Dija Integrates Jev: Bringing Dedicated Decision Models to Sales

Most of what an autonomous sales agent does is judgment.

Every day, your pipeline asks dozens of discrete evaluation questions: Does this inbound email show buying intent? Does this account meet your technical prerequisites? Which contact actually holds budget authority?

Dija's agent architecture has always kept those two kinds of work apart. Reasoning and writing, such as reading the web, researching an account, or drafting outreach, run through one path. Bounded judgments, such as classifying, scoring, gating, and ranking, run through another. Until now both paths were served by chat models: the judgment side prompted a generative model for explanatory prose, which software then parsed back into numbers, enums, and boolean flags. Functional, but slow, expensive, and fragile for what is essentially a typed answer.

Dija now supports Jev, the purpose-built decision model from TypeSafe, as the model behind the judgment path. Nothing was rewritten to make room for it; the routing already existed, and Jev is now the model it can route to.

Instead of generating text, a decision model accepts your source material alongside typed questions such as categorical selections, rubric scores, and boolean evaluations. It resolves all questions in parallel in roughly 100 milliseconds, returning structured answers paired with calibrated probabilities.

Your generative models keep the reasoning path: reading the web, researching accounts, synthesizing briefs, and drafting outreach. Jev serves the judgment path.

Why it matters

Putting a model built for decisions behind the judgment path changes three things:

• Sub-Second Execution: Complex qualification workflows that previously required multi-minute agent sessions now complete in approximately one second. The advanced qualification playbook executes its five scoring steps before you finish reading the initial note.
• Fractional Unit Economics: Jev bills strictly for input tokens at a fraction of a percent of a standard chat model’s cost. Workspaces can now continuously qualify every candidate discovered by scanners rather than rationing evaluation runs.
• Calibrated Uncertainty: Generative models often produce unearned confidence. Jev outputs mathematically calibrated probabilities that reflect actual statistical likelihoods. When a score falls below your confidence threshold, the software can take appropriate action: hold the record, request operator review, or trigger additional search enrichment.

Deterministic business logic remains where it belongs: in code. Scoring weights, threshold gates, tier boundaries, and date bucketing live in a single reviewed file. Scoring rationales are assembled directly from the model’s explicit rubric criteria rather than dictated after the fact by a generative model.

What can run on the decision model

Inbound and intent intelligence

• Inbound reply classification: Evaluates prospect replies across nine discrete categories to instantly identify tone, purchasing intent, and recommended next actions.
• Daily lead prioritization: Filters scanner-sourced opportunities through your workspace qualification gate, evaluates relevance and urgency, merges overlapping signals, and outputs a ranked top-10 queue for immediate outreach.
• Discovery gate validation: Vets each scanner discovery as Pass, Unclear, or Skip against workspace criteria, verifying that primary-source evidence explicitly identifies a real buyer.
• Social signal intelligence (X via Grok): Analyzes raw social activity to filter out general commentary, confirming company legitimacy, account authenticity, active vendor-search intent, and urgency.

Account and opportunity qualification

• Account ICP scoring: Evaluates funding, industry, team size, urgency, segment fit, and historical engagement to generate an aggregate 0 to 100 ICP score and account tier.
• Technical preflight checks: Audits five essential prerequisites (platform, stack, technical team, integrations, and infrastructure), marking each as Pass, Fail, or Unknown before an account advances to deep qualification.
• Contact and buying committee mapping: Scores stakeholders on authority, role relevance, engagement, and warmth (from Champion to Inactive), verifying that the opportunity includes an identified Decision Maker, Champion, and Technical Evaluator.
• Deal readiness and timing: Analyzes budget cycles, trigger recency, contract renewal windows, and stated deadlines to determine an active buying window and classify the account as Hot, Warm, Cold, or Dormant.
• MEDDPICC opportunity qualification: Scores all eight MEDDPICC dimensions against verified account facts to determine whether an opportunity should Advance, Hold, or Disqualify, updating pipeline stages accordingly.

Pipeline strategy and reactivation

• Portfolio and capacity alignment: Evaluates prospective engagements against available delivery bandwidth, internal skill alignment, and client concentration limits before acceptance.
• Stale and lost lead reactivation: Continuously monitors dormant and closed-lost opportunities for meaningful account changes, resolved blockers, and fresh market evidence, sorting accounts into Resurrect, Monitor, or Archive.

Quality control and safety guardrails

• Agent task verification: Audits end-of-session agent outputs to verify that the objective was achieved with concrete specifics, flagging vague responses or unresolved blockers and downgrading unverified runs to Partial.
• Automated score reconciliation: Cross-references an agent's written rationale against its numerical score, withholding publication whenever the qualitative summary contradicts the quantitative rating.
• Risk-weighted approval queue: Scans and ranks outbound message drafts by send risk, flagging unverified claims, explicit pricing commitments, or recipient mismatches so reviewers audit high-exposure drafts first.

Data hygiene and entity resolution

• Intelligent CSV ingestion: Evaluates uploaded column headers alongside sample data to map external files to your CRM schema while discarding non-essential fields.
• Safe entity deduplication: Classifies potential duplicate companies and contacts as Same Entity, Distinct Entity, or Inconclusive, automatically blocking merges whenever distinction or uncertainty exists.
• Committee contact matching: Reconciles detected committee names against existing CRM records, appending attributes to the correct profile rather than creating duplicate contacts.
• Entity validity screening: Screens newly created company records to ensure they represent legitimate, operating commercial businesses rather than generic industry categories or market segments.

How it works in your workspace

The decision model connects directly via your own TypeSafe API key, configured under Settings > Integrations alongside your other model providers.

Workspaces support three operating modes:

• Off (Default): All workflows continue running entirely on your existing configuration.
• Shadow: The decision model evaluates every question beside the chat-model path and logs both outcomes without modifying production data. You can export comparison logs to evaluate alignment, discrepancies, and confidence distributions before routing live traffic.
• Live: The decision model actively handles decisions. If the endpoint is unreachable or misconfigured, the workspace automatically falls back to the previous path and logs the fallback event.

All questions and evaluation thresholds are versioned against specific model releases, ensuring upstream updates never introduce silent scoring drift into your pipeline.

What stays with the chat model

Generative models remain dedicated to what they do best: unstructured reading and synthesis. Open-web research, company dossiers, contact intelligence, strategy briefs, proposals, and personalized outreach drafts all remain with your chat models.

The division of labor is clean:

• Chat models research the web and draft prose.
• The decision model resolves evaluations and outputs calibrated probabilities.
• Dija software enforces deterministic thresholds and arithmetic.

Available now

To get started, add your TypeSafe API key in Settings, switch your workspace to Shadow Mode, and review your evaluation benchmarks over a few days. Once the outputs match your standards, switch to Live.

Every workspace defaults to Off, giving your team full control over when and how you deploy.