Back to AI GatewayHome
AI Gateway · Model guide

Every model in an engagement, explained plainly.

No black boxes, no magic. These are the five kinds of models we bring into a Netyield.ai engagement — what each one does, the data it needs, and exactly how your subject-matter experts train it to their standard.

The five models we deploy

One engagement usually runs three or more of these.

Reasoning & drafting model

Large language model

The workhorse of every engagement. It reads contracts, quotes and deduction histories, diagnoses where margin leaks, and drafts the work product your experts used to build by hand — pricing proposals, playbook steps, opportunity briefs.

Data it needs
  • Past quotes, contracts and deal records
  • Deduction, chargeback and discount histories
  • Your pricing rules, product catalog and SOPs
  • Your experts' written playbooks and judgment notes
How your SMEs train it
  • AI consultant encodes expert judgment into prompts and rules
  • SMEs review every draft — corrections become reusable examples
  • Best-reviewed outputs build a curated exemplar library
  • Prompt and rule changes go through change control, like any validated system

Forecasting model

Time-series & regression

Predicts the numbers your teams plan against: demand by channel, price response, churn risk and capacity yield. Output always carries confidence ranges — never a single false-precision number.

Data it needs
  • 24+ months of transaction and order history
  • Promotional calendars and price-change records
  • External signals your planners already use (market, season, access changes)
  • Known one-off events flagged by your team
How your SMEs train it
  • SMEs label the drivers behind past swings — launches, shortages, tenders
  • Forecasts are back-tested against outcomes the experts lived through
  • Planners flag misses monthly; every flag retrains the model
  • Accuracy is reported in plain language: where it's sharp, where it's not

Knowledge retrieval model

Embeddings & semantic search

Turns thousands of contracts, pricing memos and policy documents into something the team can actually ask questions of — and gives every answer a citation back to the source document.

Data it needs
  • Contracts, amendments and rebate agreements
  • Pricing memos, policy documents and SOPs
  • Regulatory filings and payer rules where relevant
  • A 'golden set' of documents your experts trust
How your SMEs train it
  • SMEs curate the golden set — what's authoritative, what's superseded
  • They tag which sources win when documents conflict
  • Every answer is rated on faithfulness to its citation
  • Document refresh cycles keep the index current and auditable

Anomaly detection model

Statistical & pattern models

The always-on watchers. They scan order logs, chargebacks and rebate claims for leakage signals — discount creep, price gaps, duplicate claims — ranked by revenue impact, not by noise.

Data it needs
  • Order logs, invoices and credit memos
  • Chargeback and deduction records with reason codes
  • Rebate and royalty claims
  • Your experts' list of known, accepted exceptions
How your SMEs train it
  • SMEs confirm or reject each flag — rejections are as valuable as confirmations
  • Thresholds tuned to your risk appetite, documented and versioned
  • Known exceptions are whitelisted so the model stops crying wolf
  • Every cleared exception keeps its evidence trail for audit season

Next-best-action model

Recommender

Points reps and account teams at the highest-yield move: which customer to call, which account is at churn risk, which offer fits which segment. Your specialists approve every recommendation before it reaches the field.

Data it needs
  • CRM activity — calls, meetings, outcomes
  • Win/loss records and deal outcomes
  • Customer segments, tiers and entitlements
  • Past recommendations and what happened next
How your SMEs train it
  • SMEs run win/loss reviews that become labeled training cases
  • Field feedback on every recommendation is captured, not discarded
  • Recommendations are A/B tested against current practice before rollout
  • Adoption and lift are measured in revenue, not clicks
How SME training actually works

One loop, every model, every month.

1

Extract

The AI consultant interviews your SMEs and turns their judgment — which deals leak, which forecasts broke, which exceptions matter — into rules, labels and examples.

2

Teach

That knowledge is encoded into the model: curated prompts, labeled training cases, golden document sets and tuned thresholds. Nothing is trained on data you haven't reviewed.

3

Correct

Your experts work with the model daily. Every correction they make — a rejected flag, a rewritten draft, a missed forecast — is captured as a new training example.

4

Retrain

On a documented cycle, corrections flow back into the model through change control. Each retraining run is versioned, tested and logged for SoX and GxP audit.

The loop never stops at go-live. Your experts' corrections are the training data — so the model gets sharper the longer it runs inside your walls, and the knowledge stays yours.

See which models fit your business.

Describe your business area and revenue challenge — we'll map the models, the data and the training plan your experts would run.

Try the AI opportunity finder Start an inquiry