Fixed price per step
You know the cost of Discover, Prove and Build before each one starts.
How we work
Four steps. A decision point after each one.
Discover → Prove → Build → Run. Each step has a fixed scope, a fixed price and a clear output, so you never commit to more than you can judge.
If we cannot measure it, we do not ship it.
We agree what “good” means before we build.
Most failed AI projects fail on one question nobody asked at the start: how will we know it works?
So we answer it first. Before any production code, you and we sign off a success metric and an evaluation set. The evaluation set is a few hundred real examples from your business, with the right answer for each, checked by your people. Every version of the system is scored against it. The score decides whether we move to the next step, not a demo or a feeling.
The four steps at a glance
01
Discover
- Length
- 2 to 3 weeks
- Price model
- Fixed price
- You decide
- Which use case to prove
- Output
- Roadmap and business case
02
Prove
- Length
- 4 to 8 weeks
- Price model
- Fixed price
- You decide
- Whether to build
- Output
- Scored pilot on real data
03
Build
- Length
- 2 to 6 weeks
- Price model
- Fixed price, by milestone
- You decide
- When to go live
- Output
- System in production
04
Run
- Length
- Monthly
- Price model
- Monthly retainer
- You decide
- Whether to continue, each month after the minimum term
- Output
- Weekly quality report and improvements
Step 1Discover
2 to 3 weeks. Usually sold as the AI Opportunity Sprint.
Find the 20% of the problem that matters.
What happens
We interview 6 to 12 stakeholders and run 2 to 3 working workshops with the people who do the work: the agents on the support desk, the clerks who key in invoices, the branch managers, the finance team at month end. Not only the leadership team. We watch the real workflow, look at sample data, and map the systems it touches.
Then we size each opportunity. How many hours or rupees does it cost today? How good would AI need to be to help? Is the data there? Would a rule or a report do the job more cheaply? We rank the options and recommend one to prove first.
What you get
- A map of the workflows we examined and where time and money go.
- A ranked list of AI opportunities, each with expected value, effort, risk and data readiness.
- A business case for the top one or two, with a cost range for Prove, Build and Run.
- A draft success metric and evaluation-set plan for the first pilot.
- A list of what we recommend you do not use AI for, and why.
- A read-out session with your leadership team.
What we need from you
- A sponsor who can make decisions, and 2 hours a week of their time.
- Access to 6 to 12 people who do the work, for interviews of 60 to 90 minutes each, and their time for 2 to 3 working workshops.
- Sample data or screenshots of the workflow, under NDA.
- A contact in IT who can describe the systems involved.
Step 2Prove
4 to 8 weeks. Fixed price.
A pilot on your real data, scored against an agreed test.
What happens
We build a working version of the system for one use case, on your real data, in a safe environment. Week one is spent finalising the evaluation set with your team and agreeing the target score in writing. From then on, every change we make is scored against it, and you see the score.
We test the hard cases on purpose: the blurred scan, the caller who switches from Marathi to Hindi mid-sentence, the customer who asks for something the policy does not cover. We test what the system must refuse to do, too.
What you get
- A working pilot your team can use and try to break.
- The evaluation set, owned by you, with a weekly score against the target.
- A written evaluation report: what works, what does not, the failure patterns, and the cost per transaction at expected volume.
- A recommendation: build, change the scope, or stop.
- A fixed-price proposal for Build if the pilot meets the target.
What we need from you
- Real data for the use case, with the access approvals in place before week one.
- 2 to 3 subject experts who can label examples and judge outputs, about 3 hours a week each.
- A weekly 30-minute review with the sponsor.
- A decision within two weeks of the evaluation report.
Step 3Build
2 to 6 weeks after a successful pilot. Fixed price, paid by milestone.
From pilot to a system your business can rely on.
What happens
We connect the system to the tools your team already uses: WhatsApp Business, your CRM, Tally or SAP, the core banking system, the document store. We deploy in your cloud account or on your premises. We add what a pilot leaves out.
Production-readiness checklist
- Security review with your IT or security team, including prompt-injection testing.
- Access control that follows your existing roles.
- Guardrails: what the system may say and do, and what it must hand to a person.
- Personal data handling mapped to the DPDP Act and any sector rules.
- Logging that records what happened without leaking personal data.
- Monitoring dashboards for quality, latency, cost and hand-off rates.
- Spend limits and alerts on model and cloud usage.
- A rollback plan and a staged launch, usually to one team or region first.
- Training for the people who will use it and the people who will support it.
What you get
- The system live in production, in stages you approve.
- Source code, prompts, configuration and evaluation sets in your repository.
- Runbooks for your IT team: how it works, how to restart it, who to call.
- Training sessions and short guides for end users.
- A go-live report with the final evaluation score.
What we need from you
- An IT owner who can approve access, environments and deployment.
- Time from your security team for the review, usually 2 to 3 sessions.
- Users available for training in the launch week.
- Sign-off at each milestone within five working days.
Step 4Run
Monthly. Minimum term of six months, then month to month with 60 days’ notice. Usually sold as Managed AI.
SampleWeekly quality report
- Evaluation score
- 91.4% Sample value, target agreed in writing
- Hand-off rate
- 18.2% Sample value
- Cost per transaction
- ₹2.40 Sample value
- Open issues
- 3 Sample value
Sample values for illustration only. Not a client or a result.
We stay, and we keep it getting better.
What happens
AI systems drift. Your products change, your policies change, customers find new ways to ask, and model providers update their models. So every week we re-run the evaluation set, review a sample of live conversations or documents, and fix what has slipped. Each month we propose improvements, ranked by value.
What you get
- A weekly quality report against the agreed metric.
- Monitoring and incident response within agreed hours.
- Regular updates to prompts, retrieval, rules and evaluation sets.
- Testing of new model versions before anything changes in production.
- A monthly review of cost, with recommendations to bring it down.
- A named lead at Envisstack.
What we need from you
- A business owner for the system who attends a 30-minute monthly review.
- A channel to report issues, and someone to approve changes.
- Notice of changes to products, policies or systems the AI depends on.
How long does it take overall?
From first workshop to production, most systems take 8 to 17 weeks: 2 to 3 weeks to Discover, 4 to 8 weeks to Prove, 2 to 6 weeks to Build. Simpler systems, such as a Knowledge Assistant on a well-kept document set, sit at the short end. Voice agents and anything touching core banking sit at the long end.
If you already know the use case and have the data ready, we can skip the full Sprint and start with a one-week scoping exercise before Prove.
Discover
Prove
Build
Run
04812162024
How we price each step
40% in advance, the rest on milestones
Milestones are tied to things you can see, such as the evaluation report or go-live, and each milestone invoice is due within 15 days. If you sign a pilot within 60 days of the Sprint read-out, 50% of the Sprint fee is credited against it.
Running costs passed through, never hidden
Model, API and cloud costs are charged at cost plus a 10% handling margin when we hold the accounts, on a separate line. If you pay the providers directly from your own accounts, we add no margin.
Frequently asked questions
Can we skip Discover?
Yes, if you already know the workflow, have the data, and can state the metric. We will run a one-week scoping exercise to agree the evaluation set and the target, then start Prove.
Can we stop after Prove?
Yes. Every step ends with a decision, and you are never committed to the next one. If you stop, you keep the evaluation set, the report and the pilot code.
Who from Envisstack will be on site?
The engineers building your system, and a founder at every milestone review. We tell you names before the step starts.
Do you work on site or remotely?
Both. We spend the Discover workshops and key Build weeks on site with your team, and work remotely in between on your systems through access you control.
Can our IT vendor build part of it?
Yes. Many mid-market companies have a systems integrator who owns SAP or the core platform. We agree the interfaces with them early, and we document everything so they can support it later.
Start with a 30-minute conversation.
Tell us the workflow and the problem. We will tell you which step to start with, and what it would cost.