Guide · Published November 14, 2026 · Reviewed November 14, 2026 · Ingenious AI team · Not legal advice
Colorado Developer Duties: The Technical Documentation You Owe Your Deployers
The short answer: Colorado's ADMT framework speaks to both sides of the market. Deployers owe the person-facing duties; developers — businesses that build or sell the tool — owe their deployers two things, per the kit's checker data: technical documentation covering the tool's intended uses, the categories of training data, its known limitations, and instructions for appropriate use and human review; and notice of material updates. Developers and deployers share one more duty: records that show compliance, kept for at least 3 years. This guide is for the developer side of that exchange.
Check your duties and dates — free Build your first notice — free
What is a developer, here? The business that builds or sells an automated decision-making tool that deployers use for consequential decisions — as distinct from the deployer who uses it. If your company sells screening, scoring, ranking, or assessment software into Colorado consequential decisions, this guide and the checker result's developer note apply to you. The deployer side of the relationship is at Vendor vs Deployer: Who Owes the Notice?
Documentation item one: intended uses, stated as boundaries
Intended uses means the decisions and contexts the tool is built and warranted for — and, by implication, the ones it is not. Write it as deployers must use it: which decision types (hiring, screening, lending support), which populations, which deployment patterns (score-only versus score-plus-review). A deployer's purpose statement in their risk assessment is only as good as this document; if your intended-uses statement is marketing prose, every downstream assessment inherits the vagueness.
Boundaries are the valuable half. 'Not intended as the sole basis for a housing denial' is an intended-use statement a deployer can build a workflow around — it tells them the human step is structural, not decorative. Silence about boundaries leaves each deployer to discover them from a person's complaint.
Documentation item two: categories of training data
State the categories of training data — the checker data's word is categories, not the data itself: what kinds of records, from what kinds of sources, covering what periods and populations. Deployers need categories for two duties of their own: the data element of California's risk assessment and the data-categories field of Connecticut's written notice both draw on what the tool is and was trained on, as distinct from what it consumed in a given case.
Categories also carry the representativeness facts a deployer cannot see from outside: which populations the training data covers well and which it covers thinly. That is the bridge to the third item — a category list without its coverage notes invites deployers to assume a universality no training set has.
Worked example: Clearpath Screening Systems
Example — fictional business, real duty structure. Clearpath is a fictional vendor selling tenant-screening software to Colorado landlords. It built its developer-duty pack for the January 1, 2027 start:
- Intended uses: a two-page statement naming tenant screening as the intended decision, the populations covered, and the express boundary — not intended as the sole basis for a denial; a human decision step is structural.
- Training-data categories: 6 categories listed with sources and coverage notes, including the thin-coverage populations (applicants with under two years of address history) named as such.
- Known limitations: 5 limitations written operationally — each with its deployer consequence, including a doubled second-review sampling rate for thin-file applicants.
- Use and review instructions: 11 numbered instructions, from reviewer-role configuration to the monthly 5% adverse-outcome second-review sample.
- Updates and records: material-update notices now go to all deployer accounts with a 30-day pre-production window; documentation versions and notice dates are retained on a 3-year schedule — this quarter, 2 material updates were notified this way.
Documentation item three: known limitations, written plainly
Known limitations is the section vendors resist and deployers need most: the conditions under which the tool's output degrades — sparse-history applicants, populations underrepresented in training, input data of poor quality, decision types at the edge of the intended uses. Written plainly, with the operational consequence of each: what the deployer should watch, sample, or override more readily when the limitation is in play.
Limitations documentation is also what makes reviewer training real. California's carve-out requires reviewers who know how to interpret the output; a reviewer who knows the tool's limitations interprets differently from one who has seen only its accuracy headline. If your documentation helps deployers train reviewers, it is doing the work this item exists for.
Documentation item four: instructions for appropriate use and human review
The final item is instructions: how the tool should be deployed and reviewed to be used appropriately — the review design you warrant, the sampling or audit cadence you recommend, the configurations that void your assumptions. Deployers build their queues and safeguard lists from this; where your instructions specify a review step, a deployer who skips it has been told, in writing, what the tool's appropriate use requires.
Write instructions as instructions — numbered, testable, versioned — rather than as principles. 'Reviewers should be able to overturn outcomes' is a principle; 'configure the reviewer role with outcome-change permission, and sample 5% of adverse outcomes monthly for second review' is an instruction a deployer can implement and an auditor can check.
Material updates, and the shared 3-year record
Notice of material updates is the second developer duty: when the tool changes materially — model, inputs, scoring, limitations — deployers must hear about it, because their assessments, notices, and reviewer training were written against the previous version. Good practice gives the notice content (what changed, what it affects, when it reaches production) and a window before production effect, so deployers can re-run the affected duties rather than discover the change in their outcomes.
The shared floor is records: Colorado's checker states that developers and deployers must retain records that show compliance for a minimum of three years. For a developer, that is the documentation versions, the update notices, and their dates — the evidence that each deployer received what the duty requires, when they required it. Version your documentation the way deployers version their notices, and keep the versions for the same three years.
- Write intended uses as boundaries: decisions, populations, patterns covered — and the uses expressly not intended
- List training-data categories with sources, periods, and honest coverage notes per population
- State known limitations operationally: the condition, the degradation, and what the deployer should do differently
- Give numbered, testable instructions for appropriate use and human review — configuration, sampling, cadence
- Notify material updates with content and a pre-production window, and retain documentation versions and notices for at least 3 years
Sources, as carried in the kit's own data: the duty summaries in app/rules.py and the checker data in app/checker_data.py, verified against primary or authoritative sources on October 28, 2026. Run the free duties checker for your jurisdiction, decision type, and the tool's role, and compare the kit's plans for the record, builder, and review queue. More from Ingenious AI is on the hub.
Related guides
- Vendor vs Deployer: Who Owes the ADMT Notice?
- Colorado AI Act Automated Decision Notices: What to Send
- Adverse-Outcome Explanations in Plain Language
- ADMT Deadline Tracker: Which State Duties Land When
See all ADMT guides for the full set.
Frequently asked questions
What documentation does a Colorado ADMT developer owe deployers?
Technical documentation covering intended uses, categories of training data, known limitations, and instructions for appropriate use and human review — plus notice of material updates, per the kit's checker data.
Must developers share their actual training data?
The duty as stated concerns categories of training data — kinds, sources, and coverage — not the dataset itself.
What counts as a material update?
A change that could affect how the tool should be used or reviewed: model changes, new input categories, changed scoring behaviour, or new known limitations. Deployers' assessments and reviewer training are written against the previous version, which is why notice is owed.
Do developers have a records duty too?
Yes. The checker data states developers and deployers must retain records that show compliance for a minimum of three years — for developers, documentation versions, update notices, and their dates.
Does good documentation transfer our deployers' duties to us?
No. The person-facing duties — notice, explanation, correction, review — remain the deployer's. Documentation supports their discharge of those duties; it never absorbs them.