Agentic AI · Phase III · Vector

Agentic AI

V4 Phase III · Vector f(σ²,λ,ε) on identified literature · not cold retrieval

The workhorse of the pipeline — automating screening, extraction, and repetitive tasks on literature that has already been identified. Not a search engine. An engine of precision labour.

Autonomous task execution on a defined corpus

Agentic AI refers to AI systems that can plan, execute, and iterate through multi-step tasks without continuous human direction. In evidence synthesis, this means working through a pre-screened corpus of documents to extract specified data fields, apply inclusion/exclusion criteria at full-text level, and populate extraction templates — repeatedly, consistently, at volume.

The critical qualifier in the original table is on identified literature. Agentic AI does not retrieve cold literature. It operates downstream of Perception AI, on a corpus that has already been assembled and relevance-filtered.

The repetitive backbone of Phase III

01 Full-text screening — applying eligibility criteria to full PDFs, not just abstracts
02 Data extraction — pulling ICER values, sample sizes, outcome measures, DALY estimates, costing methods
03 Quality assessment — running GRADE, Cochrane RoB, or CHEERS checklists across included studies
04 Cross-study consistency checks — flagging duplicate data, discordant outcome reporting
05 Table population — filling standardised extraction templates at scale
WHO India · Practical example

For a UHC benefit package review, 140 full-text papers must each yield: intervention cost, comparator, ICER, population, setting, and perspective. Agentic AI populates a structured extraction sheet for all 140 papers in hours — a task that consumes weeks of human effort and introduces inter-extractor variability at every step.

Data extraction is where most reviews break down

Phase III (Vector, f(σ²,λ,ε)) is where the most consequential errors in evidence synthesis occur — not because reviewers are careless, but because repetition degrades precision. A team extracting data from paper 87 of 140 under deadline is not performing at the level they were at paper 12.

Agentic AI does not fatigue. It applies the same extraction logic to every document. This is not a minor efficiency gain — it is a systematic quality improvement that changes what is possible in a six-week evidence brief cycle.

Reliable on structure, unreliable on judgment

Agentic AI excels at extracting clearly labelled fields and fails on papers with non-standard reporting, ambiguous outcome definitions, or methods described narratively rather than in tables. It will extract what it finds — and confidently report absence as absence, even when the data exists in a footnote it did not parse.

Human review of a random 10–20% sample of agentic extractions is not optional — it is the quality control mechanism that makes the rest trustworthy.