Either someone has sold you agents and you cannot tell what makes them different from automation, or your pilot works in the demo and fails internal audit. Both questions have the same answer, and it is about which steps can survive a system choosing its own sequence.
In your words rather than ours.
This is the whole page. An agent decides what to do next; most processes require the same input to give the same output — not only under a regulator, but in tendering, inspection and invoice matching, wherever someone can disagree with the outcome and ask how it was reached. That conflict is structural and no prompt resolves it, so autonomy gets spent where it buys something and withheld where it costs you the answer.
One test resolves most cases: does a person check the output, or is the cost of an imperfect one low? If neither, the step is deterministic. The full argument, with what to log →
Given eight of our competitors have built a content practice around agents, it is worth being explicit about what we will not sell.
A deterministic skeleton with agents confined to individual steps. Three constraints make it hold, and each is a limitation rather than a feature.
The agent runs inside one box. It cannot decide that validation is unnecessary today, because the sequence is not its to choose — only the work within its own step.
A structured result the deterministic layer validates before acting on, not free text handed downstream. Failing validation is an exception, not a silent pass-through.
Which steps it chose and on what basis — not just the answer. That is what makes a decision reconstructable eight months later, and it is the part everyone omits until an auditor asks.
| We build | We build second | We decline |
|---|---|---|
| The deterministic skeleton, agents scoped to exception resolution and context gathering, schema validation at every boundary, intent-level logging | Draft generation for human review, routing with stated reasons, cost-per-case monitoring, a ground-truth set and evaluation harness | An agent in place of a written process. Autonomous final actions on outcomes someone can dispute. Agents replacing rules that already work. |
The reason we are confident about where the line sits is that we have shipped the same shape under four different rulebooks. Client names withheld under NDA. See full case studies →
A retrieval-augmented pipeline that reads large, inconsistent document sets and drafts structured memoranda with citations back to source, plus LLM-assisted triage and next-step suggestions wired into the existing case-management system. The model drafts; the case manager decides. That division is the reason it reached production rather than staying a demo. Read the case →
Page classification, signature and stamp detection, empty checklist cells and cross-pack number agreement — deterministic checks producing an annotated report, with a specialist signing off. No autonomy anywhere near the sign-off. Read the case →
Document authenticity, face match and liveness as a separable service that verifies an applicant and passes them to the process that decides — the same preparing-versus-deciding boundary, drawn in a different regulation. Read the case →
The fourth rulebook is GMP, where the same completeness-and-signature mechanic applies to pharmaceutical batch records — covered in AI batch record review.
On the visa platform the difficulty was not drafting quality, it was deciding what the model was allowed to assert. An early version produced fluent memoranda that stated things the source documents merely implied — plausible, useful-looking, and exactly the failure mode you cannot ship into a process where a person signs underneath. The fix was not a better model but a narrower contract: every claim carries a citation to a specific source passage, and anything without one does not appear in the draft at all. That constraint cost some fluency and is the reason the system survived review.
The stack runs on-premises or air-gapped, with no document leaving your network and GPU hardware quoted in the same contract. Details in security and compliance and private AI infrastructure.
Self-hosted open weights with pinned versions, because a re-run only means something if the model that produced the original answer still exists. A hosted endpoint can move underneath you and take your audit trail with it.
Fixed price against a scope agreed before the build starts — including the per-step decision on where an agent goes, which is the thing that determines cost per case. No time and materials.
Scope first, price second, and both numbers are published. The specification is a product: the AI Solution Blueprint, one system, 5–6 weeks, from $30,000, credited in full against the build if implementation starts within 90 days. Its main output here is the step-by-step decision on where an agent goes and where the process stays deterministic, plus a fixed price and a timeline. Builds of this kind typically start around $150,000 over 4–6 months.
Because the demo never asked the question audit asks. An agent selects its own sequence, so the same input can take a different path and produce a different answer, and internal audit wants reproducibility plus the ability to reconstruct a specific decision months later. No prompt fixes that — lowering temperature narrows the distribution without making the sequence deterministic. The fix is architectural: put the agent inside one bounded step of a deterministic process rather than in charge of the process.
Where the path is genuinely unknown in advance and a person checks the result: exception resolution, gathering context across several systems where the next lookup depends on the previous answer, drafting for review, and routing with a stated reason. Everywhere else stays deterministic, because rules are cheaper to run, fail loudly rather than silently, and explain themselves by construction.
Not where the outcome can be disputed, and we will say so before the proposal. Payment, approval, decline, release and award are the actions you have to be able to explain case by case — to a regulator in the strictest case, and far more often to a supplier, a bidder or a contractor, and an autonomous sequence cannot promise that the same facts would have produced the same action. If you want that anyway, it becomes an explicit scope item with its own risk discussion rather than something folded quietly into a document project.
Inputs as received, the model and prompt versions in effect, the intent trace — which steps the agent chose and on what basis, not merely the final answer — and the human review step with the reviewer's identity and what they were shown. The third item is the one usually missing and the one that turns a stored answer into a reconstructable decision.
Yes, and for agentic work the argument is stronger than usual. Self-hosted open weights let you pin the model version, which is what makes a re-run meaningful; a hosted endpoint can move underneath you and take your reproducibility with it. The whole stack runs on-premises or air-gapped with the GPU hardware quoted in the same contract.
An engineer replies, not an account manager. If rules would do the job, you will hear that first.
Want the spec first? AI Solution Blueprint — from $30,000, credited against the build.