How long does AI discovery sprint take? A realistic timeline
How long does AI discovery sprint take?
Ten working days, two calendar weeks, from kickoff to a written recommendation. That is the Eazyware AI discovery sprint at $3,250 or ₹2,00,000. Elapsed time stretches when data access is slow, so the honest planning figure is three to four weeks from first call to decision.
Ten working days, two calendar weeks, from kickoff to a written recommendation. That is the Eazyware AI discovery sprint at $3,250 or ₹2,00,000. The elapsed time stretches when data access is slow or stakeholders cannot meet, so the honest planning figure is three to four weeks from first call to decision.
Below is the day-by-day schedule we actually run, what consumes the elapsed time between those days, which parts can overlap, and the conditions under which ten days is not enough and we say so before taking the work.
The ten-day schedule, day by day
The sprint is sequenced so that the expensive question, whether the data supports the use case, is answered before anyone designs anything. Each block ends with an artefact, not a status update.
| Days | What happens | What exists at the end |
|---|---|---|
| Day 0, before the clock starts | NDA signed, access requested, interviewees named, success criteria drafted | A kickoff agenda and a live access request |
| Days 1 to 2 | Kickoff, process observation, interviews with the people doing the work today | Current-state map with volumes, handling times and pain points |
| Days 3 to 5 | Data inspection: real tables, documents and transcripts opened and measured | Data findings note with null rates, duplicates and label quality |
| Day 6 | Golden set assembled from one hundred to three hundred real items and labelled | A scoreable evaluation set you keep |
| Days 7 to 8 | Technical options assessed against your constraints, architecture sketched | Approach recommendation with models, retrieval and integration points |
| Day 9 | Cost, risk and compliance modelling for build, run and support | Costed estimate with a monthly inference forecast |
| Day 10 | Written recommendation and a readout where the decision gets argued | A decision document and a sequenced next step |
Two things about that schedule are deliberate. The golden set is built in the middle, not at the end, so the technical assessment has something to be scored against. And the readout is a working session rather than a presentation, because the recommendation has to survive disagreement from the people who will fund it.
Day 0 is not padding. Almost every sprint that overruns did so because the NDA, the access ticket and the interview invitations were treated as kickoff activities instead of prerequisites. We sign NDAs before the first working session precisely so that this week can start in parallel with contracting rather than after it.
What actually consumes the elapsed time
Ten working days of our effort rarely equals ten calendar days of your project. Three things create the gap.
Data access approvals
This is the single largest source of slippage. In a bank, a hospital or a university, read access to a production table can take one to three weeks to approve, and the request cannot start until the NDA is signed. Start the access request on the day you decide to run a sprint, not on the day it kicks off.
Stakeholder calendars
A sprint needs four to six people for roughly two hours each, plus a data owner at about one day per week. If your operations lead is on leave in week one, the current-state map is built on hearsay and the whole sprint degrades. Book the interviews before kickoff, with named substitutes.
Data that has to be extracted before it can be read
Where records sit in a legacy system with no export path, someone has to build one first. That is engineering work, quoted separately, and it typically adds one to three weeks. It is a frequent finding in legacy modernisation contexts where the information exists but is not reachable.
What can run in parallel
A sprint compresses well because several workstreams do not depend on each other. Overlapping these is the difference between a fortnight and a month.
- Access requests and interviews. Interviews need people, not permissions. Run them while security reviews the access ticket.
- Data inspection and the current-state map. One engineer opens tables while the product lead sits with the team doing the work. They reconcile on day five, and the disagreements between the two views are usually the most valuable finding.
- Golden set labelling and technical assessment. Labelling can start as soon as a representative sample exists, and your subject matter expert can label while our engineer benchmarks approaches.
- Compliance review and architecture. If DPDP, RBI or sector rules constrain where data may be processed, that assessment runs alongside the design rather than after it, because it changes the design.
- Cost modelling and the write-up. The inference forecast is arithmetic once the approach is chosen, so it drafts in parallel with the recommendation itself.
What adds weeks, and roughly how many
Scope changes the number more than complexity does. Three unrelated candidate use cases across three business units is not a sprint; it is a portfolio exercise, which is what AI product strategy and use-case discovery covers in two to four weeks from $4,250 or ₹2,80,000.
Asking for something clickable at the end adds about two weeks and changes the engagement. That is an AI POC sprint at $6,250 to $10,500, or ₹4,00,000 to ₹6,80,000, and it answers a different question: not whether to build, but whether the hardest part works on your data. The distinction is set out in AI proof of concept vs demo.
Multi-region or multilingual scope adds time in proportion to the number of languages that need a labelled sample. Regulated data adds calendar time rather than effort, because approvals are slower than analysis. And a stakeholder group without a decision-maker adds indefinite time, which is why we ask who signs before we quote.
How the sprint sits against everything after it
Discovery is the shortest engagement we run. Sprint Zero takes ten days, a ProofRun three weeks, and a Launch 6 MVP six weeks. Most scoped builds take eight to sixteen weeks after that, and the AI-accelerated MVP starts at $26,500 or ₹17,60,000. Planned end to end, a company starting from nothing reaches a production pilot in roughly three to five months, with discovery occupying the first two weeks of it.
That ordering is the point. Ten days spent deciding what to build changes the eight to sixteen weeks that follow, and the discipline that makes the later phases hold their dates is described in scope lock.
One planning note that saves arguments later: the sprint fee is credited against the next build, so there is no commercial reason to delay starting while the larger engagement is negotiated. Running discovery during the contracting window for the build is usually the fastest legitimate way to shorten the overall programme.
When ten days is not enough, and we will say so
If your data has never been inspected and lives across more than three systems with no common key, ten days produces a data findings note and no recommendation. That is still useful, but it is not what you paid for, so we would rather scope an integration assessment first.
If the question is genuinely technical rather than strategic, discovery is the wrong shape. Whether a model can extract fields from your specific document set at an acceptable error rate is answered by building something and scoring it, not by analysis. Go to a POC sprint.
And if the organisation cannot free a data owner for two days a fortnight, do not book the sprint yet. A sprint run on documentation instead of data produces confident conclusions from unverified assumptions, which is worse than no sprint. The preconditions are listed in AI readiness assessment.
A real example of the calendar
A hospital network needed appointment booking handled in several Indian languages. Discovery had to cover telephony constraints, language coverage, clinical safety boundaries and what the existing scheduling system would allow. Access to call recordings required a consent review before any listening began, which is what set the calendar rather than the analysis itself.
The sprint effort stayed at ten working days; the elapsed time ran to about four weeks because of that review. The system that followed is described in the multilingual voice agent case study, and the lesson for planning is simple: budget effort in days and calendar in weeks, and never conflate the two in a board update.
Two patterns repeat across engagements like that one. Consent and access reviews are almost always the critical path in regulated sectors, and they are almost always discoverable in advance. Ask, in week minus one, what approvals a stranger needs to read your data, and the answer will tell you your real timeline more accurately than any project plan.
A readiness checklist to protect the ten days
- Sign the NDA and raise data access requests before you agree a kickoff date
- Name the single decision the sprint must unblock, and the person who will take it
- Book all interviews in advance, with named substitutes for each slot
- Confirm a data owner is available roughly one day per week for the fortnight
- Check whether an export path exists for each system in scope, and flag the ones that do not
- Agree who must attend the day ten readout, and diarise it at kickoff
- Decide in advance what a negative recommendation would mean for the budget
Related reading
The AI discovery sprint: ten days to a straight answer describes the engagement from the client side, and the AI discovery sprint service page lists the deliverables and the fixed fee, with every published figure on the pricing page. Nielsen Norman Group's guidance on the discovery phase is a useful primary source on why discovery is time-boxed and what it should produce, and it applies to AI work as much as to interface design.
Plan ten working days of effort and three to four weeks of calendar, and the sprint will end on the day it was meant to.
Frequently asked questions
How long does an AI discovery sprint take from first call to decision?
▾
Ten working days of effort, but three to four weeks of calendar time in most organisations. The gap comes from NDA signature, data access approvals and stakeholder availability rather than from the analysis. Companies that raise access requests on the day they decide to run a sprint routinely finish inside two and a half weeks.
Can an AI discovery sprint be compressed below ten days?
▾
Occasionally, when the scope is one use case, the data is already accessible and the stakeholders are in one building. We have run useful five-day versions in those conditions. Below that the golden set gets skipped, and a sprint without a scoreable evaluation set leaves you no way to judge the build that follows.
What happens immediately after the discovery sprint ends?
▾
You receive a written recommendation, a data findings note, a costed estimate and the golden set, and the day ten readout is where the decision is argued. If you proceed, the fee is credited against the build, and the next step is usually a three-week ProofRun or a six-week Launch 6 MVP.