In January 2026, we assumed every documentation team was already automating the first draft. We ran a survey to check. We were wrong, and it's costing them.
So we ran a quick survey across Bangalore, targeting Directors and VPs who run these teams. Fifty-five companies responded. What we found surprised us: most were still relying on human authors to write documentation from scratch, start to finish. Not augmented. Not assisted. Written, end to end, by a person starting from a blank page.
And that number doesn't even count the second-order cost. A disproportionate share of skilled time goes into simply getting to a starting point: locating the relevant source material, reading and interpreting it, and manually turning it into a usable first version. This "cold start" repeats for every work item, in every cycle, across every team. It's high-effort, low-differentiation work, and its cost scales linearly with volume: the more you ship, the more you pay this tax.
If the technology exists, why hadn't more teams automated this? Three reasons kept coming up.
Early AI-generated documentation varied wildly in quality: sometimes excellent, sometimes hallucinated, sometimes technically correct but structurally unusable.
Comparing a writer's fully-loaded cost against token cost sounds simple, but few teams ran the comparison rigorously, especially with quality drift eating into the savings.
Almost every team had tried a script or a prompt template. These stalled because they were isolated experiments, not production-grade systems with real orchestration.
What we built is an agentic pipeline that sits upstream of the writing process, not a chat assistant bolted onto the end of it. Every draft ships with a readiness score, a measure of how completely the output reflects its source, so a reviewer can tell at a glance which drafts are nearly done and which need closer attention.
Poll Jira, Confluence, or any source system for docs-ready work items
Pull the spec, linked tickets, prior docs, and style constraints
Run a retrieval-constrained pass grounded strictly in that source
Attach a readiness score so reviewers know where to look first
On a defined schedule, the agent polls the source system (Jira, Confluence, a PRD repository, or whatever holds the functional specification) and identifies work items that have crossed into a documentation-ready state. For each item, it retrieves the full surrounding context: the spec itself, linked tickets, any prior related documentation, and the relevant style and template constraints. It then runs a retrieval-and-generation pass that is deliberately constrained to that source material, rather than the model's general knowledge, to keep the output grounded in what was actually specified.
The result is a structured, standards-compliant draft: correct heading hierarchy, terminology aligned to the existing doc set, and formatting that matches the target publishing system. Every draft ships with a readiness score, a measure of how completely the output reflects its source, so a reviewer can tell at a glance which drafts are nearly done and which need closer attention.
Critically, the architecture is agnostic at both ends. Swap the input system, swap the output format: the same underlying pattern holds. It isn't a single-purpose tool for one kind of document. It's a general pattern for any repeatable "source to structured first version" step, and it applies wherever that step currently consumes disproportionate skilled time.
Numbers make this concrete. In practice, this looks like turning a 10-hour manual drafting task into about 3 hours of review and refinement: a 70% cut in skilled time per item.
Take a representative AI infrastructure and cloud vendor: 7 documentation teams, averaging 7 people each, 49 writers total.
| Assumption | Value |
|---|---|
| First-pass drafting time today (X) | 10 hours / item |
| Drafting time with agent-generated first pass (Y ≈ 30% of X) | ~3 hours / item |
| Time saved per item (X − Y) | ~7 hours |
| Items drafted per writer, per two-week cycle | 3 |
| Items per cycle, across 49 writers | 147 |
| Cycles per year (bi-weekly) | 24 |
| Total Capacity Saving | = ~₹2.73 crore/year |
Run the math: 147 items × 7 hours saved = ~1,029 skilled hours returned every cycle. Across 24 cycles a year, that's roughly 24,700 hours annually, the equivalent of about 12 full-time writers' worth of capacity, freed up without adding headcount.
At an average fully-loaded tech writer cost of ₹23,00,000/year (≈ ₹1,106/hour), that capacity is worth approximately ₹2.73 crore a year for this one vendor: capacity they can redirect to backlog reduction, higher-complexity writing, or simply shipping documentation faster.
This is intentionally a moderate, defensible estimate. Real numbers will vary by spec complexity, team maturity, and how aggressively review time can shrink, which is why we built a way to run this with your numbers instead of ours.
Pick the numbers closest to your team. The panel below updates live, in both rupees and dollars.
This is a first-pass accelerator, not autopilot.
Outputs are grounded strictly in their source material. The readiness score is a triage aid, not a correctness guarantee. Nothing is finalized or published without human review and sign-off. People stay in control of accuracy and approval. We're removing the drudgery of the cold start, not the oversight that makes documentation trustworthy.
Every company that responded to our survey is sitting on some version of that ₹1.52 crore number, a cost that's easy to miss because it's spread across every writer, every sprint, every spec. Run the calculator above and you'll have your own figure in seconds.
The blank page, the cold start, is a cost you pay countless times per cycle, across every team, in every sprint. Automating it means your people spend their expertise where it actually moves the needle: judgment, verification, and the writing only they can do.
business@thegraas.com