Direct answer: no, a company does not automatically need 12 months of “clean” data before it starts. It needs enough evidence for one defined decision, with an observable outcome, a credible baseline, and a release risk the business accepts. A retrieval copilot that cites approved material may begin with dozens of reviewed cases. A model ranking a rare qualified outcome needs mature positives and negatives. A system choosing an offer needs a comparator if the claim is incremental lift. Ask not “how much data do we have?” but “what claim will this workflow make, about which unit, before which action, and what would prove it useful and safe enough?”
The 12-month rule blends three different needs. Calendar coverage may capture seasonality; event volume supports estimation; outcome maturity tells you what actually happened. Twelve months can contain millions of retail orders or twenty enterprise contracts. It may also combine an old pricing policy, a CRM migration, and campaigns that no longer run. More rows do not repair a wrong label, future-information leakage, or the absence of a comparison group. Use the first-workflow plan at https://makinai.co/insights/en/where-start-ai-marketing-90-day-first-workflow to pick the business problem, then require the contract below before promoting automation.
Write a Minimum Evidence Contract
- Decision — the recommendation, classification, or action; for whom; at what moment and in which channel.
- Unit — person, account, order, session, campaign, or asset; do not mix grains without an explicit rule.
- Outcome — the success, failure, and harm event; authoritative source and accountable label owner.
- Maturity — time until the outcome is stable, including delayed revenue, cancellation, return, or churn.
- Window — history needed to cover meaningful cycles, policy changes, and material seasonality.
- Prevalence — frequency of positive, negative, and critical cases; a large total can hide a rare outcome.
- Representation — markets, products, channels, segments, and operating conditions the workflow will face.
- Baseline — current rule, human decision, or non-AI process measured on the same metric and horizon.
- Separation — training, validation, test, and a future holdout when time order matters.
- Release gate — minimum metrics, harm limits, owner, review mode, rollback, and reassessment date.
This one-page contract should precede platform selection or model development. It prevents a vendor from declaring data sufficient after counting records only. It also prevents a valuable rules-led workflow from waiting for a perfect enterprise lake. The broader readiness review at https://makinai.co/insights/en/assess-data-readiness-before-hiring-ai-company still matters; this article scopes a narrower commercial question: the evidence required for one decision and one release.
Separate four workflow archetypes
A retrieval copilot can start with a small set if it uses approved sources, cites them, and does not execute actions. The evaluation set must include real questions, correct answers, refusals, ambiguous requests, and edge cases. Track factual accuracy, coverage, correction rate, and time saved. Document count is not an evaluation set.
A deterministic workflow with AI handling exceptions—such as checking campaign naming and routing discrepancies—can be proven with rules, case replay, and shadow mode. Minimum evidence is path coverage, false blocks, false passes, integration behavior, and rollback. Keep the critical rule in code when possible instead of hiding it in a prompt.
A predictive model, such as conversion propensity, needs examples whose outcomes have matured, positive and negative cases, and a time split that resembles deployment. Google recommends separate training, validation, and test sets; a good test set represents real-world data and does not duplicate training examples. The guidance also notes that ground truth changes. A random split can look excellent while failing on next month's customers.
A causal treatment decision—discount, channel, cadence, or message—needs a comparator. Buyer history describes correlation, not what would have occurred without the treatment. Use a holdout, a randomized test where feasible, or another analytics-approved design. Without it, describe the system as prediction or prioritization, not incremental impact.
Size the window from the outcome backward
Start with label delay. If an opportunity takes 45 days to close and cancellations settle 15 days later, recent cohorts are not labeled. Calling them lost teaches the model an operational delay. Freeze an observation date, wait for the maturity window, and document events that can revise the label.
Then inspect frequency and prevalence. A table with 100,000 rows and 80 positive events is still thin for five countries and ten product lines. Show class counts by segment, uncertainty, and error in high-impact populations. For rare outcomes, combine decisions, simplify the target, or begin with human review before adding model complexity.
Seasonality matters only when it changes the decision. Holiday demand may be essential for retail and irrelevant to an enterprise renewal sequence. If annual coverage matters but only three months exist, do not fabricate it. Restrict the rollout, retain review, and deliberately collect the next cycle. Treat pricing, tracking, consent, channel, and product changes as regime markers; earlier data can provide context without pretending it is the same population.
Treat external thresholds as boundaries, not universal laws
Google's Rules of Machine Learning favors a solid pipeline and simple first approach; it emphasizes that data and features often create more value than premature complexity. Its dataset guidance calls for representative, separate test data. NIST calls for context-specific baselines, measurement, validation, monitoring, and risk tolerance. None offers a magic enterprise row count.
Even large platforms publish recommendations that vary by product. Google Ads recommends evaluating Target CPA over 30 days with at least 30 conversions, while value-based bidding guidance discusses volume, conversion delay, and value quality and cites 15 conversions in 30 days for certain Search cases. Do not transplant an ad-product recommendation into lead scoring, churn, or content review. The variation is the lesson: the decision and its delay determine the evidence.
Start before a custom model, but not before measurement
Week 1: select a frequent, reversible, economically legible decision. Define source, unit, outcome, delay, owner, and baseline. Week 2: extract a time-based sample, reconcile labels against CRM or revenue, and record gaps. Week 3: build the simplest rule and evaluation set; reserve future data. Week 4: run in shadow mode without changing customer treatment. Week 5: compare segments, review errors, and set limits. Week 6: release a small share with human approval, monitoring, and rollback.
Combine utility and harm in the gate. Lead prioritization may use precision at the top, coverage, mature revenue, saved time, and missed opportunities. Content generation needs factuality, brand compliance, rework, and cycle time. CRM decisions need eligibility errors, complaints, opt-outs, and margin. Offline performance is necessary but does not prove business lift; keep the operational baseline and an experiment when treatment changes.
Cost drivers include backfill, identity reconciliation, label definition, review, integrations, a test environment, monitoring, outcome delay, and maintenance. The complete budget model is at https://makinai.co/insights/en/first-ai-marketing-workflow-cost-complete-budget. Acquiring more history can be cheaper than operating errors; elsewhere, labeling a sample and constraining the workflow beats a broad data program. Compare cost per avoided error, time to evidence, and value of learning.
Decide who can approve sufficiency
Marketing owns the decision and expected impact; CRM and Data own unit, extraction, and quality; Analytics owns baseline, split, and comparator; Sales, Commerce, or Service owns the economic label; Privacy owns purpose and minimization; Technology owns execution and rollback; the sponsor owns risk tolerance. A provider may advise, but should not unilaterally approve the dataset used to prove its own work.
The copilot-workflow-agent choice at https://makinai.co/insights/en/copilot-workflow-ai-agent-marketing-what-build-first should follow that risk. More autonomous and irreversible actions require stronger evidence, observability, and approval. If your team is stuck between “not enough data” and “just run a pilot,” bring MAKINAI at https://makinai.co/services/en/ai-strategy-transformation-consulting one decision, an anonymized sample, and the outcome delay. We can scope the Minimum Evidence Contract and the smallest pilot that produces a defensible answer—without selling a magic number of months.