A fixed-fee, fixed-timeline pilot that puts a production-grade AI system on your data in 4 weeks and ends with a go/no-go decision and a plan for production. Something to show in week 2. You own the code from day one.
The AI Pilot in 4 Weeks is the Erbis fixed-fee entry engagement: one use case, your data, a production-grade build with evaluation, guardrails and integration, delivered in four weeks with a go/no-go decision and a production plan at the end. It is for leaders who need evidence before committing budget, and an exit if the evidence is not there.
MIT’s NANDA initiative found 95% of enterprise GenAI pilots show no measurable P&L return; the cause was a learning and workflow gap, not model quality. Gartner expects 30% of GenAI projects to be abandoned after proof of concept because of poor data quality, inadequate risk controls, escalating costs and unclear value. A pilot built as a demo tests none of those four.
The typical pilot runs on a clean sample, skips integration, has no eval set and no cost ceiling, and ends with a slide deck. Everyone liked it. Nobody can say whether it would survive real volumes, real exceptions or a security review. So it waits, then quietly stops. Plug and Play’s 2026 survey found 74% of enterprises run AI in production, and half of them cannot prove it pays off.
We designed this pilot to produce the four things a production decision needs: measured accuracy on your data, a working integration path, a cost per transaction, and a governance review. It is fixed in budget and timeline, flexible in scope, and ends in a decision either way. A no-go with a clear reason is a good outcome. It cost 4 weeks instead of 4 quarters.
We pick the use case with the clearest metric, the most available data and the shortest path to a system of record. Not the most impressive one.
The same architecture we would ship, at a bounded scope.
An eval set built from your own cases in week 1, so accuracy is a number, not an impression.
Price and end date do not move. What moves is scope, and we say so in writing when it does.
Week 4 ends with a decision meeting, not a wrap-up.
Everything produced in the pilot is yours, whether or not you continue.
Four deliverables, each one an input to a production decision. If any of them cannot be produced, that is itself the finding.
A data-flow diagram showing what the system reads, writes and stores, with classification of PII and PHI, access scopes, retention and the residency of every model call. Reviewed with your security lead. Gaps listed as production-plan items with an owner.
The eval set and its results: accuracy, straight-through rate, error rate and failure categories on your records, compared with the current baseline. Test coverage of tools and integrations. The threshold you set in week 1 and whether it was met.
Tracing of every run from the pilot, a dashboard of volume, latency, failures and token spend, and a cost per transaction at pilot volume with a projection at production volume. Budget controls and hard stops configured.
A working integration into at least one system of record, documented. A production plan with architecture, remaining integrations, team and a fixed price for phase two. Repository access, infrastructure templates and a handover session, whether you continue or not.
15 min to < 1 minaccount setup time per user
Two AI agents generate onboarding and password-reset scripts on serverless AWS; script caching cut OpenAI API costs 90% and setup time to under a minute.
$10M+annual savings across the companies using the solution
AWS Textract plus SageMaker Ground Truth human-in-the-loop review digitizes ~95% of invoices unattended, saving 10+ companies $10M+ a year.
28,800 hsaved per year
A custom-trained LLM with a verification controller generates Word contracts from 10,000+ precedents via an intranet chatbot, making drafting 10x faster.
Business analyst, solution architect and AI engineer with your owner. We define the use case, the success threshold, the risks and the data. The eval set is built from your records. Access to systems and data is arranged. Output: a one-page pilot charter, signed by both sides.
The working system on your data, end to end at a narrow scope, with first eval results. Where an interface matters, a clickable prototype is tested with the people who will use it. This is where most assumptions break, which is why it comes early.
Integration to the system of record, guardrails, confidence gate, tracing and cost controls. Architecture diagram drawn for production. Second eval run; thresholds tuned. Security review with your team.
Final eval run at the largest volume the pilot allows. Cost per transaction calculated. Production plan written: scope, milestones, schedule, team, price. Decision meeting with your owner, your CTO and ours. Go, no-go, or go with conditions.
One fixed fee, quoted after a 45-minute discovery call once we know the use case, the data and the systems involved. The price does not change during the 4 weeks; scope does, in writing, if we learn something that requires it. The fee covers a business analyst, a solution architect and AI engineers for four weeks. Cloud usage runs in your account and is small at pilot volume. The production plan at the end carries its own fixed price for phase two.
Four recur: a use case chosen for impressiveness rather than evidence, a clean sample instead of real data, no eval set so accuracy stays an impression, and no integration so nobody learns what production would take. The pilot is designed against each: the use case is scored on data readiness and integration effort, the eval set is built from your records in week 1, and the system writes to at least one real system by week 3.
A signed pilot charter with the success threshold (week 1); a working system on your data with first eval results (week 2); an integrated, hardened build with an architecture diagram and a security review (week 3); and in week 4 the final eval results, a cost per transaction, a data-governance review and a production plan with a fixed price for phase two. All code, prompts, eval sets and infrastructure templates are in your repositories throughout.
You get a no-go with the reason: the data was not ready, the accuracy threshold was not reachable at acceptable cost, the integration path was blocked, or the governance review found a stopper. You keep everything built, the eval set and the findings, which usually become the shortlist for what to fix. Nothing further is owed. A clear no-go after 4 weeks is cheaper than a slow no after 4 quarters, and we count it as a successful pilot.
Yes, and it is often the best candidate: the data access exists, the stakeholders know the problem, and the previous attempt shows where it broke. We start by reading what was built and why it stalled, then re-frame the use case around a measurable threshold and a real integration. If the earlier work is sound we build on it. If the use case cannot pass the evidence test, we say so in the discovery call rather than in week 4.
One owner with the authority to decide go or no-go, about two hours a week. A subject-matter expert who knows the process and can review eval cases, a few hours in weeks 1 and 2. Someone who can grant access to data and systems in week 1. A security or compliance contact for the review in week 3. Your engineers are welcome throughout and are needed for the handover session in week 4.
Bring one workflow and the number you want to move. In a 45-minute call we will tell you whether it fits a pilot, what it would cost and what you would know at the end of week 4.
AWS partner · Databricks partner · ISO 27001 · You own the code from day one
