to 6 weeks.
from $200K. 3 to 4 months.
from $80K. 6 to 10 weeks.
Where Do You Start?
Three situations we see most often. Pick the one closest to yours.
You're Starting Fresh
AI is a stated priority. Budget is approved. You have an eighteen-month horizon to ship something that holds up under user use, audit, and regulator review. You have not built this before and you do not want to.
Architecture And Scoping (4 to 6 weeks, from $60K). Then Copilot Production Build (3 to 4 months, $200K to $280K typical).
A production RAG system on your infrastructure with evaluation harness, citation enforcement, governance documentation, and a team that can operate it after we leave.
You Shipped a Copilot That's Not Holding Up
Your existing copilot launched, but accuracy has drifted, adoption has stalled, or your team is fielding hallucination complaints. You don't need to start over. You need to fix what's broken and lift the system to a defensible standard.
Copilot Diagnostic (3 to 4 weeks, from $40K). Then Copilot Rebuild (6 to 10 weeks, $80K to $200K typical).
A rebuilt retrieval architecture on your existing platform, restored evaluation harness, accuracy lifted to target threshold, and a clean handover.
You Have a Copilot Live and Want It Run Properly
The system is in production. Your team is stretched and ongoing tuning, content onboarding, evaluation runs, and model swap-outs are not getting done. You need embedded capacity that treats this as a system that has to keep working. Not a project that finished.
Copilot Operations Retainer (12 months rolling, from $15K per month).
Monthly evaluation runs, content-source onboarding, model upgrades, drift monitoring, and ongoing user-feedback integration. Named owners on both sides.
What a Typical Engagement Looks Like
Most clients arrive in one of the three situations above. The shape of a typical Build engagement is below. The rhythm, the team, the cadence, the milestones. Rebuilds and Operations engagements run shorter and lighter, but the discipline is the same.
Architecture and Scoping
- Weeks 1 to 2: Content estate audit. We map your sources, formats, access controls, and freshness. Your SMEs are involved from day one.
- Weeks 3 to 4: Ground-truth Q&A pair set defined with your team. Retrieval architecture document drafted. Access and identity model proposed.
- Weeks 5 to 6: Evaluation harness design completed. SOW for the production build delivered, ready for procurement.
Production Build
- Weeks 7 to 8: Foundation build. Environment, ingestion, chunking, embedding pipeline. First demo at end of sprint 1.
- Weeks 9 to 12: Retrieval layer hardening. Evaluation harness running against ground-truth set. Iterations against accuracy threshold.
- Weeks 13 to 16: Citation enforcement, hallucination controls, admin dashboard, governance documentation. Pilot user group enabled.
- Week 17: Production cutover. Hypercare begins.
- Week 18: Handover documentation delivered. Working session with your team.
Hypercare and Handover
- Senior engineers on call. Daily monitoring of accuracy, fallback rate, and user feedback. Adjustment cycles if needed.
- Operations runbook finalized against actual production patterns.
Operations Retainer
- Monthly evaluation runs against expanded ground-truth set.
- New content-source onboarding as your knowledge estate grows.
- Model swap-outs as the ecosystem evolves.
- Quarterly governance review with your sponsor.
Typical program investment: Build plus first year of Operations is $340K to $460K. Scoped against your specific content estate and use case during Architecture And Scoping.
Building a copilot program your CIO and CISO can defend?
Who's on the Team
The senior architect who scopes the work is the senior architect who delivers it. We do not rotate practitioners off engagements after kickoff. Your team is named in the SOW.
-
Lead Architect
12+ years in enterprise AI delivery. Owns retrieval architecture, evaluation strategy, and engagement quality. Single point of accountability.
-
Senior Engineer, Retrieval & Data
Deep on chunking strategy, embedding tuning, content-estate integration, and access control. Writes the production code that determines whether the copilot holds up.
-
Senior Engineer, Integration & Evaluation
Owns the evaluation harness, citation enforcement, and integration with your existing identity and permissions stack. Builds the scaffolding your team operates from.
-
Platform Engineer
Cloud-agnostic deployment. Infrastructure-as-code. Observability stack. Hypercare on-call rotation.
-
Delivery Lead
Sprint cadence, fortnightly executive steering with your sponsor, RAID log, milestone acceptance. Your single contact for engagement status.
How We Approach the Work
The retrieval layer is 80 percent of an enterprise copilot. The model underneath is commoditizing. What holds up over years is the discipline applied to data, chunking, embeddings, access model, and evaluation harness. We spend most of every engagement on those layers, not on prompt engineering.
We architect for model-swappability from day one. Whatever frontier model you run today, you should be able to swap it for a better one in twelve months without rebuilding the system. This is a specific architectural decision about where the model interface sits relative to the retrieval layer, and we make it explicitly with you on day one.
We build evaluation before we build the copilot. Most copilots ship without an evaluation harness, which is why drift is invisible until users complain. Our delivery includes a ground-truth Q&A set you own and extend, version-controlled, with regression testing on every model or prompt change.
Governance is built in, not bolted on. PII handling, access control, audit trail, citation enforcement, and red-team testing against prompt injection are part of the build. Not a follow-on engagement. Framework alignment to NIST AI RMF, ISO 42001, EU AI Act, HIPAA, and DORA delivered as part of the documentation set.
Technology and Platform Posture
You have already made decisions about your cloud, your data platform, and your model posture. We deliver against your choices, not ours. Below is what we work with most often. If your stack isn't listed, ask. We likely cover it.
Model Providers
Anthropic 
OpenAI 
Azure OpenAI Service 
Amazon Bedrock 
Google Vertex AI Open-weight models via Hugging Face (Llama, Mistral, Qwen, others).
Retrieval and Orchestration
LangChain 
LangGraph 
LlamaIndex 
vLLM Custom-built orchestration where the use case warrants it. Vector stores: Weaviate, Qdrant, Milvus, pgvector. Evaluation tooling built on open frameworks, owned by your team after delivery.
Deployment Platforms
AWS 
Azure 
Google Cloud 
Databricks 
Snowflake On-premise where data residency or sovereignty requires it. Cloud-agnostic by architectural decision, not by inheritance.
Vendor lock-in is managed, not assumed. We make the architectural decisions that determine exit cost explicitly with you on day one.
Bring the use case. We return with architecture, scope, and SOW.
What to Expect from a Copilot We Deliver
Production-grade enterprise copilots typically land in the following ranges. These are category benchmarks for this class of engagement, not blanket commitments. Specific outcome targets for your engagement are scoped during Architecture & Scoping and written into the SOW.
80 to 90 percent.
under 3 percent on monitored queries.
40 to 70 percent depending on rollout discipline.
Where a target falls outside range due to constraints in your content estate or use case, we surface that during Architecture & Scoping and scope to a defensible number. We do not commit to outcomes we cannot defend.
From the Copilot Practice
Production microservices and microfrontend platform for a European energy operator. Real-time monitoring across gas, electricity, EV charging, and carbon emissions.
Most AI agents are prompt layers, not products. Here's why most agent startups won't survive, and the execution architecture that actually lasts.
Golden datasets, regression suites, automated evaluator models, drift detection, and the handoff criteria for moving an LLM from internal use to customer-facing. Written for engineering teams about to ship.
Scope a copilot engagement
Scope a Copilot Engagement
Tell us where you are. Building, rebuilding, or operating. Thirty minutes is enough to know if there's a fit and what shape the engagement would take. If we're not the right firm for what you need, we'll point you to who is.