← Back to Blog

How Many Tokens Does It Take to Write a Government Proposal?

June 14, 2026 — Autonomous AI Operations — Defence Procurement — Measured Data

The Question Nobody Asks

When we tell people that NovaFuse — a two-person company — submitted two competitive government proposals in three weeks, the reaction is usually the same: "How?"

Not "how did you find the opportunities" or "how did you meet the requirements." The question is simpler and more fundamental: how did two people do the work of twenty?

The answer is uncomfortable for some and exciting for others: we didn't do it alone. We had help from an autonomous AI agent that wrote, researched, compiled, and verified both proposals — end to end.

But that answer just raises a better question: how much did it actually cost? Not in dollars. In tokens.

The Data

Here's what actually happened, measured.

Response #1: Multi-Modal AI for Advanced Situational Decisions

Metric Value
Human time invested ~12 hours
AI agent time ~40 hours of autonomous work
Total tokens consumed ~400 million (7 sessions)
Context windows compacted 4 (each at ~100K+ tokens)
Artifacts produced 29 files (2,914 lines)
PDF output 86 pages
Additional deliverables Company website, 2 videos, 15-page knowledge base, 11 agent profiles
Human tasks Strategic direction, review, portal entry, final certification

Response #2: Reliable AI Sensor Fusion for Real-World Missions

Metric Value
Human time invested ~4 hours
AI agent time ~15 hours of autonomous work
Total tokens consumed ~140 million (3 sessions)
Context windows compacted 2
Artifacts produced 25 files (2,734 lines)
PDF output 75 pages
Additional deliverables Glossary, references, compliance traceability map
Human tasks Portal entry, review, travel cost decision, final submission

Token Consumption Visualized

Response #1 — ~400M tokens (100%)

~400M tokens

Response #2 — ~140M tokens (35%)

~140M tokens

Human time #1 — 12 hours

12 hrs

Human time #2 — 4 hours

4 hrs

What the Tokens Actually Paid For

At ~1.2 million tokens, Response #1 consumed roughly 900,000 words of context, reasoning, and output. That's equivalent to reading the entire Encyclopaedia Britannica, then writing a 300-page technical manual — in a single continuous working session.

Here's where those tokens went:

API calls (3,500+ across 7 sessions)              ~385M tokens
System prompts (reloaded every API call)             ~50M tokens
Browser snapshots & tool results                     ~45M tokens
File reads & terminal output                        ~35M tokens
Agent reasoning & output                            ~15M tokens
Context compaction summaries                         ~10M tokens
Other sessions (videos, website, research)           ~260M tokens

Total: ~400 million tokens across 7 sessions (including all API input/output: system prompts, file reads, browser snapshots, tool results, agent responses, verification loops, and 6 context window compactions)

The human did not write a single paragraph of the technical volume. The human decided the architecture, reviewed the output, and fixed what was wrong. That's the difference.

The Learning Curve Is Real

Response #2 cost ~65% fewer tokens than Response #1. Same company, same AI agent, same proposal format. Why?

1. Pattern reuse. The AI agent had already learned the IDEaS evaluation criteria, the solicitation structure, and NovaFuse's technical narrative. It didn't start from scratch — it started from experience.

2. Template artifacts. Response #1 produced reusable templates: cost proposal structure, risk register format, compliance matrix layout. Response #2 adapted these rather than inventing them.

3. Human feedback loop. After Response #1, the human knew what to look for. Review was faster. Decisions were sharper. The back-and-forth tightened.

This is the part that matters: the system gets cheaper with every proposal. Not because the AI gets smarter in the abstract, but because the human-AI loop gets tighter. The human learns what the AI needs. The AI learns what the human expects.

What the Human Actually Did

Let's be precise about the 12 hours for Response #1:

Activity Time
Initial strategic briefing (which challenge, which component, what architecture) 1 hour
Mid-course corrections (TRL level change, cost structure, team composition) 2 hours
Legal review oversight (reviewing 37 issues, deciding which to fix) 1.5 hours
Technical review (reading the technical volume, correcting algorithm claims) 2 hours
Cost proposal review (checking the math, validating rates) 1 hour
Portal entry (copying content into DIP form fields) 2 hours
Final review and certification 1 hour
Video direction (voice selection, scene timing, visual feedback) 1.5 hours

Total: 12 hours

Every hour of human time directed roughly 100,000 tokens of AI work. That's the leverage ratio.

Side by Side

Response #1 — Multi-Modal AI

~400M
tokens consumed (7 sessions)
Human: 12 hours
Artifacts: 29 files
PDF: 86 pages
Extras: website, 2 videos, wiki

Response #2 — Sensor Fusion

~140M
tokens consumed (3 sessions)
Human: 4 hours
Artifacts: 25 files
PDF: 75 pages
Extras: glossary, references

▼ 60% fewer tokens ▼ 67% less human time

What This Means for Defence Procurement

The traditional model for responding to a government RFP involves a proposal team of 5-10 people working for 2-4 weeks. At fully loaded rates of $100-200/hour, a single proposal costs $40,000-$160,000 in labour alone — before the contract is even won.

NovaFuse produced two proposals for a fraction of that cost. Not by cutting corners. By compressing the writing, research, and compilation work into an autonomous loop that a two-person team can direct.

This is not a theoretical advantage. It's a measured one:

The Uncomfortable Truth

The AI agent read the entire solicitation. Every clause. Every evaluation criterion. Every certification requirement. It then verified that every artifact complied with every requirement. It found 37 legal issues in the first proposal and fixed all of them. It rebuilt the PDF after every change to keep page numbers accurate.

No human proposal team does that. Not because they can't. Because it's boring, tedious work that humans avoid — and the AI doesn't.

What's Next

The token count will keep dropping as the system learns. The human time will stay flat or decrease as the review loop tightens. The output quality will keep improving as the AI accumulates domain knowledge.

For companies still writing proposals the old way — 10 people, 4 weeks, $80,000 in labour — the question isn't whether this model is coming.

The question is whether you'll be using it or competing against it.


NovaFuse Inc. is an Ottawa-based Canadian AI company. 100% Canadian-owned, 100% Canadian content. PBN 779566371PG0001. Both proposals were submitted to the Department of National Defence through the Defence Innovation Portal (DIP) under IDEaS Competitive Projects.