A 2-day AI scoping workshop is a forcing function, not a softer version of a 2-week engagement. Sixteen hours, one senior engineer in the room the entire time, and a founder who arrives with three things — a one-page product description, sample data, and a domain expert. What walks out is five named files: a capability probe, a v0 eval set, a cost-per-query worksheet, an architecture sketch, and a signed go-or-no-go memo. Anything less is a slide-ware day. The 2026 price band is $8K to $15K, typically $12K, and the workshop is only the right purchase when three conditions hold.
This is the buyer-side operating manual: the hour-by-hour agenda, the artifacts that ship by Tuesday evening, the founder pre-work that determines whether the workshop produces real outputs, and the scenarios under which the workshop is correctly the wrong purchase. Companion reading: the idea-to-product manifesto, the eval-first build playbook, the AI feature scoping services worksheet, and the longer-form AI agency discovery week.
Why a 2-day workshop exists as a category
Most founders cannot defensibly commit to a $15K two-week scoping engagement until they have an afternoon of evidence the project is worth scoping at all. A 2-day workshop is the smallest possible eval-bound engagement that produces a defensible go-or-no-go — sitting below a 5-day discovery week and bypassing the four-week deck-driven scoping most agencies still sell.
The category also fails most often in the market. Default failure mode: a “2-day AI strategy workshop” turns out to be eight hours of facilitated discussion, a Friday slide deck, and a soft hand-off to a $150K build proposal. A real workshop ships code, runs models against sample data, and writes a recommendation a domain expert can argue with.
The structural test is who codes during the 16 hours. If the agency staffs the room with a facilitator and a strategist, the workshop is theatre. If a single senior engineer is in the room the full two days — and the founder can watch the capability probe run on a laptop in the third hour — the workshop is engineering.
Day 1: capability inventory and task taxonomy
Day 1 converts the founder’s product description into a structured map of what the model needs to do and a numeric probe score against a candidate frontier model. Two files commit by end of day: capability-probe.md and task-taxonomy.md.
9:00–11:00 — Product narrative. Senior engineer and founder write a 600-to-900-word narrative live in a shared editor. Prose, not a slide. It names the user, the workflow today, the failure mode that triggered the project, the constraint envelope, and the criterion under which the product is unambiguously successful. If the narrative cannot be tightened in 120 minutes, the workshop has surfaced its first finding: the founder does not yet know what they are paying to build.
11:00–13:00 — Task taxonomy. The product decomposes into discrete tasks the model performs. Each is graded on input type, output type, and the consequence of a wrong answer (recoverable, user-visible, regulatory). A typical SaaS AI feature decomposes into four to eight tasks. The taxonomy commits as task-taxonomy.md — a table with rows for tasks and a severity column. This file is what the Day 2 eval set is built around.
13:00–14:00 — Frontier-model shortlist over lunch. Two candidates picked for the afternoon probe. The default 2026 shortlist is Claude Sonnet 4.6 as primary, GPT-5 as the second-vendor pricing benchmark, with Gemini 2.5 Pro as a third probe when vision or long-context workloads are in scope. The pick records in the narrative with one paragraph of rationale.
14:00–16:00 — The capability probe. The senior engineer writes a small Python harness — 80 to 200 lines — that runs five to ten representative inputs through each candidate model and produces a numeric score on a rubric the domain expert helped draft over lunch. Typical scores: 0.7-to-0.9 on simple extraction or classification, 0.4-to-0.7 on generation with domain-specific constraints, below 0.4 on planning or multi-step agent flows without architecture work.
16:00–17:30 — Findings and Day 2 setup. The room reads probe results together and answers three questions: which tasks the model can perform at acceptable quality today, which need a routing layer or retrieval scaffold, and which are out of scope. The senior engineer drafts the Day 2 sampling rationale and commits day-2-plan.md. The founder takes the evening to pull 40 to 60 ground-truth examples from production data.
Day 2: eval set, cost model, and the go-or-no-go
Day 2 converts Day 1’s evidence into a graded eval set, a cost model the founder can take to a board meeting, and a written recommendation with a defensible call. Three artifacts ship: eval-set-v0.csv (with results), cost-per-query.xlsx, and go-no-go-memo.md.
9:00–11:30 — Eval set construction. Senior engineer and domain expert build a 30-to-50-case eval set from the founder’s overnight examples. Each case has an input, an expected output, and a rubric score breakdown. Cases are weighted by severity per the Day 1 taxonomy. This is the v0 of what a longer-form scoping engagement would expand to 80–150 cases.
11:30–13:00 — Eval run and two-grader scoring. The eval set runs against the primary model. Senior engineer scores half the outputs, domain expert scores the other half independently, and 10-to-15 cases are double-graded to compute inter-rater agreement. A first-pass IRR above 70% is acceptable for v0; below means the rubric needs sharpening — itself a finding. Scores roll up into eval-summary.md with three headlines: overall pass rate, pass rate by task, and the three worst failure modes by tag.
13:00–14:00 — Architecture sketch over lunch. A one-page list of named components — model call layer, fallback path, retrieval scaffolding if needed, logging requirement — with a one-line description of each. Not a system diagram. This is what the afternoon cost model is built around.
14:00–15:30 — Cost-per-query model. A spreadsheet that models the unit economics. Inputs: average input/output tokens, expected calls per session, sessions per active user per month, model price per million tokens. Output: cost-per-active-user per month, with a sensitivity table on the three variables that move it most. The number that matters: cost-per-active-user as a percentage of expected ARPU. Above 30% the economics are fragile; above 60% the product is unshippable as designed.
15:30–17:00 — Go-or-no-go memo. A two-to-three page memo with a signed recommendation. Three possible answers: go (build now; here is the shape and the path from v0 eval set to production bar), no-go (do not build; here is what needs to be true first), or build-after-conditions (build once these two things resolve; here is the eval bar the resolved version must clear). Signed by the senior engineer. The founder reads it in the room.
17:00–17:30 — Handoff. Repository goes to the founder. Five files plus a README naming what each file is and what the v0-to-v1 expansion costs as a separately-scoped engagement.
A workshop that ends without a written go-or-no-go has failed. The most common vendor failure mode is delivering “options” and a “next-steps proposal” instead of a recommendation. The founder paid for the recommendation. Insist on it in writing.
What walks out: the five named artifacts
The named-file list is the procurement contract. If a workshop SOW does not enumerate them, the workshop is selling something else.
| File | What it is | What it costs the buyer if missing |
|---|---|---|
capability-probe.md | Numeric probe scores for 2 candidate models against 5–10 representative inputs | No defensible answer to “which model and why” — every downstream build vendor picks their preferred stack |
task-taxonomy.md | Table of 4–8 model tasks graded by input/output type and severity | The build SOW has no coherent acceptance criterion — every task is graded against itself |
eval-set-v0.csv | 30–50 case eval set with rubric scores and inter-rater agreement | ”Production-ready” is a marketing word the build vendor self-attests |
cost-per-query.xlsx | Unit-economics model with sensitivity table | The founder discovers in month 5 the product cannot make gross margin |
go-no-go-memo.md | Two-to-three page signed recommendation: go, no-go, or build-after-conditions | The founder bought permission to spend, not permission to walk away |
A defensible workshop SOW names all five by file extension. If the proposal uses prose synonyms — “a roadmap”, “a feasibility report”, “alignment on capabilities” — the vendor is preserving optionality on what they actually ship. Ask for the file names. The same procurement logic appears in stop scoping AI features in user stories.
What the founder brings to the room
Pre-work determines whether the workshop produces a real eval set or a hypothetical one. Missing any of the three items collapses the workshop into a planning meeting for a longer engagement.
A one-page product description. Not a deck. A page naming the user, the workflow today, the failure mode, the constraint envelope, and what success looks like. If the founder cannot write this page beforehand, two of the 16 hours are spent extracting it — and one of the five artifacts gets cut.
Sample data — 100 to 300 representative rows. Real production data, anonymised as needed. The domain matters less than the realism: a small set of real conversations beats a large synthetic set. Without sample data, the capability probe runs against fabricated inputs and the eval set is hypothetical.
A domain expert in the room for both days. The highest-impact commitment the founder makes. The expert converts a rubric from “what an engineer thinks correct looks like” to “what an experienced operator scores as correct.” Required attendance: Day 1 narrative, lunch rubric drafting, Day 2 eval-set construction, and Day 2 scoring — roughly 6 to 7 hours total.
Two secondary items: frontier-model API access ($100 credit each on Anthropic and OpenAI — the workshop runs $20 to $50 of spend), and a finance contact who can validate the ARPU assumption feeding the cost worksheet.
What it costs in 2026
The honest 2026 cost band is $8K to $15K, with $12K as the typical centre.
| Line | 2026 range | What drives it |
|---|---|---|
| Senior engineer, 16 hours | $4K–$8K | U.S. senior-engineer day rates at $2K–$4K; EU often half |
| Domain consultant or grader hours | $1K–$2K | Domain difficulty; in-house expert drops this to zero |
| Frontier API spend | $50–$150 | Output-token-heavy tasks land at the top |
| Project orchestration + writeup | $1K–$2K | One coordinator runs the calendar and writes the memo polish |
| Workshop facilities / video | $0–$500 | Remote runs to zero; in-person adds a room booking |
| Vendor margin and risk buffer | $1.5K–$2.5K | Standard professional-services markup |
| Total | $8K–$15K (typical $12K) | — |
Ranges are triangulated against Stack Overflow’s 2024 developer survey, publicly posted agency rate cards (Designli, Markovate, Apexon), and BCG’s 2025 AI build cost benchmark — a 2-day workshop compresses on the lower end of BCG’s $25K–$60K six-week discovery band by roughly four-to-one.
Below $8K typically means junior-rate labour or a cut artifact list. Above $15K usually means scope-creep into Day 3 prototyping. What the $12K does not buy: a working prototype, a full architecture document, a system diagram, or any code commit beyond the capability probe and eval harness. Those belong in a 2-week paid scoping at $10K–$25K (see the scoping services worksheet) or M1 build at $30K–$60K.
One pricing pattern to recognise: some agencies “credit” the workshop fee against a downstream build SOW. That is not a discount — it is a lead-gen offer, bundled below cost on the assumption you will sign the build. Walk if the workshop is not priced to stand alone.
When 2 days is enough, and when to upgrade
Three scenarios in which a 2-day workshop produces a defensible go-or-no-go, and three in which it cannot — in which case the founder should upgrade to a 5-day discovery week or a full 2-week scoping.
When 2 days is enough:
- Single-task AI features. Classification, extraction, or generation against a clear input type with a domain expert who can grade correctness. The eval set is tractable in three hours of construction.
- Frontier-model-only stacks. Architecture is a single model call (plus prompt and minimal retrieval). No agent loop, no tool routing, no fine-tune. The capability probe carries most of the model-selection weight.
- Founders with a domain expert in-house. The most expensive workshop line — the grader — is supplied by the founder. The senior engineer’s 16 hours are conserved for capability work and eval construction.
When to upgrade past 2 days:
- Multi-step agent flows. When the product requires the model to plan, call tools, and recover from intermediate failures, the eval set must cover trajectories, not just final outputs. That is a week of work, not an afternoon.
- Regulatory or compliance footprint. When refusal-policy design, audit-logging, or PII handling is on the critical path, the fallback design alone takes a day. The workshop’s Day-2 afternoon does not fit it.
- Cross-functional decisions involving data engineering. When the model is only viable after a data pipeline is built or a retrieval layer is in place, the workshop produces a no-go for the wrong reason — the model can do the task; the data system cannot feed it.
The cleanest test: write a single sentence describing the AI feature. If the sentence has one noun-verb-object structure, the 2-day workshop fits. Two or more clauses linked by “and then” or “while also” — the workshop is too short. Either expand the engagement or scope the feature smaller.
Vendor diagnostic questions
Eight questions a founder asks before signing the workshop SOW. The vendor’s answers — not the brochure — tell you what you are buying.
- Who specifically is the senior engineer in the room for the full 16 hours? Get a name and a LinkedIn. “A senior engineer” without a named individual means the room gets backfilled with a junior.
- What is the exact file list that ships at end of Day 2? A real workshop names files. Brochure language is a tell.
- Will the eval set use my data or hypothetical examples? The answer must be “your data, 100-to-300 rows by Day 1 morning.”
- Which two frontier models will the capability probe run? A vendor who refuses to name models has not designed the probe. One model named means a preferred stack is being preserved.
- What inter-rater agreement bar do you score the v0 eval set against? A real answer is a number between 0.7 and 0.85. “We score it carefully” is not a number.
- Can I see a redacted sample go-or-no-go memo from a prior workshop? A vendor with a portfolio shares one. A vendor who deflects has not written one.
- What is the workshop fee if I do not sign a downstream build with you? “It doubles” or “we do not offer it that way” — walk.
- What happens if the memo recommends no-go? The answer must be “we sign and deliver it; you keep the artifacts.”
Eight clean answers means a real workshop. Five or fewer means theatre dressed in engineering vocabulary.
FAQ
How much does a 2-day AI scoping workshop cost in 2026?
$8K to $15K, with $12K as the typical centre. The split: $4K–$8K for the senior engineer (16 hours at U.S. day rates), $1K–$2K for the domain consultant or grader, $1K–$2K for orchestration and writeup, $50–$150 for frontier API spend, and $1.5K–$2.5K vendor margin. Below $8K means junior-rate labour or a cut artifact list. Above $15K usually means Day-3 scope creep.
What is the difference between a 2-day workshop and a 2-week scoping engagement?
The 2-day workshop produces a v0 eval set (30-to-50 cases), one architecture sketch, and a signed go-or-no-go on a single-task AI feature. The 2-week scoping produces an 80-to-150-case eval set, a full capability map with probe scores across two-to-three frontier vendors, a fallback design with refusal taxonomy, and a complete cost-per-query worksheet. Buy the workshop when the feature is single-task and the founder has a domain expert; buy the 2-week scoping when the feature has multi-step agent flows or a regulatory footprint.
Can I do an AI scoping workshop in-house?
Sometimes. A technical founder with prior eval-design experience and a domain expert on the team can produce the five artifacts in a focused two-day pair-programming sprint. A non-engineer founder usually cannot — the capability probe and eval harness require frontier-model-API fluency. The honest test: can you write a Python harness that runs 10 inputs through Claude Sonnet 4.6 and scores the outputs against a rubric this afternoon? If yes, run the workshop in-house. If no, pay for it.
What if the workshop ends with a no-go recommendation?
That is the workshop working as designed. A vendor who refuses to write a no-go is selling permission to spend, not procurement protection. The artifacts you take home — the capability probe, the task taxonomy, the eval-set scaffold — tell the next vendor or the next model release exactly what bar to clear. Treat a $12K no-go as a $150K build mistake averted.
Should the same vendor that runs the workshop also build the product?
They might, but the engagement does not require it. The five artifacts are portable — any builder downstream can grade their work against the v0 eval set and the go-or-no-go memo. A vendor who insists the build must follow the workshop is preserving margin. The artifacts belong to the founder.
What sample data do I need to bring?
100 to 300 representative rows in a CSV or JSON file, anonymised as needed. Real production data beats large synthetic sets. The data must cover the input types the model will see. Pre-launch products bring 50 to 100 hand-curated examples from competitor products, internal tooling, or domain experts. Hypothetical data produces a hypothetical eval set, and the go-or-no-go memo will say so.
What if my domain expert cannot attend both full days?
The expert must attend at least four of the eight blocks: Day 1 narrative (9:00–11:00), Day 1 rubric drafting (13:00–14:00), Day 2 eval-set construction (9:00–11:30), and Day 2 scoring (11:30–13:00) — roughly 6 to 7 hours across the two days. Outside those blocks the senior engineer runs probes, builds the cost model, and drafts the memo without expert presence. Zero hours of expert time is a workshop killer.
How does a 2-day workshop compare to a 5-day discovery week?
The 5-day discovery week costs $25K–$40K and adds a full data audit, a runnable prototype against sample data, an architecture decision record, and an eval-bound proposal for the build. The 2-day workshop is a tighter forcing function on a smaller surface area. Choose the workshop when you need a go-or-no-go before committing to discovery. Choose the discovery week when the go is already implied and you need the build proposal.
Ready to scope an AI feature in 48 hours? Book a 30-minute idea review — we will tell you on the call whether a 2-day workshop, a 5-day discovery week, or a 2-week scoping engagement is the right shape for your project, and what the founder pre-work needs to look like.
Arthur Wandzel
