Decision analytics for small teams

See what your decision depends on.

Fork turns messy business decisions into inspectable models of options, assumptions, unknowns, constraints and breakpoints, so small teams can see what matters, what is fragile, and what to learn next.

Get Early Access →

fork · decision analysis 10,000 scenarios · 4 node types · 9 edges

FRAGILE

EPISTEMIC BREAKDOWN

Known

2 facts · used deterministically

Estimated

4 ranges · sampled in simulation

Assumed

3 beliefs · no evidence

Unknown

2 unresolved · blocking decision

WHAT ACTUALLY DRIVES THIS

Unknown variable A

31%

Assumed variable B

21%

Estimated variable C

13%

Known variable D

Unknown A drives 31% of all instability, and it's unresolved.

FRAGILITY

B wins δ = 0.18 A wins

Preference reverses at a smaller shift than it looks.

UNSUPPORTED CERTAINTY

1 unverified belief

↓

conclusion A conclusion B conclusion C conclusion D

4 conclusions downstream of 0 evidence.

For real business choicesPricing, hiring, launches, channels, build vs buy.

Builds an inspectable modelNot a one-shot recommendation hidden in prose.

Shows what could flipFind the assumptions and thresholds that change the choice.

Tells you what to learn nextTurn the most important unknown into a concrete test.

Fork, in plain English / 01

A decision workspace for teams that don’t have a decision science team.

Give Fork a messy, high-stakes choice. It turns the context into an inspectable model, shows which assumptions are carrying the decision, tests what would make the answer flip, and tells you what evidence is worth collecting next.

Bring the messy decision.

Paste the brief, research, constraints, rough numbers and competing opinions. Fork does not require a clean spreadsheet first.

See what the choice depends on.

Options, objectives, variables, constraints, assumptions and unknowns become explicit, with “known," “estimated," “assumed" and “unknown" kept separate.

Stress-test the current preference.

Fork looks for the nearest threshold, constraint or unsupported relationship that would change which option wins.

Learn before you commit.

Instead of “do more research," Fork identifies the unresolved unknown most likely to change the decision and proposes a concrete test.

Built for: founders, PMs, growth teams and small operators making consequential choices without dedicated strategy, analytics or decision-science infrastructure.

The problem / 02

AI made answers cheap.
Decisions are still expensive.

Enterprise teams buy decision-intelligence software and expensive consulting. Small teams and founders, often making equally consequential calls with less data and less specialist support, get a chat window.

$30K+/yearEnterprise decision software

Six figuresStrategy and analytics consulting

A chat boxWhat most small teams get

What you get from AI chat

Confident answers to uncertain inputs.

The interface optimizes for a useful response. The uncertainty, assumptions and scenario conditions are usually buried inside prose.

Recommendations stated as conclusions, not scenario-dependent behavior

Assumptions hidden inside reasoning, not labeled or inspectable

Qualitative relationships can become invented numeric precision

No direct answer to: what would change this recommendation?

Static reasoning that must be regenerated when inputs change

TYPICAL AI RESPONSEONE-SHOT

“Based on your situation, I’d recommend the guided cohort at ₹2,499. This balances revenue potential with learner outcomes and positions you well for word-of-mouth growth."

OPTION B RECOMMENDED · HIGH CONFIDENCE

What Fork gives you

Uncertainty visible. Fragility shown. Yours to interrogate.

The output is not merely an answer. It is a persistent model whose inputs, assumptions, evidence and failure points remain inspectable.

✓

Every variable classified: known, estimated, assumed or unknown

✓

Qualitative relationships preserved, no invented math

✓

10,000 Monte Carlo scenarios, preference as frequency, not a magic score

✓

Adversarial threshold search to find what flips the decision

✓

Persistent and editable. The same decision evolves as evidence arrives

FORK OUTPUTMODEL V3

₹2,499 WTPUNKNOWN
31% DECISION SENSITIVITY

Guided leads in 68% of scenarios, but flips below 2.7% qualified conversion. This unknown dominates the decision. Resolve it before committing.

A different primitive

Not a smarter answer.
A decision you can inspect.

Chatbots collapse uncertainty into prose. Prompt chains make the prose longer. Fork creates a persistent decision object that can be challenged, recomputed and updated.

Chat / agent defaultADVICE IN, PROSE OUT

Context disappears into a prompt. Assumptions stay implicit. Confidence sounds persuasive. When an input changes, the reasoning is regenerated and trusted again.

MESSY CONTEXT→LLM→RECOMMENDATION

ForkA LIVING DECISION MODEL

AI proposes structure. Humans review it. Deterministic systems test eligible relationships. Every output remains traceable to the model state that produced it.

CONTEXT→TYPED MODEL→TEST→BREAK→LEARN

Intelligence layer / 03

A better decision starts
before the model.

Fork researches the landscape, understands your operating context, and assembles the strategic, product, growth, and decision-science lenses relevant to the choice at hand.

Decision Intelligence Workbench

Should a vertical AI startup move upmarket now, or keep winning SMB?

Company context

NRR 108%Founder-led sales6-month runway3 engineersSupport load rising

Customer evidence

Enterprise feature requestsSMB churn reasonsSales call transcriptsExpansion behavior

Market evidence

Category consolidationBuyer budget shiftsProcurement cyclesAdjacent platform moves

Competitive patterns

9 relevant companiesICP movementPackaging changesDistribution models

Analytical lenses assembled

Strategycapability fit × strategic commitment × reversibility

Productunmet job × evidence of pull × roadmap displacement

Growthsales-cycle economics × expansion × CAC payback

Decision sciencevalue of information × sensitivity × cost of delay

Fork's current read

The decision is not primarily “SMB vs enterprise.”

It depends on whether observed enterprise pull is strong enough to justify the operational commitment required to serve it.

Current evidence supports a narrower move: test one enterprise wedge before repositioning the company.

Why: procurement latency and roadmap displacement dominate downside; expansion behavior among existing larger accounts is the strongest available signal; broad market growth is comparatively low-value evidence for this decision.

Sources synthesized

Customer signals mapped

Analytical lenses selected

Variables proposed

Contradictions surfaced

Relationships provisional

Decision-critical unknowns identified

Company context
+ customer evidence
+ industry research
+ competitive patterns

→

AI deep research + synthesis

→

Decision archetype
+ relevant lenses
+ missing variables
+ contradictions

→

Provisional decision model

→

Human review
+ bounded computation

Fork doesn't assume you already know how to model the decision.

Research broadly. Model selectively.

How it works / 04

Five layers.
No black boxes.

Most tools collapse interpretation and computation into one opaque step. Fork makes each layer explicit, so you can see what the AI proposed, what you approved, and what the code actually ran.

Messy decision

Options, context, constraints, numbers, uncertainties

→

02 · Claude

Decision compiler

Extracts a provisional typed graph. Never silently invents evidence or weights.

→

03 · Human

Model review

Accept, edit or reject AI-proposed variables and relationships.

→

04 · Code

Simulation engine

Validates formulas, samples uncertainty, applies constraints, reruns scenarios.

→

Decision intelligence

Sensitivity, breakpoints, fragile assumptions and learning priorities.

02 · Claude / Decision compiler

Extracts a typed graph from messy context

Reads natural language, rough numbers, competing opinions, unstated assumptions, and proposes options, variables, constraints and relationships. Every object carries an epistemic status. Nothing is invented silently.

Provisional typed output Epistemic status on every object Never fills unknowns silently

→

03 · Human / Model review

You decide what the model is allowed to claim

Before any computation runs, you accept, edit or reject the AI proposals. High-impact assumptions are flagged first. Epistemic honesty is enforced here, not automated away.

Known ≠ Estimated ≠ Assumed ≠ Unknown No silent imputation Partial execution

→

04 · Code / Simulation engine

Deterministic analysis on what you approved

Valid subgraphs run. Unsupported relationships stay blocked. Validates formulas, samples uncertainty, only on what has been explicitly permitted.

Formula validation Uncertainty sampling Conclusion permissions

The integrity contract

The model can know less than the chatbot.

That is a feature. Fork preserves uncertainty instead of laundering it into confidence.

Unknown inputMay block downstream computation. Never silently receives a substitute value.

Qualitative edgeStays visible and labelled. Never becomes an invented percentage.

Simulation resultDescribes model behavior, not stated confidence.

AI-proposed objectBegins provisional. High-impact assumptions reviewed before computation runs.

Four product moments / 05

Don’t tell me the answer.
Show me where it breaks.

Four interactions make the thesis tangible. Each attacks a failure mode of conventional AI advice.

01 / COMPILE

Messy context becomes a model you can inspect.

Claude proposes structure without pretending its interpretation is ground truth. Epistemic status stays attached to every object. The team can challenge a specific node, not a paragraph.

WHY THIS MATTERSThe team can disagree with a specific assumption or input instead of arguing with a wall of AI advice.

COMPILER / PROVISIONAL MODEL

"We have 340 waitlisted users. Product works for 80% of use cases. 2 rough edges remain. Should we launch paid now, build for 2 more months, or run a paid beta? 6 weeks of runway buffer."

340KNOWN · waitlist size

8–18%ESTIMATED · beta conversion

WTP ?UNKNOWN · at current quality

Polish → retentionASSUMED · 4 dependents

02 / ATTACK

Find the smallest change that flips your choice.

Fork does not defend its current preference. It searches for the nearest permissible boundary: a threshold, constraint, or relationship removal that changes which option wins.

WHY THIS MATTERSYou see whether a recommendation is robust, or one small, plausible shift away from being wrong.

BREAK THIS / ADVERSARIAL MODE

NEAREST DECISION BOUNDARY

If beta conversion falls below 11%, waiting 2 months wins.

WAIT & BUILDLAUNCH NOW

03 / REMOVE

Delete one belief. Watch four conclusions lose their basis.

Not every assumption has defensible evidence. Fork removes the relationship and traces exactly which downstream claims survive, and which don't.

WHY THIS MATTERSA persuasive assumption can no longer quietly support a chain of confident downstream conclusions.

UNSUPPORTED CERTAINTY / TRACE

ACTIVE ASSUMPTION

"More polish leads to meaningfully better retention."

Lower early churndirect

Stronger referralsdownstream

Premium pricing justifieddownstream

Lower CAC long-termsecond-order

Current model: waiting retains a long-term retention advantage.

04 / LEARN

New evidence updates the same decision, visibly.

Fork identifies the highest-value unknown, proposes the cheapest experiment to resolve it, then versions the model when evidence arrives. The team learns cumulatively.

WHY THIS MATTERSEvidence changes the model instead of disappearing into another chat thread. The decision history is preserved.

EVIDENCE DIFF / V3 → V4

V3 · BEFORE BETA

Launch is fragile

WTP · unknown

Conversion · 8–18% est.

Launch wins · 52% of runs

→

V4 · 14 DAYS LATER

Launch strengthens

Beta result · 6/20 paid

WTP · evidenced at ₹999

Launch wins · 81% of runs

Live data · collaboration / 06

Feed it anything.
Keep it alive.

Fork connects to the tools your team already uses. Your decision model updates as your data changes. No manual re-entry, no stale assumptions. Run an experiment, watch the model recompute. Share it with your team so everyone reasons from the same model, not the same gut feel.

CRMHubSpot

AdsMeta

AnalyticsMixpanel

RevenueStripe

SpreadsheetGoogle Sheets

WorkspaceNotion

CommsSlack

Any formatCSV / API

CHANNEL MIX DECISION · MODEL V7

Where should we allocate next quarter's growth budget?

LIVE MODEL

Live data feeds

META ADS

₹340 CPL

Paid social, last 30 days

↻ updated 4 min ago

MIXPANEL

6.2% → 8.1%

Trial→paid conversion, this vs last week

↻ updated 12 min ago

STRIPE

₹4.2L MRR

Up 18% vs last month

↻ updated 1 min ago

HUBSPOT

34 SQLs

This week: 11 from content

↻ updated 8 min ago

Decision model: auto-updated nodes

ESTIMATED · AUTO FROM META + MIXPANEL

Paid CAC

₹340 CPL × 4.2 touches = ₹1,428

KNOWN · FROM STRIPE

LTV:CAC ratio

₹12,400 LTV ÷ ₹1,428 = 8.7×

ASSUMED · UNVERIFIED

Content CAC

Estimated ₹380, no clean attribution yet

ESTIMATED · FROM MIXPANEL

Conversion lift from content

8.1% vs 5.9% baseline, small sample still

Live output

Decision state

FRAGILE

Paid leads in 61% of scenarios. Content attribution is the highest-leverage unknown. Resolve it before committing budget.

What drives this

Content CAC assumption

29%

Paid conversion rate

19%

LTV estimate

10%

Learn next

Set up UTM tracking for content leads. Resolve the attribution unknown in 2 weeks. The model updates automatically when data flows in.

Collaboration

The whole team reasons from the same model, not the same meeting.

Share a live Fork model instead of a slide deck. Your co-founder can challenge a specific assumption. Your investor can see exactly what the decision depends on. New evidence updates the model for everyone, instantly.

Shared decision workspace. Invite your team. Everyone sees the same model, the same node statuses, the same simulation. No version confusion.

Challenge specific assumptions. Instead of arguing about a conclusion, a team member flags an Assumed node and proposes an alternative. The model reruns.

Evidence updates for everyone. Run the experiment. Update the node. Simulation reruns. Everyone sees the new decision state, immediately.

Decision history is preserved. Every version is saved. Six months later, see exactly what was Known, Estimated, and Assumed when the call was made.

Where teams use it / 07

Every area where consequential calls get made.

Fork works for any decision where the data is available but scattered, the options are real, and the cost of a wrong call matters. Live integrations make it a tool teams return to daily, not just once.

Growth

Where should this budget go?

Model channel mix, CAC, LTV, and conversion across paid, content, and community. Decision auto-updates as Meta and Mixpanel data flows in.

Meta AdsMixpanelStripe

Product

What do we build next?

Prioritize features by sensitivity × retention impact × dev cost. Mixpanel usage data drives variable values, not gut feel.

MixpanelNotionSheets

Experiments

Did the A/B test change the decision?

Fork versions the model before and after any experiment. See exactly which nodes updated, how the simulation shifted, whether the decision flipped.

A/B resultsMixpanelCSV

Sales

Which deals deserve our attention?

Pull live pipeline from HubSpot. Model deal prioritization across close probability, ACV, strategic fit, and resource cost, live.

HubSpotSheets

Finance & Ops

Runway, headcount, burn: when do we act?

Connect Stripe MRR data. Model timing decisions live as revenue and burn numbers update. The model flags when the decision state changes from Robust to Fragile.

StripeSheetsCSV

Strategy

Market entry, pricing, pivot, or stay.

The highest-stakes calls rarely have clean data. Fork models what you know, what you're estimating, and what you're assuming, and surfaces what matters most before you commit.

Any sourceManual input

The stack / 08

Six primitives.
Not one big prompt.

Typed heterogeneous decision graph

Numeric, ordinal, categorical, qualitative, objective, constraint and outcome nodes coexist without forced arithmetic.

MODEL

Epistemic provenance

Known, estimated, assumed and unknown travel with downstream conclusions.

INTEGRITY

Computational eligibility resolver

Valid subgraphs run; unsupported transformations remain blocked.

RUNTIME

Mixed analytical regimes

Calculation, simulation, constraints, ordinal comparison and qualitative reasoning route separately.

ENGINES

Conclusion permission system

A claim is shown only when its prerequisites are satisfied.

GUARDRAIL

Decision-learning loop

Sensitivity, uncertainty, evidence gap, downstream reach and learning cost shape what to learn next.

LOOP

FORK / DECISION ANALYTICS FOR SMALL TEAMS