This page compares two ways of producing the words in a registration. It is not about which firm or product to choose, and it starts with what AI cannot do.
The limits of AI drafting
A language model writes from what it is given. If an experiment, a failed approach or a conclusion never reached a ticket, a pull request or a document, the model cannot know it happened. The bigger danger runs the other way: a model asked to describe a hypothesis will produce a plausible one even when the records do not contain it. That is fabrication, and it is the opposite of what the program requires.
The registration form adds pressure. Every narrative field allows up to 4,000 characters, and six core activity fields have minimums, for example 650 characters for the hypothesis. A tool told to reach a minimum can pad, repeat or inflate the uncertainty. A careful method treats a short field as a sign that evidence is missing, not as a gap to fill with words.
Finally, a tool cannot sign. When applying, the company declares that the information is complete and correct, that it kept appropriate records while doing the work, and that the activities meet the requirements. That declaration is a person's.
The limits of consultant drafting
An interview-led draft is written by someone who can ask why, which is a real strength. Its weakness is timing. Interviews often happen months after the work, and memory fills gaps confidently. The ATO says paperwork prepared afterwards is generally not enough by itself unless some contemporaneous records back it up. A consultant also sees only what people raise in the time booked, so a dead end that taught the team something may go unclaimed, and a routine feature may be described as an experiment because it was memorable.
Four tests for any draft
- Supported. Each statement of fact about a hypothesis, test or result rests on records made at the time. Explanation written later can interpret those records and set out the reasoning; it cannot add an experiment or a result the records do not show.
- Honest about gaps. Missing evidence is flagged for a person to answer, never filled in.
- Not inflated. The uncertainty described is the one the team actually faced, judged against what a competent professional could have known.
- Reviewed by someone qualified. Grouping activities, for example, is only proper where they resolve the same technical uncertainty, and that call needs judgement.
The same field, drafted two ways
An illustrative hypothesis field, written from the same records: a ticket and two pull requests showing a per-tenant cache that missed its latency target in two load tests.
- Beyond the records: "The team hypothesised that a novel distributed caching architecture would cut latency by 80% across all services."
- Within the records: "The team expected per-tenant caching of ranked results to keep p95 latency under 200 ms at 5,000 requests a second. The first two load tests, recorded in the pull requests, did not."
The second is plainer, and each figure in it can be checked. Neither sentence makes the work eligible: that still depends on whether a competent professional could have known the outcome in advance.
Which method suits which company
AI drafting suits teams whose work already leaves a trail, typically software, data and health technology companies using issue trackers and version control. Read how records support claims for training ML models. Interview drafting suits work that happens at the bench, in the field or on paper, and first claims where nobody yet knows what to record.
Rand uses the first method with review on top: AI drafts the registration from GitHub, Jira, Linear and Notion records and Xero payroll, the team reviews it, and an Expert Reviewer finalises it before the company submits. For the commercial side of the same choice, see R&D tax software vs a consultant.